A dual-graph adaptive multi-modal fusion battery soc prediction method and system

Abstract

The application relates to the technical field of battery management systems, and particularly discloses a double-graph adaptive multi-modal fusion battery SOC prediction method and system. The application aims to solve the problems of poor adaptability and insufficient model stability in the existing state of charge prediction method. The main scheme comprises the following steps: constructing a battery channel feature tensor and a temperature channel feature tensor; inputting the feature tensors into parallel battery graph neural networks and temperature graph neural networks respectively to extract spatial features; each graph neural network dynamically constructs the correlation between nodes through an adaptive learning mechanism and performs graph convolution operation to output battery spatial features and temperature spatial features respectively; performing cross-modal attention interaction and gate-weighted fusion on the battery spatial features and the temperature spatial features to obtain fusion features; and performing multi-period time series feature extraction on the fusion features, and predicting the state of charge of the power battery based on the extracted time series features.

Description

Technical Field

[0001] This invention relates to the field of battery management system technology, and provides a dual-graph adaptive multimodal fusion battery SOC prediction method and system. Background Technology

[0002] Accurately predicting the state of charge (SOC) of a power battery is a core function of the battery management system (BMS), directly affecting the estimation of the driving range of electric vehicles, the safety control of charging and discharging, and the management of battery life. SOC cannot be directly measured by sensors and must be estimated based on measurable parameters such as voltage, current, and temperature through algorithmic models.

[0003] Currently, SOC estimation methods mainly include: (1) electrochemical model-based methods, such as open-circuit voltage method and ampere-hour integration method. These methods rely on idealized assumptions and accumulate large errors under dynamic operating conditions; (2) data-driven methods, such as support vector machine and standard recurrent neural network (RNN). These methods are difficult to capture the spatial correlation and long-term temporal dependence between battery cells; (3) graph neural network (GNN)-based methods, which use graph structures to model the electrochemical coupling relationship between battery cells. However, existing schemes mostly use a single graph structure or a preset fixed graph topology (such as chain graph or fully connected graph), which fails to effectively distinguish between the two heterogeneous relationships of electrochemical coupling between battery cells and thermal conduction between temperature probes. Furthermore, the fixed graph structure is difficult to adapt to the evolution of topological relationships caused by battery aging.

[0004] In the data preprocessing stage, battery management systems (BMS) face unique challenges in cleaning streaming data. Onboard BMS uploads sampled data in real time via T-Box (typically at 0.1Hz), forming a continuously arriving data stream characterized by large data volume, long time span, and concept drift. Existing technologies mostly employ batch statistical methods (such as global Z-Score) or offline algorithms like isolated forests for anomaly detection, which cannot meet the real-time requirements of streaming scenarios. Furthermore, existing methods typically continue to update the statistical model after detecting anomalies (Always-Update strategy), leading to outliers continuously contaminating subsequent detection thresholds. Although the Welford online algorithm supports streaming mean and variance calculation, and the SPOT algorithm supports streaming extreme value detection, their collaborative application in battery SOC prediction, especially the conditional update mechanism based on the Inspect-then-Update strategy, has not yet been reported.

[0005] In recent years, time series analysis methods have made progress. TimesNet proposed using Fast Fourier Transform (FFT) to discover multi-period patterns and reshape one-dimensional time series signals into two-dimensional periodic matrices. It then utilizes two-dimensional convolution to simultaneously capture intra-period and inter-period dependencies, significantly improving the accuracy of time series prediction. However, existing TimesNet applications mostly target single time series and fail to consider the end-to-end coupling between multimodal spatial features (voltage, temperature) and time series features in battery management systems. Existing SOC prediction methods based on single-graph structures model battery cells and temperature probes uniformly, ignoring the fundamental differences in their physical spatial distribution and coupling mechanisms: battery cells are in series with electrochemical coupling; temperature probes are distributed and have thermal conduction coupling. This forced uniform modeling leads to node feature confusion and fails to accurately capture the local impact of temperature anomalies on specific battery cells.

[0006] Therefore, there is an urgent need for a SOC prediction method that can adapt to real-time cleaning of streaming data (based on Welford online statistics, SPOT extreme value detection and Inspect-then-Update strategy), model the heterogeneous relationship between battery and temperature separately (dual-graph adaptive structure), and achieve end-to-end joint optimization of spatial and temporal data (integrating TimesNet multi-period temporal modeling and GAT adaptive graph learning), in order to solve the technical problems of existing technologies such as easy contamination of statistical models in streaming anomaly detection, insufficient expressive power of single graph structure, poor adaptability of fixed topology, and insufficient fusion of spatiotemporal features. Summary of the Invention

[0007] The technical problem this invention aims to solve is to overcome the shortcomings of existing SOC prediction methods, such as the inability of a single graph structure to simultaneously capture electrochemical coupling and thermal conduction relationships, the difficulty of adapting fixed graph topologies to changes in battery dynamic characteristics, and the low prediction accuracy, poor adaptability, and insufficient model stability caused by outlier contamination statistical models. This invention provides a power battery state-of-charge prediction method based on a dual-graph adaptive structure and multimodal time-series fusion, achieving accurate modeling of heterogeneous spatial relationships, robust cleaning of streaming data, and high-precision prediction under dynamic operating conditions.

