Battery SOH (state of health) prediction method integrating signal physical denoising and space-time correlation modeling

Abstract

The invention discloses a battery SOH (state of health) prediction method integrating signal physical denoising and space-time correlation modeling. The method comprises the following steps: extracting original health features in battery charging and discharging data; key features are screened through a Pearson's correlation coefficient, and outliers are cleaned by using a Hampel filter; constructing a sliding window VMD model based on mirror image continuation, performing online multi-scale decomposition on the feature sequence to obtain a trend item, a fluctuation item and a noise item, and generating an enhanced feature vector in combination with position coding; constructing a dynamic graph topology based on the correlation between the features; constructing a GAT-GRU deep learning model, carrying out space-time fusion through a space attention and gating circulation unit, and outputting an SOH predicted value; real-time data is processed and then input into the trained model for online reasoning; according to the method, a mirror image continuation sliding window VMD strategy is adopted, high-fidelity feature extraction is realized, multi-scale health features are fused, a GAT-GRU model architecture is designed, the online prediction precision is improved, and the prediction stability is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of battery management system and new energy storage technology, and particularly relates to a battery SOH prediction method that integrates signal physical denoising and spatiotemporal correlation modeling. Background Technology

[0002] With the widespread adoption of new energy vehicles and energy storage power stations, the accurate monitoring of the State of Health (SOH) of lithium-ion batteries, as the core energy storage unit, has become a crucial aspect of ensuring the safe, stable, and economical operation of the system. Battery aging is a complex electrochemical nonlinear time-varying process, manifested not only as long-term capacity degradation but also as local fluctuations such as "capacity regeneration" caused by mechanisms like electrochemical relaxation. This places higher demands on the accuracy and adaptability of SOH monitoring technology. Against this backdrop, efficient and reliable SOH prediction technology has become an important research direction in the development and engineering application of battery management systems (BMS).

[0003] Currently, data-driven SOH prediction methods have become mainstream because they do not rely on complex internal mechanism models. A typical implementation usually includes the following steps: First, external monitoring data such as voltage, current, and temperature are collected during the battery charging and discharging process, and multiple aging-related health features are extracted from them, such as constant current charging time and voltage curve integral. Second, signal decomposition methods (such as Empirical Mode Decomposition (EMD) or Variational Mode Decomposition (VMD) are used to perform offline noise reduction and trend separation on the full life cycle data. Finally, a deep learning model (such as Long Short-Term Memory Network (LSTM) or Graph Convolutional Network (GCN)) is constructed, and the preprocessed features are concatenated and input into the network. The model is then trained to fit the battery capacity degradation trajectory, thereby completing the SOH estimation.

[0004] The existing methods described above still have significant shortcomings in practical applications, especially in online real-time prediction scenarios: First, existing VMD or EMD decompositions typically rely on complete lifetime data for offline processing, while online applications can only decompose based on a finite-length sliding window of data. This leads to severe endpoint effects (Gibbs phenomenon) at the end of the data sequence, causing distortion of extracted trend features and failing to meet the requirements for high-precision real-time SOH prediction. Second, existing models mostly employ simple feature splicing or static weighted fusion strategies. When a sensor data becomes abnormal due to interference, the model lacks mutual verification and noise shielding mechanisms between features, making the prediction results susceptible to single-point noise and exhibiting poor robustness against interference. Furthermore, traditional time-series models have limited ability to capture non-stationary features such as "long-term degradation" and "short-term capacity fluctuations" during battery aging, especially prone to tracking lag at capacity regeneration inflection points. Finally, existing graph network methods often use fixed-structure adjacency matrices, failing to reflect the dynamic coupling relationship between battery aging stage features and SOH, limiting the model's adaptability and prediction accuracy. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a battery SOH prediction method that integrates signal physical denoising and spatiotemporal correlation modeling, thereby solving the problems of endpoint effect, feature noise, and non-stationary dynamic feature capture in existing technologies.

[0006] Technical solution: The battery SOH prediction method based on the fusion of signal physical denoising and spatiotemporal correlation modeling described in this invention includes the following steps:

[0007] Step S1: Collect charge and discharge monitoring data of the lithium-ion battery throughout its entire life cycle, and extract original health characteristic indicators from the charge and discharge monitoring data that reflect the increase in internal polarization resistance, loss of active material and thermodynamic properties of the battery.

