A lithium battery soh prediction method and system based on dual-source channel enhancement
Patent Information
- Application Number
- CN202610973863.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-01
- Publication Date
- 2026-10-02
AI Technical Summary
[0005]本申请针对现有锂电池SOH预测方法中多变量特征贡献差异难以显式表征、关键退化信息难以动态强化以及输出层线性映射难以拟合复杂非线性退化轨迹的技术问题，提供一种基于双源通道增强的锂电池SOH预测方法及系统
本申请提供一种基于双源通道增强的锂电池SOH预测方法，通过反转嵌入层将多变量退化时序数据中的各变量对应的历史时间序列独立编码为变量级表示，改变了传统Transformer以时间步为基本建模单元的模式，使得各健康因子变量的身份信息和语义差异得以保留；在此基础上，通过双源通道增强模块引入两组可学习通道嵌入，对变量级表示进行门控融合增强，其中第一组通道嵌入生成先验门控权重对原始变量级表示进行自适应预调制，第二组通道嵌入经动态门控融合机制实现不同通道语义信息的自适应加权分配，有效解决了现有技术对多变量贡献差异及退化过程中变量重要性动态变化缺乏显式刻画的问题；同时，通过非线性预测头替代传统单层线性投影结构，增强了模型从高维特征空间到未来退化趋势的非线性映射能力。上述技术方案相互配合、协同作用，显著提升了对复杂非线性退化轨迹的表征能力与预测精度，有效避免了关键健康因子贡献被平均化以及输出端映射能力不足导致预测精度受限的问题。
Smart Images
Figure CN122860108A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of lithium battery life assessment technology, specifically to a lithium battery SOH prediction method and system based on dual-source channel enhancement. Background Technology
[0002] Lithium-ion batteries, due to their high energy density, long cycle life, and low self-discharge rate, have been widely used in various devices such as electric vehicles, smartphones, and energy storage systems. With increasing cycle count, lithium batteries exhibit aging phenomena such as capacity decay and increased internal resistance, manifesting as a gradual decline in State of Health (SOH), ultimately leading to a shortened Remaining Useful Life (RUL). Accurately predicting the SOH of lithium batteries is crucial for ensuring equipment operational safety and developing reasonable maintenance plans. Currently, lithium battery life prediction methods are mainly divided into two categories: mechanistic modeling-based methods and data-driven methods. Mechanism-based modeling methods rely on the battery's internal electrochemical mechanisms and experimental priors, offering a degree of interpretability. However, due to the complex operating environment of lithium batteries and the highly nonlinear degradation process, these methods often struggle to achieve high-precision modeling under real-world conditions. In contrast, data-driven methods can directly learn degradation patterns from historical operating data, and therefore have attracted widespread attention.
[0003] In data-driven approaches, Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTMs) were among the first to be applied to predict the remaining lifespan of lithium-ion batteries. However, methods such as RNNs and LSTMs still have limitations when dealing with high-dimensional multivariate coupling relationships and complex global dependency modeling, making it difficult to fully capture the long-range dependencies between variables. In recent years, Transformer models have been introduced into lithium-ion battery lifespan prediction tasks due to their excellent long-sequence dependency modeling capabilities. Existing research has shown that Transformer and its variants have great potential in modeling battery degradation trends and predicting battery lifespan. However, existing Transformer-based methods still have limited ability to directly model multivariate degradation information, specifically in the following ways: First, they lack explicit characterization of the differences in the contributions of variables. During lithium battery degradation, multiple features such as voltage, current, temperature, internal resistance, and capacity constitute multivariate time series data. Different variables have different characterization abilities for battery degradation states, while existing methods use a uniform modeling approach for different features, ignoring the differential contributions of each variable in the battery degradation process. Second, they lack the ability to dynamically enhance key degradation information. The battery degradation process has significant dynamic characteristics, and the importance of key features may change in different degradation stages. For example, voltage fluctuations may be more sensitive in the early degradation stage, while internal resistance growth may become the dominant indicator in the late degradation stage. Existing methods lack effective characterization of this dynamic characteristic and cannot adaptively adjust the degree of attention to different variables according to the degradation stage. Third, the output layer's expressive power is insufficient to fit complex degradation trajectories. The degradation process of lithium-ion batteries typically exhibits significant nonlinearity, stages, and complex coupling characteristics. Existing methods employ a single-layer linear projection structure at the output end, which makes it difficult to fully characterize the complex mapping relationship between high-dimensional encoded features and future degradation trends.