[0008] To achieve the above objectives, the present invention employs the following technical means:

[0009] This invention provides a dual-graph adaptive multimodal fusion battery SOC prediction method, comprising the following steps:

[0010] Step 1: Obtain the operating data of the power battery system, including individual battery cell data and temperature data; preprocess the operating data to construct a battery channel feature tensor. and temperature channel feature tensor ;

[0011] Step 2: Input the battery channel feature tensor and the temperature channel feature tensor into a parallel dual-graph neural network for spatial feature extraction. The dual-graph neural network includes a battery graph neural network and a temperature graph neural network. Each graph neural network dynamically constructs the relationship between nodes through an adaptive learning mechanism and performs graph convolution operations to output battery spatial features and temperature spatial features respectively.

[0012] Step 3: Perform cross-modal attention interaction and gated weighted fusion on the battery spatial features and temperature spatial features to obtain fused features;

[0013] Step 4: Extract multi-cycle time-series features from the fused features, and predict the state of charge of the power battery based on the extracted time-series features.

[0014] In existing technologies, vehicle-mounted BMS streaming data suffers from technical problems such as outlier contamination of statistical models and difficulty in uniformly representing heterogeneous electrochemical-thermal conduction data. Specifically, conventional batch statistical methods cannot meet the real-time requirements of streaming; if a conventional Always-Update strategy is used, outliers will continuously contaminate the mean and variance statistics, causing subsequent graph structure learning to be based on incorrect node similarity; and if voltage and temperature are directly concatenated and input into a single network, heterogeneous physical relationships will be confused.

[0015] Due to the aforementioned technical problems, this invention employs a synergistic combination of "flow-based robust cleaning" (S1.1-S1.2) and "heterogeneous feature separation and construction" (S1.3-S1.5) in step 1: S1.1-S1.2 achieves physical isolation between outliers and the statistical model through Welford online algorithm + SPOT extreme value detection + Inspect-then-Update strategy; S1.3-S1.5 preserves the independence of heterogeneous physical relationships by constructing dual tensor structures for battery channels and temperature channels respectively.

[0016] The combination of S1.1-S1.5 above produces a synergistic effect of "constructing heterogeneous features with anomaly immunity": the cleaned data ensures the accuracy of subsequent graph learning, and the dual-channel separation ensures the decoupling of electrochemical and thermal conduction mechanisms. If only S1.1-S1.2 is used without dual-channel separation, although the data can be cleaned, the problem of heterogeneous relationship confusion cannot be solved; if only S1.3-S1.5 is used without streaming cleaning, outliers will contaminate the statistical distribution of the two channels. It is precisely because of the above combination that robust data preprocessing in the vehicle-mounted streaming scenario is inevitably achieved.

[0017] In the above scheme, step 1 includes the following steps:

[0018] Step 1.1: Obtain the operating data of the power battery system, including the voltage data of M individual battery cells, the temperature data of N temperature probes, and the total current data;

[0019] Step 1.2: Perform robust anomaly cleaning on the running data based on the Welford online algorithm and SPOT extreme value detection, specifically including:

[0020] Maintaining the mean statistic of the data stream based on the Welford online algorithm With variance statistics The update formula for the Welford algorithm is:

[0021]

[0022] in Indicates deviation, For the current data point, This is the current average. The current variance, It is a second-order moment accumulator;

[0023] Calculate the Z-Score vector for each data point based on the mean and variance. And calculate the abnormal score:

[0024]

[0025] in The number of feature dimensions. For the first VI's Z-Score;

[0026] The SPOT extreme value detection algorithm is used to dynamically set the anomaly score based on extreme value theory. Detection threshold ;

[0027] Execute the Inspect-then-Update strategy: only when a decision is made When the current data point is included in the update of the mean and variance statistics, if... If the value is not found, it is considered abnormal and removed, and the update of the statistics at that point is blocked.

[0028] Step 1.3: Based on the cleaned individual battery cell data, form the battery channel feature tensor. ,in For time step, The battery characteristics are defined as follows: voltage, current, and power. ;

[0029] Step 1.4: Generate temperature channel feature tensors based on the cleaned temperature data ,in Temperature is a characteristic dimension;

[0030] Step 1.5: Perform Z-Score normalization on the battery channel feature tensor and the temperature channel feature tensor respectively. The normalization formula is as follows:

[0031]

[0032] in and These are the mean and standard deviation of the training set, respectively.