[0008] Step S2: The original health characteristic indicators are screened using the Pearson correlation coefficient PCC analysis method, and the features with an absolute value of Pearson correlation coefficient with battery capacity that is greater than a preset threshold are retained to obtain a feature set;

[0009] Step S3: Use the Hampel filter to clean the feature sequences corresponding to the feature set, and remove bursty outliers from the feature sequences;

[0010] Step S4: Construct a sliding window variational mode decomposition (VMD) model based on mirror extension, perform online multi-scale frequency domain enhancement on the cleaned feature sequence, separate the trend term, fluctuation term and noise term, and generate enhanced feature vectors by combining normalized position encoding;

[0011] Step S5: Construct a graph topology based on the correlation strength between features in the feature set, and generate the corresponding weighted adjacency matrix;

[0012] Step S6: Construct a graph attention network-gated recurrent unit (GAT-GRU) deep learning prediction model containing a spatial attention aggregation layer and a temporal memory evolution layer. Perform spatiotemporal fusion on the enhanced feature vector through the spatial attention aggregation layer and the temporal memory evolution layer, and output the SOH prediction value.

[0013] Step S7: After processing the real-time collected battery data through steps S1 to S6, input it into the trained GAT-GRU deep learning prediction model for online inference to obtain the current SOH estimate of the battery.

[0014] This invention effectively solves the problems of endpoint effects, feature noise, and difficulty in capturing non-stationary dynamic features in existing battery SOH prediction: It cleans the feature sequence using a Hampel filter, significantly suppressing interference from sudden outliers; it utilizes a mirror-extended sliding window VMD model to achieve online multi-scale frequency domain decomposition, accurately separating trend, fluctuation, and noise terms while avoiding endpoint effects, thus improving the feature representation quality; furthermore, it constructs a graph topology based on feature correlation and combines it with a deep learning model that integrates spatial attention and gated recurrent units, achieving in-depth mining and adaptive fusion of complex spatiotemporal correlations and non-stationary dynamic evolution laws during battery degradation, thereby significantly improving the robustness and accuracy of SOH prediction.

[0015] Preferably, the original health characteristic indicators mentioned in step S1 include voltage characteristic, current characteristic and thermodynamic characteristic; the voltage characteristic includes the slope of the constant current charging voltage curve, the integral of the voltage curve during the constant current charging stage, the constant current charging time, and the peak value of the incremental capacity analysis ICA curve; the current characteristic includes the integral of the charging current curve; the thermodynamic characteristic includes the time to reach the temperature peak.

[0016] This invention extracts multi-dimensional and multi-physically meaningful original health characteristic indicators from battery charge and discharge monitoring data. These indicators not only cover voltage characteristics reflecting changes in the battery's internal polarization resistance and current characteristics reflecting the degree of participation of active materials in reactions, but also include temperature characteristics characterizing the stability of thermodynamic processes. This comprehensive feature extraction strategy can characterize the complex electrochemical and thermodynamic degradation behavior within the battery from multiple angles and at a deeper level, providing a rich and complementary information foundation for subsequent accurate modeling. This effectively enhances the ability to describe and broaden the information coverage of battery health state degradation modes and evolution trajectories.

[0017] Preferably, step S3, which involves using a Hampel filter to clean the feature sequences corresponding to the feature set, includes:

[0018] S31: For a feature sequence of length N, set the sliding window length to be... ;

[0019] S32: Calculate the median of the feature data within the sliding window. and absolute median :

[0020]

[0021]

[0022] in, For feature data within the sliding window, This is a function for calculating the median;

[0023] S33: Calculate the standard deviation estimate ;

[0024] S34: If a data point within the sliding window is... The absolute value of the difference is greater than That is, the data point is determined to be an outlier, and is then... Replace the outlier value.

[0025] By employing a Hampel filter based on the median and absolute median difference for data cleaning, it is possible to effectively identify and suppress sudden outliers caused by measurement interference or instantaneous sensor failures. The core of this approach lies in utilizing the natural robustness of the median to replace the mean, and combining it with a robust absolute median difference to construct an adaptive anomaly detection threshold. This allows for the smooth filtering of singular values ​​without significantly disturbing the overall data distribution, significantly improving the local stability and smoothness of the feature sequence. This lays a high-quality data foundation for subsequent feature decomposition and model learning, and enhances the anti-interference capability of the SOH prediction model.

[0026] Preferably, step S4, which involves constructing a sliding window variational mode decomposition (VMD) model based on mirror extension, includes:

[0027] S41: Sliding window construction: At prediction time t, extract a historical feature window sequence of length W. ;

[0028] S42: Mirror continuation processing: using the feature data corresponding to the predicted time t Using the axis of symmetry, the data at the end of the historical feature window sequence is mirrored and concatenated to the historical feature window sequence to generate an extended sequence. :

[0029]

[0030] in, For the extension length;

[0031] S43: VMD Decomposition: Perform variational mode decomposition on the extended sequence to obtain the low-frequency trend mode. Secondary low frequency fluctuation mode and high-frequency noise modes ;

[0032] S44: Feature Fusion: Calculate Normalized Position Encoding This is used to supplement the global lifetime process information lost by the sliding window:

[0033]

[0034] in, This indicates the current cycle number of the battery. It is a priori value of the maximum battery life based on statistics from the training set;

[0035] The original observations of the historical feature window sequence, the low-frequency trend mode, the second-lowest frequency fluctuation mode, the high-frequency noise mode, and the normalized position code are concatenated to generate an enhanced feature vector. :

[0036] .