[0004] The aforementioned shortcomings limit the accuracy of existing methods in predicting the state of health (SOH) of lithium batteries, making it difficult to meet the high-precision requirements for health management in practical applications. Summary of the Invention
[0005] This application addresses the technical problems in existing lithium battery SOH prediction methods, such as the difficulty in explicitly characterizing the differences in multivariate feature contributions, the difficulty in dynamically enhancing key degradation information, and the difficulty in fitting complex nonlinear degradation trajectories with linear mapping of the output layer. It provides a lithium battery SOH prediction method and system based on dual-source channel enhancement.
[0006] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, this application provides a lithium battery SOH prediction method based on dual-source channel enhancement, comprising the following steps: S1, acquire multivariate degradation time series data of lithium battery within a historical time window, input the multivariate degradation time series data into the inverted embedding layer, and independently embed the historical time series corresponding to each variable in the multivariate degradation time series data into a variable-level representation to obtain a variable-level embedding representation; S2, the variable-level embedding representation is input into the dual-source channel enhancement module, and the variable-level embedding representation is enhanced by gating fusion through two sets of learnable channel embeddings to obtain the channel-enhanced representation; S3, the channel enhancement representation is input into the encoder, processed by a multi-head attention mechanism, to establish the global dependency relationship between variables in the multivariate degenerate time series data, and to obtain the encoded feature representation; S4. The encoded feature representation is input into the nonlinear prediction head, and after nonlinear mapping processing, the SOH prediction value of the lithium battery within the future prediction window is obtained.
[0007] Furthermore, the multivariate degradation time series data in S1 includes the monitored values of at least two variables among voltage, current, temperature, internal resistance, and capacity at each historical time step.
[0008] Furthermore, the process of embedding the historical time series corresponding to each variable in the multivariate degenerate time series data into a variable-level representation independently in S1, to obtain the variable-level embedded representation, specifically includes: Each variable in the multivariate degraded time series data is treated as an independent variable-level token, resulting in N variable-level tokens, where N is the number of variables in the multivariate degraded time series data. Each of the variable-level tokens is projected onto a D-dimensional latent space through a linear mapping to obtain the variable-level embedding representation.
[0009] Furthermore, the process in S2 specifically includes: Channel gating weights are generated by the first set of learnable channel embeddings, and the variable-level embedding representation is pre-modulated using the channel gating weights to obtain the channel-premodulated feature representation. Adaptive gating fusion is performed on the second set of learnable channel embeddings and the channel premodulated feature representations based on the channel premodulated feature representations to obtain the fused channel enhancement information; The fused channel enhancement information is superimposed onto the variable-level embedding representation in a residual manner to obtain the channel enhancement representation.
[0010] Furthermore, the step of generating channel gating weights through the first set of learnable channel embeddings specifically includes: Obtain the first set of learnable channel embedding parameters, and perform max pooling and average pooling on the first set of learnable channel embedding parameters in the variable dimension to obtain the global max pooling result and the global average pooling result. The global max pooling result and the global average pooling result are respectively input into independent linear mapping layers. The outputs of the two linear mapping layers are summed and then activated by the Sigmoid function to generate recalibrated weights as the channel gating weights.
[0011] Further, the adaptive gating fusion of the second set of learnable channel embeddings and the channel premodulated feature representations based on the channel premodulated feature representations to obtain the fused channel enhancement information specifically includes: The pre-modulated feature representation of the channel is input into the first gated mapping layer for nonlinear transformation to obtain the semantic representation of the intermediate channel. The semantic representation of the intermediate channel is input into the second gating mapping layer and activated by the Sigmoid function to generate dynamic gating weights. The channel information after nonlinear mapping is multiplied by the dynamic gating weight to obtain the first weighted channel representation; The second set of learnable channels is embedded and multiplied by the dynamic gating weights to obtain the second weighted channel representation; The first weighted channel representation is added to the second weighted channel representation to obtain the fused channel enhancement information.
[0012] Furthermore, the encoder in S3 includes multiple encoder layers, each encoder layer including a multi-head attention sublayer, a first residual connection and layer normalization sublayer, a feedforward network sublayer, and a second residual connection and layer normalization sublayer.