[0033] Existing single-graph structures cannot simultaneously capture electrochemical coupling and thermal conduction relationships, and fixed adjacency matrices are ill-suited to adapting to topological evolution caused by battery aging. Therefore, in step 2 of this invention, a combination of the "dual-graph parallel architecture" in S2.1-S2.2 and the "IDGL dynamic graph learning" in S2.3-S2.4 is adopted: the dual-graph parallel architecture processes the relationships between individual battery cells and the temperature probe relationships respectively, avoiding confusion of heterogeneous features; IDGL dynamically learns the adjacency matrix based on the features of the current batch, replacing the fixed topology.

[0034] This approach yields a synergistic effect of "heterogeneous decoupling and topology adaptation": the separation of the two graphs ensures independent modeling of electrochemical-thermal conduction, while IDGL dynamic learning ensures adaptive evolution of the topology as the battery ages. Using a single graph for unified modeling simplifies the structure but leads to confusion regarding heterogeneous relationships; using a fixed adjacency matrix ensures stable training but cannot adapt to changes in operating conditions. It is precisely this combination of two graphs and dynamic learning that inevitably achieves accurate representation of heterogeneous relationships and adaptive optimization of the topology.

[0035] In the above scheme, step 2 includes the following steps:

[0036] Step 2.1: Convert the battery channel feature tensors separately. and temperature channel feature tensor The data is input to a parallel battery graph attention network and a temperature graph attention network, wherein the battery graph attention network is used to handle the electrochemical coupling relationship between battery cells, and the temperature graph attention network is used to handle the thermal conduction relationship between temperature probes;

[0037] Step 2.2: In each graph attention network, execute the iterative deep graph learning (IDGL) module and the graph attention convolution (GAT) module in sequence;

[0038] Step 2.3: In the IDGL module, the similarity matrix between nodes is adaptively calculated based on the features of the current node, specifically including:

[0039] Node features are projected into the hidden space using a multi-head perceptron, and the attention score between node pairs is calculated as a similarity measure.

[0040] Based on the similarity matrix, the node with the highest similarity is retained. Connect each neighbor to construct a sparse adjacency matrix;

[0041] The adjacency matrix is ​​dynamically updated in each training batch to adapt to the dynamic changes in the battery system's characteristics.

[0042] Step 2.4: In the GAT module, perform a multi-head attention graph convolution operation based on the dynamic adjacency matrix;

[0043] Step 2.5: Output the battery space characteristics respectively Temperature spatial characteristics ,in For the hidden layer dimension.

[0044] The static weighted fusion of existing technologies cannot adapt to the dynamic changes in the importance of battery features and temperature features under different operating conditions. Therefore, in step 3 of this invention, a combination of "cross-modal attention interaction" in S3.2 and "gated dynamic weighting" in S3.3-S3.4 is adopted: cross-modal attention enables the bimodal features to enhance each other; the gating mechanism dynamically calculates the fusion weights through sigmoid.

[0045] This resulted in a synergistic effect of "condition-adaptive modal coupling": cross-modal attention solved the problem of insufficient information from a single modality, while gating mechanisms addressed the issue of weight rigidity caused by changes in operating conditions. If only cross-modal attention is used without gating, interaction is possible but dynamic adjustment of contributions is impossible; if only gating is used without cross-modal interaction, dynamic weighting is possible but inter-modal information enhancement is lacking. It is precisely this combination of interaction and gating that inevitably achieves robust fusion under extreme operating conditions.

[0046] In the above scheme, step 3 includes the following steps:

[0047] Step 3.1: Compress the battery spatial features and temperature spatial features into global feature vectors using attention pooling layers. and ;

[0048] Step 3.2: Perform cross-modal attention interaction, specifically including:

[0049] Calculate the battery feature query matrix Temperature-related feature bond matrix The similarity is used to obtain the cross-modal attention weights:

[0050]

[0051] Based on the attention weights of the temperature feature matrix Weighted polymerization was performed to obtain temperature-enhanced battery characteristics:

[0052]

[0053] Similarly, calculate the enhanced temperature characteristics of the battery. ;

[0054] in, The feature dimension scaling factor;

[0055] Step 3.3: Dynamically calculate the dual-modal fusion weights using a gating mechanism. , ,satisfy ;

[0056] Step 3.4: Based on the weights, perform weighted fusion of the dual-modal features to obtain the fused features:

[0057] .