[0037] By employing a mirror-extended sliding window VMD model, global lifetime location information is cleverly combined with local multi-scale feature analysis. Mirror extension effectively mitigates the endpoint effect problem at the window edges of signal decomposition, ensuring the accuracy of the decomposition. VMD decomposition adaptively separates the local sequence into different modes representing long-term degradation trends, medium-term fluctuation patterns, and short-term noise interference, achieving a multi-scale refined representation of the non-stationary dynamics of battery degradation. The introduced normalized location encoding compensates for the limitations of the sliding window's field of view and supplements the battery's progress information throughout its global lifetime. The resulting enhanced feature vector combines local evolution details with key information about the global degradation stage, providing robust and information-rich input for subsequent spatiotemporal models.

[0038] Preferably, the graph topology structure constructed in step S5 includes:

[0039] S51: Define each feature index in the feature set obtained in step S2 as a node in the graph network;

[0040] S52: Calculate the Pearson correlation coefficient between the features corresponding to any two nodes. When the absolute value of the Pearson correlation coefficient is greater than a preset correlation threshold, establish a connection edge between the two nodes and use the Pearson correlation coefficient as the weight of the connection edge. When the absolute value of the Pearson correlation coefficient is not greater than the preset correlation threshold, do not establish a connection edge between the two nodes and the corresponding weight is 0, thus obtaining the weighted adjacency matrix.

[0041] By constructing a graph topology based on the Pearson correlation coefficient, the intrinsic physical relationships between multi-source feature indicators are explicitly modeled in the form of a weighted adjacency matrix. Nodes represent each feature indicator, and weighted edges quantify the synergistic or coupling strength of feature degradation. This graph structure not only naturally expresses the complex spatial relationships of different features influencing each other during battery degradation, providing a topological foundation for subsequent graph neural network models to capture information transmission and dependencies between features, but also highlights the structured information flow between strongly correlated features through reasonable pruning of weakly correlated or uncorrelated relationships. This enhances the model's ability to learn the synergistic evolution of multi-dimensional degradation patterns in battery health.

[0042] Preferably, the input tensor dimension of the graph attention network-gated recurrent unit (GAT-GRU) deep learning prediction model described in step S6 is (Batch, Seq, Nodes, Fea), where: Batch is the batch size, Seq is the time step of the sliding window, Nodes is the number of nodes in the graph network, and Fea is the dimension of the augmented feature vector corresponding to each node.

[0043] By constructing a high-dimensional input tensor, the feature-enhanced multivariate time-series data is highly structured and represented within a unified feature space. This enables the model to efficiently extract information from multiple dimensions simultaneously: the time dimension (Seq) captures the dynamic evolution of features, the spatial dimension (Nodes) models the topological relationships between multiple features through a graph structure, and the feature dimension (Fea) aggregates the rich information from multi-scale decomposition and location encoding at each feature point. This tensor structure allows the model to fully explore the complex spatiotemporal correlations and patterns hidden in the battery degradation process in a collaborative and parallel manner, providing crucial underlying data structure support for achieving accurate and robust SOH prediction.

[0044] Preferably, the processing procedure of the spatial attention aggregation layer in step S6 includes:

[0045] S61: For each time step Input node feature set Through a learnable weight matrix Perform a linear transformation to reduce the feature dimension. Mapping to hidden layer dimension , where N is the number of feature nodes;

[0046] S62: Introducing a weighted adjacency matrix As a mask, calculate the current node. with neighboring nodes Attention coefficient at the current time step :

[0047]

[0048] in, For activation function, Let W be the attention vector, and W be the learnable weight matrix. For splicing operations;

[0049] S63: Use the Softmax function to adjust the attention coefficients. Normalization is performed to obtain the attention weights. The features of neighboring nodes are then weighted and aggregated based on the attention weights to obtain a node representation that incorporates spatial information. :

[0050]

[0051] in, Let i be the set of neighboring nodes of node i;

[0052] S64: Concatenate all node representations aggregated at each time step to form the comprehensive spatial feature vector for each time step. .