[0013] Furthermore, in S4, the nonlinear prediction head includes a first linear transformation layer, a first GELU activation function, a second linear transformation layer, a second GELU activation function, and an output linear layer connected in sequence. The nonlinear mapping process yields the predicted SOH value of the lithium battery within the future prediction window, specifically including: The encoded feature representation is input into the first linear transform layer and nonlinearly mapped with the first GELU activation function to obtain the first intermediate prediction feature; The first intermediate prediction feature is input into the second linear transformation layer and nonlinearly mapped with the second GELU activation function to obtain the second intermediate prediction feature; The second intermediate prediction feature is input into the output linear layer and mapped to the target prediction step space to obtain the SOH prediction value.
[0014] Furthermore, the feedforward network sublayer includes two linear transformation units and a nonlinear activation function located between the two linear transformation units.
[0015] Secondly, this application also provides a lithium battery SOH prediction system based on dual-source channel enhancement, comprising: The inversion embedding module is used to acquire multivariate degradation time series data of lithium batteries within a historical time window, input the multivariate degradation time series data into the inversion embedding layer, and independently embed the historical time series corresponding to each variable in the multivariate degradation time series data into a variable-level representation to obtain a variable-level embedding representation; A dual-source channel enhancement module is used to input the variable-level embedding representation into the dual-source channel enhancement module, and perform gated fusion enhancement on the variable-level embedding representation through two sets of learnable channel embeddings to obtain a channel-enhanced representation. The encoder module is used to input the channel enhancement representation into the encoder, process it through a multi-head attention mechanism, establish the global dependency relationship between variables in the multivariate degenerate time series data, and obtain the encoded feature representation. The nonlinear prediction head module is used to input the encoded feature representation into the nonlinear prediction head, and after nonlinear mapping processing, obtain the SOH prediction value of the lithium battery within the future prediction window.
[0016] Compared with the prior art, this application has the following beneficial effects: This application provides a lithium battery SOH prediction method based on dual-source channel enhancement. By reversing the embedding layer, the historical time series corresponding to each variable in the multivariate degradation time series data are independently encoded into variable-level representations. This changes the traditional Transformer model, which uses time steps as the basic modeling unit, thus preserving the identity information and semantic differences of each health factor variable. Based on this, a dual-source channel enhancement module introduces two sets of learnable channel embeddings to perform gated fusion enhancement on the variable-level representations. The first set of channel embeddings generates prior gating weights to adaptively premodulate the original variable-level representations, while the second set of channel embeddings uses a dynamic gating fusion mechanism to adaptively weight the semantic information of different channels. This effectively solves the problem of existing technologies lacking explicit characterization of the differences in multivariate contributions and the dynamic changes in variable importance during degradation. Simultaneously, a nonlinear prediction head replaces the traditional single-layer linear projection structure, enhancing the model's nonlinear mapping capability from high-dimensional feature space to future degradation trends. These technical solutions work together synergistically to significantly improve the representation ability and prediction accuracy of complex nonlinear degradation trajectories, effectively avoiding the problems of averaging of key health factor contributions and insufficient output mapping capabilities that limit prediction accuracy.
[0017] Furthermore, max pooling and average pooling are performed on the variable dimension using the first set of learnable channel embeddings. The pooling results are then fed into two independent linear mapping layers and recalibrated using the Sigmoid function to generate recalibrated weights as channel gating weights. Compared to directly using fixed channel weights or single-path pooling, this approach can simultaneously capture the most significant feature responses and overall distribution features in the channel embeddings, enhancing the expressive power and robustness of the channel gating weights. Building upon this, dynamic gating weights are generated by sequentially passing the pre-modulated feature representations of the channels through two gating mapping layers. These dynamic gating weights are then used to weight and fuse the nonlinearly mapped channel information and the second set of learnable channel embeddings. This allows the model to dynamically adjust the contribution weights of each variable channel based on different input samples and different degradation stages. Compared to existing methods that only use fixed channel embeddings or unified attention mechanisms, this approach effectively characterizes the phased changes in the importance of variables during battery degradation, enhancing the model's adaptability and ability to filter key degradation information in complex degradation scenarios.