[0058] The existing concatenated architecture of "spatial pooling first, then temporal modeling" results in a significant loss of spatial topology information before temporal modeling. Therefore, in step 4 of this invention, a multi-period extraction using TimesNet directly coupled to S3 is employed, instead of the traditional spatial-temporal decoupling architecture: preserving the topology-rich fusion features H of the S3 output. f The temporal pattern and spatial features are jointly processed by two-dimensional convolution through FFT periodicity detection.

[0059] This approach yields a synergistic effect of "spatiotemporal end-to-end joint optimization": it avoids the loss of spatial details before temporal modeling and achieves deep coupling between topological relationships and temporal cycles. While a traditional serial architecture simplifies the network, it loses spatial locality information; while using only one-dimensional convolution to process temporal data captures the temporal sequence, it cannot utilize the intra-cycle to periodic modeling capabilities of two-dimensional convolution. It is precisely because of the end-to-end architecture employing spatial feature preservation and TimesNet multi-cycle extraction that high-precision SOC tracking under dynamic conditions is inevitably achieved.

[0060] In the above scheme, step 4 includes the following steps:

[0061] Step 4.1: Input the fused features into the temporal feature extraction network, perform Fast Fourier Transform (FFT) on the input temporal features, and calculate the frequency amplitude spectrum;

[0062] Step 4.2: Select the Top-P frequency components with the largest amplitudes and calculate the corresponding period lengths. ;

[0063] Step 4.3: For each cycle , with a length of One-dimensional sequence rearrangement A two-dimensional matrix, where rows represent the number of periods and columns represent the positions within a period;

[0064] Step 4.4: Process the two-dimensional matrix in parallel using multi-scale two-dimensional convolutional kernels to extract the dual dependencies within and between periods. The size of the multi-scale two-dimensional convolutional kernels includes... , , , , , ;

[0065] Step 4.5: Perform temporal pooling on the extracted temporal features to obtain a temporal feature vector, and input the temporal feature vector into the multilayer perceptron (MLP) regressor to output a normalized SOC prediction value.

[0066] In the above scheme, the SPOT extreme value detection algorithm includes:

[0067] During the preheating phase, initial samples are collected. One outlier score is used to initialize the statistics;

[0068] During the calibration phase, the initial anomaly scores collected during the warm-up phase are used. First calculate its quantiles as the initial threshold And extract the set of peak values ​​that exceed the threshold. .

[0069] The Grimshaw algorithm is used to perform maximum likelihood estimation of the parameters of the generalized Pareto distribution (GPD): by solving the log-likelihood function. The extreme values ​​are obtained by iterative optimization to obtain the estimated shape parameters. and scale parameter estimates .

[0070] Extreme value quantiles are calculated based on the estimated GPD parameters and used as dynamic detection thresholds. :

[0071] when When using this formula:

[0072] when When using this formula:

[0073] Where n is the number of preheated samples, N is the number of peak values ​​exceeding the initial threshold, and q is the preset anomaly probability threshold.

[0074] During the online detection phase, as streaming data continues to arrive, the GPD parameters are updated incrementally based on the newly accumulated peak data at preset time intervals. and And recalculate the extreme value quantiles. This enables dynamic adaptive adjustment of the detection threshold.

[0075] The present invention also provides a dual-graph adaptive multimodal fusion battery SOC prediction system, including a processor and a storage medium, wherein the processor executes a program in the storage medium to implement the method described therein.

[0076] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0077] 1. This invention solves the technical problem of confusion between the statistical model of outlier contamination and the electrochemical-thermal conduction heterogeneous relationship in the context of vehicle-mounted BMS streaming data by combining the "streaming robust cleaning and heterogeneous feature construction" techniques in steps S1.1-S1.5 of step 1, and achieves the heterogeneous data characterization effect of anomaly immunity.

[0078] 2. By combining the "dual-graph parallel architecture and IDGL dynamic graph learning" techniques in steps S2.1-S2.5 of step 2, this invention solves the technical problems that a single graph structure cannot distinguish heterogeneous physical coupling relationships and that a fixed topology is difficult to adapt to battery aging evolution, thus achieving accurate spatial modeling effects of heterogeneous decoupling and topology adaptation.

[0079] 3. By combining the "cross-modal attention interaction and gating dynamic weighting" techniques in steps S3.1-S3.4 of step 3, this invention solves the technical problems that static fusion weights cannot adapt to extreme working condition changes and that single modal information is insufficient, and achieves a modal collaborative enhancement effect that is adaptive to working conditions.

[0080] 4. By combining the "spatial feature preservation and multi-period temporal end-to-end fusion" techniques in steps 3 and 4, this invention solves the technical problem of spatial topology information loss caused by the traditional "space first, then temporal" serial architecture, and achieves end-to-end optimization effect of deep spatiotemporal coupling.