[0053] The spatial attention aggregation layer constructed in this invention can dynamically assign importance weights between each feature node and its different neighboring nodes. This process not only effectively utilizes the prior graph topology constructed by Pearson correlation coefficients as a constraint and guide for information transmission, but also adaptively adjusts the influence intensity between nodes through an end-to-end learning mechanism. This design enables the model to selectively focus on key information flows from complex feature association networks, deeply mine and fuse the dynamic spatial dependencies between features, thereby accurately capturing the internal physical mechanism of multi-dimensional feature co-evolution during battery degradation, and significantly enhancing the model's ability to represent and understand the complex spatial states of battery systems.

[0054] Preferably, in step S62, for any two feature nodes, if the corresponding connection weight in the weighted adjacency matrix is ​​0, then when calculating the attention coefficient, the attention coefficient is forcibly set to negative infinity to ensure that the attention mechanism only takes effect between feature nodes with strong correlation.

[0055] By forcing the attention coefficients corresponding to connections with zero weights in the weighted adjacency matrix to negative infinity during the computation of the spatial attention aggregation layer, a priori physical correlation knowledge and an adaptive attention mechanism are organically integrated. This constraint ensures that the model's information propagation path strictly follows the graph topology structure composed of strongly correlated features. This effectively avoids noise interference and invalid information transmission that may occur between irrelevant or weakly correlated features, and guides the model to concentrate computational resources on mining deep interaction patterns between feature nodes that have been verified to have significant correlations. This enhances the interpretability and modeling accuracy of the attention mechanism, enabling the model to focus more on capturing the key collaborative relationships that truly exist in the battery degradation process.

[0056] Preferably, the temporal memory evolution layer described in step S6 achieves state updates through reset gates and update gates; wherein the reset gate Determine how much historical information to discard to adapt to sudden changes such as capacity regeneration; update gates The decision is made on how much of the historical state to retain in order to maintain a long-term downward trend; the final state update is as follows:

[0057]

[0058]

[0059] in, To update the learnable weight matrix of the gate, This is the comprehensive spatial feature vector output by the spatial attention aggregation layer. This is the hidden state from the previous moment. Let σ be the hidden state at the current time step, and σ be the Sigmoid activation function. Let be the learnable weight matrix of the candidate hidden states. The tanh activation function is used. This is element-wise multiplication.

[0060] The temporal memory evolution layer constructed in this invention achieves intelligent modeling and adaptive tracking of non-stationary temporal dynamics during battery degradation through reset and update gate mechanisms. The reset gate sensitively captures and responds to short-term abrupt changes or mode switching, such as capacity regeneration, thereby flexibly adjusting the degree of forgetting historical information. The update gate, on the other hand, focuses on maintaining the consistency of long-term degradation trends, stably representing the slow evolutionary pattern of overall battery degradation through long-term memory of historical states. This synergistic gating mechanism enables the model to simultaneously consider both stable tracking of long-term trends and sensitive response to short-term dynamics. Thus, in the complex and ever-changing temporal evolution, it accurately captures the gradual and abrupt changes in battery health status, significantly enhancing the robust representation and prediction capabilities for various degradation behaviors throughout the entire lifespan.

[0061] Preferably, the training process of the GAT-GRU deep learning prediction model in step S7 includes: dividing the battery data into a training set and a test set; using the mean squared error (MSE) as the loss function, training the GAT-GRU deep learning prediction model based on the training set, and using the test set to verify the model performance.

[0062] This invention employs mean squared error (MSE) as the loss function during the model training phase. By minimizing the difference between predicted and actual values, it drives the model to learn battery degradation mapping patterns from the training set. Combined with a strategy for partitioning the training and test sets, this ensures the model can fully learn complex spatiotemporal correlation patterns while effectively avoiding over-reliance on training data, thereby improving its generalization performance and robustness. Ultimately, this training scheme enables the trained model to accurately capture the evolutionary trend of battery health status, exhibiting good generalization ability and stability.

[0063] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. To address the distortion problem at the end of VMD decomposition caused by limited sliding window data in online prediction, this invention adopts a sliding window VMD strategy based on mirror extension. This strategy eliminates Gibbs oscillations generated at data boundaries by performing symmetrical extension at prediction points, thereby enabling online frequency domain decomposition based on short windows to achieve accuracy similar to offline full-cycle decomposition, thus providing distortion-free and high-fidelity trend and fluctuation features for subsequent modeling; 2. To address the problem of unstable prediction caused by abnormal sensor data or environmental noise interference, this invention dynamically constructs a graph topology based on Pearson correlation coefficient and introduces a masked graph attention mechanism, enabling the model to aggregate information based on the real-time correlation strength between features. When a single feature is abnormal, the model can automatically reduce its attention weight and use other strongly correlated features for correction and supplementation, realizing active mutual verification and noise shielding between features, greatly improving the prediction robustness under complex working conditions; 3. To address the problem of battery aging... To address the challenge of accurately capturing the "long-term smooth decay" and "short-term mutations" in battery aging, this invention uses Virtual Mode Decomposition (VMD) to physically decompose the original signal into modal components representing different time scales, and designs a time-series memory evolution layer composed of GRU units. The low-frequency trend term separated by VMD works in conjunction with the update gate of the GRU to focus on learning and tracking long-term aging trends; while the second-highest frequency fluctuation term works in conjunction with the reset gate of the GRU to sensitively perceive and adapt to local mutations such as capacity regeneration. This combination of "decomposition-memory" dual mechanisms effectively solves the problem of prediction lag at inflection points in traditional methods. Furthermore, to address the issue that fixed graph structures cannot reflect the dynamic changes in feature correlation during battery aging, this invention abandons the fixed adjacency matrix and instead dynamically generates and updates the connection edges and weights of the graph network based on the real-time calculated Pearson correlation coefficient between features. This design allows the graph topology to adaptively adjust to different aging stages of the battery, thereby more accurately modeling the dynamic coupling relationship between features and SOH as lifespan evolves, improving the model's adaptability and prediction accuracy throughout its entire lifespan. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0065] Figure 2 This is a schematic diagram of the feature enhancement process of the present invention;

[0066] Figure 3 This is a schematic diagram of the GAT-GRU model structure of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0068] This invention provides a battery SOH prediction method that integrates signal physical denoising and spatiotemporal correlation modeling, such as... Figure 1 As shown, this method achieves highly robust battery health state prediction through a dual mechanism of "signal-level physical filtering" and "model-level feature verification". The specific steps are as follows:

[0069] Step 1: Acquisition and Feature Extraction of Battery Aging Data

[0070] Data Acquisition: Charge-discharge cycle tests were performed on the lithium-ion battery, and basic data were collected during each cycle, including voltage ( ), current ( ),temperature( ) and time ( ).

[0071] Initial Feature Selection: Based on the battery aging mechanism (such as increased internal resistance and loss of active material), raw feature indicators (FIs) are extracted from the voltage, current, and temperature curves. The candidate feature indicators extracted in this embodiment include, but are not limited to:

[0072] Voltage characteristics: Slope of constant current charging voltage curve ( ), voltage curve integral during constant current charging stage ( Constant voltage charging time ( ), Incremental Capacity Analysis (ICA) curve peak ( ).

[0073] Current characteristics: Integral of charging current curve ( ), Discharge current curve integral ( Constant current charging time ( ).

[0074] Thermodynamic properties: maximum temperature rise rate, average temperature rise over cycles, time to peak temperature ( )wait.

[0075] Integrating the charging current curve (Taking the amount of charge that the battery can hold in the current cycle) as an example, the calculation method is as follows:

[0076]

[0077] in, This represents the number of data sampling points for this loop. The sampling interval is... For the first The current value at each sampling point.

[0078] Label definition: The discharge capacity of each charge-discharge cycle is used as the true SOH value (or maximum remaining capacity) for that cycle. ).

[0079] Step 2: Feature Optimization Based on Correlation Analysis

[0080] To construct a feature space with redundancy verification capability, the Pearson correlation coefficient (PCC) is used to screen candidate features.

[0081] Calculate each candidate feature sequence With the true capacity sequence PCC values ​​between:

[0082]

[0083] in, For covariance, Standard deviation This is the mean.

[0084] Screening strategy: Set a relevance threshold (in this example, it is set to 1). Features highly correlated with SOH evolution were selected. Example selection: Seven feature nodes were ultimately selected, including the slope of the constant current charging voltage curve. Peak value of ICA curve Integral charging current curve Constant current charging time Constant current charging voltage integral Time to reach peak temperature and the capacity value at the previous time step. .

[0085] Although the selected features have different physical meanings (voltage, current, temperature), they are all highly correlated with SOH. This combination of "multi-source heterogeneous" features provides a data foundation for subsequent efforts to combat single-sensor noise.

[0086] Step 3: Data cleaning based on Hampel filtering

[0087] To address the burst noise and outliers in the collected data, and to preserve the true physical phenomenon of battery "capacity regeneration," this invention employs a Hampel filter for cleaning.

[0088] For length of The feature sequence, with the sliding window length set to... (For example, take) ).

[0089] Calculate the median within the window and absolute median :

[0090]

[0091]

[0092] Calculate the standard deviation estimate .

[0093] Anomaly detection and handling: If Then determine It is an outlier, and is used as an exception. Replace this point. This method only eliminates transient measurement noise while retaining the capacity regeneration fluctuation characteristics that have a certain degree of persistence.