[0081] 5. This invention solves the chain of technical problems of "data pollution-topology rigidity-modal confusion-temporal drift" in the vehicle BMS scenario by combining the "streaming robust cleaning" in step 1, the "dual-graph dynamic learning" in step 2, the "cross-modal gating fusion" in step 3, and the "multi-cycle temporal extraction" in step 4 in the end-to-end process, and achieves a high-precision and stable technical effect throughout the entire life cycle. Attached Figure Description

[0082] Figure 1This is an overall flowchart provided in the embodiments of the present invention;

[0083] Figure 2 This is a system diagram provided in an embodiment of the present invention;

[0084] Figure 3 This is a model diagram for predicting the state of charge of a power battery based on a dual-graph adaptive structure and multimodal time-series fusion, provided in an embodiment of the present invention. Detailed Implementation

[0085] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0086] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0087] The technical problem to be solved by this invention is to overcome the technical defects of existing SOC prediction methods, such as the inability of a single graph structure to simultaneously capture electrochemical coupling and thermal conduction relationships, the difficulty of adapting fixed graph topologies to changes in battery dynamic characteristics, and the susceptibility of outlier detection to contamination. This invention provides a power battery state of charge prediction method based on a dual-graph adaptive structure and multimodal time-series fusion, which achieves accurate modeling of heterogeneous spatial relationships, robust cleaning of streaming data, and high-precision prediction under dynamic operating conditions.

[0088] This invention provides a method for predicting the state of charge of a power battery based on a dual-graph adaptive structure, comprising the following steps:

[0089] S1: Heterogeneous Data Preprocessing and Dual-Channel Feature Construction

[0090] Obtain operational data of the power battery system, including Voltage data of individual battery cells, Temperature data from each temperature probe and total current data; robust anomaly removal and normalization processing using Welford online algorithm and SPOT extreme value detection; battery channel feature tensor formed based on the battery cell data. ,in For time step, For battery feature dimensions; a temperature channel feature tensor is formed based on the temperature data. ,in Temperature is a characteristic dimension;

[0091] The robust anomaly cleaning based on the Welford online algorithm and SPOT extreme value detection includes:

[0092] Maintaining the mean statistic of the data stream based on the Welford online algorithm With variance statistics The Welford algorithm, described above, updates statistics online through a single data iteration, avoiding the need to store all historical data. Unlike batch computation methods, the Welford algorithm updates statistics online with a single iteration, eliminating the need to store all historical data and offering the advantage of good numerical stability. The update formula for the online Welford algorithm is as follows:

[0093]

[0094] in, For the current data point, This is the current average. The current variance, It is a second-order moment accumulator.

[0095] Calculate the Z-Score vector for each data point based on the mean and variance. and calculate abnormal scores. This represents the mean of the absolute Z-scores for each dimension. The anomaly score integrates the degree of deviation across all feature dimensions, enabling the identification of cases where multiple dimensions exhibit slight anomalies simultaneously.

[0096] The SPOT extreme value detection algorithm is used to dynamically set the detection threshold for the outlier score based on extreme value theory. Unlike fixed threshold methods, the SPOT algorithm can adapt to the time-varying characteristics of data distribution and dynamically adjust the detection threshold.

[0097] The SPOT extreme value detection algorithm includes:

[0098] Preheating phase: collecting initial... One outlier score is used to initialize the statistics; Calibration phase: based on the initial outlier scores... The initial threshold is set for quantiles. The parameters of the generalized Pareto distribution (GPD) are estimated using the Grimshaw algorithm, and the extreme value quantiles are calculated as dynamic detection thresholds. In the online detection phase, for each new data point, if the anomaly score is lower than the dynamic threshold, it is determined to be NORMAL; if it is higher, it is determined to be ALARM, and the threshold is dynamically updated.

[0099] The Inspect-then-Update strategy is implemented: the current data point is only included in the update of the mean and variance statistics if the outlier score is determined to be below the dynamic threshold; if it is determined to be an outlier, the point is removed and its contribution to the statistical model update is blocked. It is precisely because of the Inspect-then-Update conditional update mechanism that the isolation between outliers and the statistical model is inevitably achieved, solving the technical problem of outliers continuously polluting subsequent detections and ensuring the purity of the training data.

[0100] In the above scheme, before step S2, the following steps are also included:

[0101] Feature attention weighting is applied to the battery channel feature tensor, and the importance weights of the three feature dimensions of voltage, current, and power are automatically learned through a learnable attention network. The weights are calculated using a temperature-scaling sigmoid function and satisfy a minimum weight constraint to prevent feature degradation.

[0102] S2: Adaptive Spatial Feature Extraction from Two Graphs

[0103] Construct battery relationship diagrams respectively Relationship with temperature , where the node set , These correspond to individual battery cells and temperature probes, respectively; the battery channel feature tensor and temperature channel feature tensor are input into parallel battery graph attention networks and temperature graph attention networks, respectively.