[0094] Step 4: Online multi-scale frequency domain feature enhancement (first layer of noise reduction mechanism)

[0095] This invention introduces variational mode decomposition (VMD) technology to enhance features, such as... Figure 2 As shown, VMD, as an adaptive non-recursive signal processing method, has excellent frequency domain separation capabilities and can effectively decompose complex signals adaptively into trend, fluctuation, and noise components. However, the VMD algorithm is essentially based on frequency domain analysis, and it has inherent defects in signal processing: when the processed data length is short and the two ends are not closed, Gibbs effect will occur at the signal ends, causing severe spurious oscillations and divergence in the decomposed modes at the endpoints. Especially in the online SOH prediction task targeted by this invention, in order to meet real-time requirements, the input data must be a short sequence of a finite-length "sliding window". The data at the very end of the window (current time) The endpoint divergence at the endpoints is the most critical information source for predicting future capacity, and this is precisely the region most severely affected by the endpoint effect. If VMD decomposition is directly performed on the sliding window data, the divergence at the endpoints will directly lead to severe distortion in the feature extraction at the current moment, introducing human error. To address this technical contradiction, this invention designs the following online sliding window VMD decomposition strategy based on mirror extension, as follows:

[0096] Sliding window construction: at the prediction time The construction length is Historical feature window sequence In this embodiment, the window length is... Set it to 16.

[0097] Mirror extension processing: In order to suppress VMD at the end of the window (i.e., the latest time). The boundary effect of ) Using the axis of symmetry, the data at the end of the window is mirrored and then concatenated. The extended sequence... Length is Its formula is expressed as:

[0098]

[0099] in For the extension length (in this embodiment, it is taken as...) This operation ensures a smooth transition at the endpoints of the sequence to be decomposed, avoiding the Gibbs phenomenon.

[0100] Variational Mode Decomposition (VMD): for extended sequences Perform VMD decomposition and set the number of modes. Punishment factor The decomposition yields three intrinsic mode functions (IMFs):

[0101] Low-frequency trend term, reflecting the long-term decline trajectory of SOH.

[0102] The second-lowest frequency fluctuation term reflects the capacity regeneration phenomenon.

[0103] High-frequency noise term, reflecting random interference.

[0104] Location coding and fusion:

[0105] Position encoding calculation: To supplement the global lifetime process information lost by the sliding window, normalized position encoding is introduced. :

[0106]

[0107] in It is the maximum lifetime prior value based on the statistics of the training set (e.g., 168 for the NASA dataset). This encoding provides the model with the current "lifetime progress bar".

[0108] Residual fusion: To prevent the loss of original information during the decomposition process, the concept of "residual connections" is used to construct enhanced features. For time step... For each feature node, its final input vector The dimension is 5, and the structure is as follows:

[0109]

[0110] Step 5: Graph Construction

[0111] This invention models the SOH prediction problem as a spatiotemporal prediction problem on a graph structure and constructs a multi-feature interconnected topology. First, the graph structure is constructed:

[0112] Node definition: Define the 7 feature indicators selected in step two as nodes in the graph network. .

[0113] Adjacency Matrix generate:

[0114]

[0115] Edge and Adjacency Matrix Definition: Using a weighted adjacency matrix Describe the topological relationships between nodes, and the correlation between two features. If the value is 0.8, then a connection is established.

[0116] This highly connected graph structure means that any feature node is connected to multiple other "ally" nodes, providing a pathway for correcting its own deviations through neighboring nodes.

[0117] Step Six: Spatiotemporal Graph Attention Network (GAT-GRU) Prediction Model (Secondary Noise Reduction Mechanism)

[0118] This step constructs an end-to-end deep learning model that integrates Graph Attention Network (GAT) and Gated Recurrent Unit (GRU), enabling it to simultaneously capture the spatial coupling relationships and temporal evolution patterns between features. Specifically, as follows... Figure 3 As shown, the implementation process consists of four specific sub-steps: data tensor construction, spatial feature aggregation, temporal evolution modeling, and regression prediction.

[0119] Input Tensor Construction: The enhanced feature data obtained in step four is organized into a four-dimensional tensor format and used as the network input. Tensor dimension is defined as ,in: Batch size, taken as 32 in this example; The time step of the sliding window, corresponding to the aforementioned sliding window length. In this embodiment, 16 is used; The number of graph nodes corresponds to the 7 preferred features; The feature dimension of each node corresponds to the 5-dimensional enhanced feature (original value + trend + fluctuation + noise + location encoding) constructed in step four.

[0120] Spatial attention aggregation layer with time-by-time feature mutual verification: This layer is located before temporal processing and is designed to perform mutual verification of features at each time step. ( Spatial convolution is performed on the feature maps to achieve feature cross-validation. This process shares parameters on the time axis and can be processed in parallel or cyclically.

[0121] Linear transformation: for time step Input node feature set First, through a learnable weight matrix Perform a linear mapping to change the feature dimension from Mapping to hidden layer dimension .

[0122] Attention coefficient calculation (introducing adjacency matrix constraints): Calculate the node with neighboring nodes Attention coefficient at the current moment The formula is as follows:

[0123]

[0124] in For attention vectors, This is a concatenation operation. Key implementation point (Mask operation): In this step, the adjacency matrix generated in step five is introduced. As a mask. If Then force setting This operation ensures that the attention mechanism only works between strongly related feature nodes, physically blocking interference from irrelevant features.

[0125] Dynamic weighting and aggregation: Attention weights are obtained by normalization using the Softmax function. The node representation, which incorporates spatial information, is obtained by weighted summation of neighbor features. :

[0126]

[0127] This method can achieve feature mutual verification. When a feature node (such as temperature) undergoes a sudden change due to noise, its feature vector... Distortion will occur, affecting its correlation score with normal nodes. The decrease, in turn, reduces the weight assigned to it. Automatically decreases.

[0128] Output processing: All node features aggregated at this time are concatenated (Flattened) or pooled to form the comprehensive spatial feature vector at this time. Repeat this process for the 16 time steps within the window to obtain the sequence. .

[0129] Temporal Memory Evolutionary Layer (GRU Layer): This layer processes the spatial feature sequence output by the GAT layer. Input the data into a GRU network to capture the long-term decline trend of SOH. Perform the following steps in the GRU network:

[0130] Looping update: For each time step in the sequence (from arrive The GRU unit receives the spatial characteristics at the current time. and the hidden state of the previous moment Update the state through the gating mechanism:

[0131] Reset door : Decide how much historical information to discard to accommodate sudden changes such as capacity regeneration.

[0132] Update Gate : Decide how much of the historical state to retain in order to sustain a long-term downward trend.

[0133] Status Update:

[0134]

[0135]

[0136] Output: After After the last iteration, output the hidden state of the last time step. (or This vector encapsulates the spatiotemporal information within the entire time window.

[0137] Predictive output layer: Predicts the final output of the GRU. The input is fed into the regression layer. It consists of two fully connected network (MLP) layers, with a ReLU activation function and a Dropout layer (e.g., Dropout=0.2) in between to prevent overfitting. The output layer is a single neuron with no activation function (or linear activation), outputting a scalar value. This is the predicted SOH value at the current moment.

[0138] Step 7: Model Training and Prediction

[0139] The model is trained and predicted through the following steps: 1) Dataset partitioning: The dataset is divided into training and test sets. 2) Loss function: Mean squared error (MSE) is used as the loss function. 3) Online prediction: Battery data is collected in real time, processed through steps one to six, and then input into the trained GAT-GRU model.

Claims

1. A battery SOH prediction method integrating signal physical denoising and spatiotemporal correlation modeling, characterized in that, Includes the following steps: Step S1: Collect charge and discharge monitoring data of the lithium-ion battery throughout its entire life cycle, and extract original health characteristic indicators from the charge and discharge monitoring data that reflect the increase in internal polarization resistance, loss of active material and thermodynamic properties of the battery. Step S2: The original health characteristic indicators are screened using the Pearson correlation coefficient PCC analysis method, and the features with an absolute value of Pearson correlation coefficient with battery capacity that is greater than a preset threshold are retained to obtain a feature set; Step S3: Use the Hampel filter to clean the feature sequences corresponding to the feature set, and remove bursty outliers from the feature sequences; Step S4: Construct a sliding window variational mode decomposition (VMD) model based on mirror extension, perform online multi-scale frequency domain enhancement on the cleaned feature sequence, separate the trend term, fluctuation term and noise term, and generate enhanced feature vectors by combining normalized position encoding; Step S5: Construct a graph topology based on the correlation strength between features in the feature set, and generate the corresponding weighted adjacency matrix; Step S6: Construct a graph attention network-gated recurrent unit (GAT-GRU) deep learning prediction model containing a spatial attention aggregation layer and a temporal memory evolution layer. Perform spatiotemporal fusion on the enhanced feature vector through the spatial attention aggregation layer and the temporal memory evolution layer, and output the SOH prediction value. Step S7: After processing the real-time collected battery data through steps S1 to S6, input it into the trained GAT-GRU deep learning prediction model for online inference to obtain the current SOH estimate of the battery.

2. The method according to claim 1, characterized in that, The original health characteristic indicators mentioned in step S1 include voltage characteristic, current characteristic and thermodynamic characteristic; the voltage characteristic includes the slope of the constant current charging voltage curve, the integral of the voltage curve during the constant current charging stage, the constant current charging time, and the peak value of the incremental capacity analysis ICA curve; the current characteristic includes the integral of the charging current curve; the thermodynamic characteristic includes the time to reach the temperature peak.