[0104] In each graph attention network, an iterative deep graph learning (IDGL) module and a graph attention convolution (GAT) module are executed sequentially: In the IDGL module, a similarity matrix between nodes is adaptively calculated based on the current node features. Specifically, a multi-head perceptron projects the node features into the hidden space, calculates the attention score between node pairs as a similarity metric, and retains the K most similar neighbors for each node based on the similarity matrix to construct a sparse adjacency matrix; In the GAT module, a multi-head attention graph convolution is performed based on the dynamic adjacency matrix to output the battery space features. Temperature spatial characteristics , where D is the dimension of the hidden layer.

[0105] The adjacency matrix is ​​dynamically updated in each training batch to adapt to changes in the dynamic characteristics of the battery system. It is precisely this synergistic combination of IDGL adaptive graph learning and GAT graph convolution that inevitably enables dynamic optimization of the graph topology as the battery state evolves and the extraction of high-order spatial features, solving the technical problem of performance degradation in long-term prediction of fixed graph structures and their inability to adapt to battery aging and changes in operating conditions.

[0106] S3: Gated cross-modal adaptive fusion

[0107] The battery spatial features and temperature spatial features are compressed into global feature vectors through attention pooling layers. and Perform cross-modal attention interaction: Calculate temperature-enhanced battery features using battery features as queries and temperature features as keys, and calculate battery-enhanced temperature features using temperature features as queries and battery features as keys; Dynamically calculate bimodal fusion weights through a gating mechanism. , ,satisfy The dual-modal features are weighted and fused based on the aforementioned weights to obtain the fused features:

[0108]

[0109] It is precisely because of the adoption of the gated cross-modal adaptive fusion mechanism that dynamic modeling of the complex nonlinear coupling relationship between battery and temperature is inevitably realized, which significantly improves the prediction robustness of the model under different thermal management conditions.

[0110] Performing cross-modal attention interactions specifically includes:

[0111] Calculate the battery feature query matrix Temperature-related feature bond matrix The similarity is used to obtain the cross-modal attention weights:

[0112]

[0113] Based on the attention weights of the temperature feature matrix Weighted polymerization was performed to obtain temperature-enhanced battery characteristics:

[0114]

[0115] Similarly, calculate the enhanced temperature characteristics of the battery. ;in, This is the feature dimension scaling factor.

[0116] S4: Multi-period temporal feature extraction and SOC prediction

[0117] The fused features are input into a time-series feature extraction network, which automatically discovers the dominant period in the time-series data based on the Fast Fourier Transform (FFT), reconstructs the one-dimensional time-series signal into a two-dimensional periodic matrix, extracts multi-scale time-series patterns using a convolutional neural network, and outputs a time-series feature vector; the time-series feature vector is input into a multilayer perceptron (MLP) regressor, which outputs a normalized SOC prediction value.

[0118] Step S4, which involves automatically identifying the dominant period in time series data based on the Fast Fourier Transform, specifically includes:

[0119] Perform an FFT transform on the input time-series features to calculate the frequency amplitude spectrum; select the Top-order features with the largest amplitudes. Calculate the corresponding period length for each frequency component. For each cycle , with a length of One-dimensional sequence rearrangement A two-dimensional matrix is ​​formed, where rows represent the number of periods and columns represent the positions within a period. The two-dimensional matrix is ​​processed in parallel using multi-scale two-dimensional convolution kernels to extract the dual dependencies between periods.

[0120] The temporal feature extraction network adopts a TimesNet-based temporal network, which uses a two-layer temporal block structure, with each layer having two dominant periods (Top-...). The kernel sizes are respectively , , , , , .

[0121] Example 1: Basic Example (corresponding to the widest scope of claim 1)

[0122] This embodiment uses real-vehicle operating data of a certain ternary lithium battery system, including... A series-connected battery cell and There are several temperature probes, with a sampling frequency of [missing information]. .

[0123] S1: Data preprocessing and feature construction: Obtain the original CSV data, which includes the fields: terminaltime (time stamp), batteryvoltage (96 individual voltage connection strings), probetemperatures (32 temperature probe connection strings), and totalcurrent (total current).

[0124] Perform robust anomaly cleaning based on Welford online algorithm and SPOT extreme value detection:

[0125] Preheating phase: Settings Initialize the mean Second-order moment accumulator ,counter ;

[0126] Online updates and detection: for each new data point (Including characteristics such as voltage, current, and power):

[0127] counter ;

[0128] Calculate the deviation ;

[0129] Update the mean ;

[0130] Update second moment ;

[0131] Calculate variance (when );

[0132] Calculate Z-Score ;

[0133] Calculate abnormal scores ;

[0134] SPOT threshold determination: 100 samples collected during the preheating phase. Used to initialize SPOT (set) (Initial threshold quantile 0.98), calculate the extreme value quantile as the dynamic threshold. During the online testing phase, if If it is determined to be NORMAL, Inspect-then-Update (update using the current point) is executed. and );like If the value is identified as ALARM, the point is removed and the statistics are not updated.

[0135] The cleaned voltage string is divided into 96 columns, and the temperature string is divided into 32 columns.

[0136] Constructing battery channel characteristics: Calculating power from the voltage of 96 individual cells and the total current after broadcasting ( ), forming a shape The tensor has three feature dimensions: voltage, current, and power. Temperature channel features are constructed using 32 temperature probe values, with the shape of... .

[0137] Z-score normalization is performed on the two-channel features separately, with parameters calculated only from the training set:

[0138]

[0139] Build a sliding window: Settings , The generated input sample shape is and The label represents the SOC value at the last moment of the window.

[0140] S2: Dual-graph adaptive spatial feature extraction battery graph attention network configuration: , , , (Top-K retains 10 neighbors). Temperature map attention network configuration: , , , .

[0141] IDGL Graph Learning: Node similarity is calculated using 4-head attention. For each node, the top-10 similar neighbors are retained, constructing a sparse adjacency matrix (sparseness approximately...). GAT Convolution: Performs 4-head graph attention convolution based on a dynamic adjacency matrix, outputting dimensions... After RMSnorm normalization, the 96-node / 32-node vector is compressed into a 128-dimensional global vector through attention pooling.

[0142] S3: Gated cross-modal adaptive fusion execution cross-modal attention: Battery features (Query) interact with temperature features (Key / Value), and temperature features (Query) interact with battery features (Key / Value) to obtain enhanced bimodal features.

[0143] Gated computation: splicing bimodal features and using MLP ( The original gate value is calculated and then normalized using Softmax. , ,satisfy Weighted fusion:

[0144]

[0145] And add residual connections (fusion residual weights) ).

[0146] S4: Temporal network configuration for multi-period temporal feature extraction and SOC prediction: input dimension 128, hidden dimension 64, sequence length 32, number of temporal block layers 2, number of dominant periods retained. Convolution kernel number 6 (size) ).

[0147] Performing an FFT reveals the dominant cycle. The temporal features are reshaped into a 2D matrix, and after feature extraction through multi-scale convolution, mean-over-time pooling is used to obtain 64-dimensional features.

[0148] MLP regressor: Activate with GELU and Dropout ( Finally, the Sigmoid outputs the SOC prediction value (range). ).

[0149] Training configuration: AdamW optimizer, initial learning rate Cosine annealing to , , Stop early .

[0150] Experimental results: On the test set, the root mean square error (RMSE) is... The mean absolute error (MAE) is .

Claims

1. A dual-graph adaptive multimodal fusion battery SOC prediction method, characterized in that, Includes the following steps: Step 1: Obtain the operating data of the power battery system, including individual battery cell data and temperature data; The operational data is preprocessed to construct a battery channel feature tensor. and temperature channel feature tensor ; Step 2: Input the battery channel feature tensor and the temperature channel feature tensor into a parallel dual-graph neural network for spatial feature extraction. The dual-graph neural network includes a battery graph neural network and a temperature graph neural network. Each graph neural network dynamically constructs the relationship between nodes through an adaptive learning mechanism and performs graph convolution operations to output battery spatial features and temperature spatial features respectively. Step 3: Perform cross-modal attention interaction and gated weighted fusion on the battery spatial features and temperature spatial features to obtain fused features; Step 4: Extract multi-cycle time-series features from the fused features, and predict the state of charge of the power battery based on the extracted time-series features.

2. The method according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Obtain the operating data of the power battery system, including the voltage data of M individual battery cells, the temperature data of N temperature probes, and the total current data; Step 1.2: Perform robust anomaly cleaning on the running data based on the Welford online algorithm and SPOT extreme value detection, specifically including: Maintaining the mean statistic of the data stream based on the Welford online algorithm With variance statistics The update formula for the Welford algorithm is: in Indicates deviation, For the current data point, This is the current average. The current variance, It is a second-order moment accumulator; Calculate the Z-Score vector for each data point based on the mean and variance. And calculate the abnormal score: in The number of feature dimensions. For the first VI's Z-Score; The SPOT extreme value detection algorithm is used to dynamically set the anomaly score based on extreme value theory. Detection threshold ; Execute the Inspect-then-Update strategy: only when a decision is made When the current data point is included in the update of the mean and variance statistics, if... If the value is not found, it is considered abnormal and removed, and the update of the statistics at that point is blocked. Step 1.3: Based on the cleaned individual battery cell data, form the battery channel feature tensor. ,in For time step, The battery characteristics are defined as follows: voltage, current, and power. ; Step 1.4: Generate temperature channel feature tensors based on the cleaned temperature data ,in Temperature is a characteristic dimension; Step 1.5: Perform Z-Score normalization on the battery channel feature tensor and the temperature channel feature tensor respectively. The normalization formula is as follows: in and These are the mean and standard deviation of the training set, respectively.