3. The method according to claim 1, characterized in that, Step S3, which involves using a Hampel filter to clean the feature sequences corresponding to the feature set, includes: S31: For a feature sequence of length N, set the sliding window length to be... ; S32: Calculate the median of the feature data within the sliding window. and absolute median : ; ;in, For feature data within the sliding window, This is a function for calculating the median; S33: Calculate the standard deviation estimate ; S34: If a data point within the sliding window is... The absolute value of the difference is greater than That is, the data point is determined to be an outlier, and is then... Replace the outlier value.

4. The method according to claim 1, characterized in that, Step S4, which involves constructing a sliding window variational mode decomposition (VMD) model based on mirror extension, includes: S41: Sliding window construction: At prediction time t, extract a historical feature window sequence of length W. ; S42: Mirror continuation processing: using the feature data corresponding to the predicted time t Using the axis of symmetry, the data at the end of the historical feature window sequence is mirrored and spliced ​​to the historical feature window sequence to generate an extended sequence. : ;in, For the extension length; S43: VMD Decomposition: Perform variational mode decomposition on the extended sequence to obtain the low-frequency trend mode. Secondary low frequency fluctuation mode and high-frequency noise modes ; S44: Feature Fusion: Calculate Normalized Position Encoding This is used to supplement the global lifetime process information lost by the sliding window: ;in, This indicates the current cycle number of the battery. It is a priori value of the maximum battery life based on statistics from the training set; The original observations of the historical feature window sequence, the low-frequency trend mode, the second-lowest frequency fluctuation mode, the high-frequency noise mode, and the normalized position code are concatenated to generate an enhanced feature vector. : 。 5. The method according to claim 1, characterized in that, The graph topology structure described in step S5 includes: S51: Define each feature index in the feature set obtained in step S2 as a node in the graph network; S52: Calculate the Pearson correlation coefficient between the features corresponding to any two nodes. When the absolute value of the Pearson correlation coefficient is greater than a preset correlation threshold, establish a connection edge between the two nodes and use the Pearson correlation coefficient as the weight of the connection edge. When the absolute value of the Pearson correlation coefficient is not greater than the preset correlation threshold, do not establish a connection edge between the two nodes and the corresponding weight is 0, thus obtaining the weighted adjacency matrix.

6. The method according to claim 1, characterized in that, The input tensor dimension of the graph attention network-gated recurrent unit (GAT-GRU) deep learning prediction model described in step S6 is (Batch, Seq, Nodes, Fea), where: Batch is the batch size, Seq is the time step of the sliding window, Nodes is the number of nodes in the graph network, and Fea is the dimension of the augmented feature vector corresponding to each node.

7. The method according to claim 1, characterized in that, The processing procedure of the spatial attention aggregation layer in step S6 includes: S61: For each time step Input node feature set Through a learnable weight matrix Perform a linear transformation to reduce the feature dimension. Mapping to hidden layer dimension , where N is the number of feature nodes; S62: Introducing a weighted adjacency matrix As a mask, calculate the current node. with neighboring nodes Attention coefficient at the current time step : ;in, For activation function, Let W be the attention vector, and W be the learnable weight matrix. For splicing operations; S63: Use the Softmax function to adjust the attention coefficients. Normalization is performed to obtain the attention weights. The features of neighboring nodes are then weighted and aggregated based on the attention weights to obtain a node representation that incorporates spatial information. : ;in, Let i be the set of neighboring nodes of node i; S64: Concatenate all node representations aggregated at each time step to form the comprehensive spatial feature vector for each time step. .

8. The method according to claim 7, characterized in that, In step S62, for any two feature nodes, if the corresponding connection weight in the weighted adjacency matrix is ​​0, then when calculating the attention coefficient, the attention coefficient is forcibly set to negative infinity to ensure that the attention mechanism only takes effect between feature nodes with strong correlation.

9. The method according to claim 1, characterized in that, The temporal memory evolution layer described in step S6 achieves state updates through reset gates and update gates; wherein the reset gate Determine how much historical information to discard to adapt to sudden changes such as capacity regeneration; update gates The decision is made on how much of the historical state to retain in order to maintain a long-term downward trend; the final state update is as follows: ; ;in, To update the learnable weight matrix of the gate, This is the comprehensive spatial feature vector output by the spatial attention aggregation layer. This is the hidden state from the previous moment. Let σ be the hidden state at the current time step, and σ be the Sigmoid activation function. Let be the learnable weight matrix of the candidate hidden states. The tanh activation function is used. This is element-wise multiplication.

10. The method according to claim 1, characterized in that, The training process of the GAT-GRU deep learning prediction model described in step S7 includes: dividing the battery data into a training set and a test set; using the mean squared error (MSE) as the loss function, training the GAT-GRU deep learning prediction model based on the training set, and verifying the model performance using the test set.

Images