3. The method according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Convert the battery channel feature tensors separately. and temperature channel feature tensor The data is input to a parallel battery graph attention network and a temperature graph attention network, wherein the battery graph attention network is used to handle the electrochemical coupling relationship between battery cells, and the temperature graph attention network is used to handle the thermal conduction relationship between temperature probes; Step 2.2: In each graph attention network, execute the iterative deep graph learning (IDGL) module and the graph attention convolution (GAT) module in sequence; Step 2.3: In the IDGL module, the similarity matrix between nodes is adaptively calculated based on the features of the current node, specifically including: Node features are projected into the hidden space using a multi-head perceptron, and the attention score between node pairs is calculated as a similarity measure. Based on the similarity matrix, the node with the highest similarity is retained. Connect each neighbor to construct a sparse adjacency matrix; The adjacency matrix is ​​dynamically updated in each training batch to adapt to the dynamic changes in the battery system's characteristics. Step 2.4: In the GAT module, perform a multi-head attention graph convolution operation based on the dynamic adjacency matrix; Step 2.5: Output the battery space characteristics respectively Temperature spatial characteristics ,in For the hidden layer dimension.

4. The method according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Compress the battery spatial features and temperature spatial features into global feature vectors using attention pooling layers. and ; Step 3.2: Perform cross-modal attention interaction, specifically including: Calculate the battery feature query matrix Temperature-related feature bond matrix The similarity is used to obtain the cross-modal attention weights: Based on the attention weights of the temperature feature matrix Weighted polymerization was performed to obtain temperature-enhanced battery characteristics: Similarly, calculate the enhanced temperature characteristics of the battery. ; in, The feature dimension scaling factor; Step 3.3: Dynamically calculate the dual-modal fusion weights using a gating mechanism. , ,satisfy ; Step 3.4: Based on the weights, perform weighted fusion of the dual-modal features to obtain the fused features: 。 5. The method according to claim 1, characterized in that: Step 4 includes the following steps: Step 4.1: Input the fused features into the temporal feature extraction network, perform Fast Fourier Transform (FFT) on the input temporal features, and calculate the frequency amplitude spectrum; Step 4.2: Select the Top-P frequency components with the largest amplitudes and calculate the corresponding period lengths. ; Step 4.3: For each cycle , with a length of One-dimensional sequence rearrangement A two-dimensional matrix, where rows represent the number of periods and columns represent the positions within a period; Step 4.4: Process the two-dimensional matrix in parallel using multi-scale two-dimensional convolutional kernels to extract the dual dependencies within and between periods. The size of the multi-scale two-dimensional convolutional kernels includes... , , , , , ; Step 4.5: Perform temporal pooling on the extracted temporal features to obtain a temporal feature vector, and input the temporal feature vector into the multilayer perceptron (MLP) regressor to output a normalized SOC prediction value.

6. The method according to claim 1, characterized in that: The SPOT extreme value detection algorithm includes: During the preheating phase, initial samples are collected. One outlier score is used to initialize the statistics; During the calibration phase, the initial anomaly scores collected during the warm-up phase are used. First calculate its quantiles as the initial threshold And extract the set of peak values ​​that exceed the threshold. , Represents the q-quantile; The parameters of the generalized Pareto distribution (GPD) are estimated using the Grimshaw algorithm by solving the log-likelihood function. The extreme values ​​are obtained by iterative optimization to obtain the estimated shape parameters. and scale parameter estimates ; Extreme value quantiles are calculated based on the estimated GPD parameters and used as dynamic detection thresholds. : when When using this formula: when When using this formula: Where n is the number of preheated samples, N is the number of peak values ​​exceeding the initial threshold, and q is the preset anomaly probability threshold; During the online detection phase, as streaming data continues to arrive, the GPD parameters are updated incrementally based on the newly accumulated peak data at preset time intervals. and And recalculate the extreme value quantiles. This enables dynamic adaptive adjustment of the detection threshold.

7. A dual-graph adaptive multimodal fusion battery SOC prediction system, comprising a processor and a storage medium, characterized in that, When the processor executes a program in the storage medium, it implements the method as described in any one of claims 1-6.

8. A storage medium, characterized in that, When the processor executes a program in the storage medium, it implements the method as described in any one of claims 1-6.

Images