[0018] By performing two layers of nonlinear mapping on the encoded feature representation output by the encoder, first through a linear transformation layer and the GELU activation function, and second through a linear transformation layer and the GELU activation function, and then mapping it to the target prediction step space through the output linear layer, compared with the output structure of the existing iTransformer model which only uses a single-layer linear projection, richer high-order nonlinear features can be extracted in the prediction stage. This allows the model to more fully fit the complex nonlinear relationships and stage characteristics in the lithium-ion battery life degradation process, effectively making up for the lack of expressive power of single-layer linear mapping in complex prediction tasks, thus providing more accurate evaluation results for lithium battery SOH prediction. Attached Figure Description
[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a schematic diagram of a lithium battery SOH prediction method based on dual-source channel enhancement provided in an embodiment of the present invention.
[0021] Figure 2 This is a schematic diagram of the overall framework of the DVCEformer model provided in an embodiment of the present invention.
[0022] Figure 3 This is a schematic diagram of the structure of the dual-source adaptive variable channel enhancement module provided in an embodiment of the present invention.
[0023] Figure 4 This is a schematic diagram illustrating the change of information about battery B0005 over time and cycles in an embodiment of the present invention.
[0024] Figure 5 A comparison chart of prediction results for batteries in NASA dataset B0005 is provided for embodiments of the present invention.
[0025] Figure 6 This is a comparison chart of the prediction results for batteries in the CALCE dataset CS2_35 in this embodiment of the invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0027] See Figure 1 This application provides a method for predicting the state of harm (SOH) of lithium batteries based on dual-source channel enhancement, comprising the following steps: S1, acquire multivariate degradation time series data of lithium battery within a historical time window, input the multivariate degradation time series data into the inverted embedding layer, and independently embed the historical time series corresponding to each variable in the multivariate degradation time series data into a variable-level representation to obtain a variable-level embedding representation; In some embodiments of this application, the multivariate degradation time series data includes the monitored values of at least two of the following variables at each historical time step: voltage, current, temperature, internal resistance, and capacity.
[0028] Specifically, operational data of lithium batteries during charge-discharge cycles are collected through a Battery Management System (BMS) or battery testing platform. The collected data includes, but is not limited to: terminal voltage (V), charge / discharge current (A), battery surface temperature (°C), DC internal resistance (mΩ), and available capacity (Ah). This data is organized according to the number of cycles and the time steps within each cycle to form multivariate degradation time-series data.
[0029] Suppose the collected multivariate degradation time series data are:
[0030] Where B represents the batch size, i.e. the number of samples processed by the model at one time; L represents the historical time step, i.e. the number of historical observation data used for prediction; and N represents the number of feature variables, i.e. the number of variables in the selected multivariate degenerate time series data, such as the number of variables such as voltage, current, and temperature.
[0031] In step S1, the process of independently embedding the historical time series corresponding to each variable in the multivariate degraded time series data into a variable-level representation to obtain the variable-level embedded representation specifically includes: Each variable in the multivariate degraded time series data is treated as an independent variable-level token, resulting in N variable-level tokens, where N is the number of variables in the multivariate degraded time series data. That is, the input sequence is... Transform into This means treating the historical time series corresponding to each variable as an independent token.
[0032] Each of the variable-level tokens is projected onto the D-dimensional latent space through a linear mapping to obtain the variable-level embedding representation. The transformation process is as follows:
[0033] in, The weight matrix is the linear mapping. For bias terms, For variable-level embedding representation, D is the embedding dimension.
[0034] S2, the variable-level embedding representation is input into the dual-source channel enhancement module, and the variable-level embedding representation is enhanced by gating fusion through two sets of learnable channel embeddings to obtain the channel-enhanced representation; like Figure 2 and Figure 3 As shown, in one embodiment provided in this application, the dual-source channel enhancement module explicitly enhances the variable-level representation by constructing a dual-branch channel embedding mechanism; at the same time, it combines a gating fusion strategy to dynamically adjust the semantic information of different channels, thereby improving the model's ability to capture key degradation information.
[0035] The process in S2 specifically includes: Channel gating weights are generated by the first set of learnable channel embeddings, and the variable-level embedding representation is pre-modulated using the channel gating weights to obtain the channel-premodulated feature representation. Specifically, the variable-level embedding obtained in step S1 represents the Z-input dual-source channel enhancement module, in which two sets of learnable channel embedding parameters are introduced for each variable channel. , , , These represent the channel prior representations from two different sources: the variable-level embedding representation Z serves as the input feature of the dual-source channel enhancement module.
[0036] Embedding with the first set of learnable channels Generate channel gating weights. Specifically, on the variable dimension... Perform average pooling and max pooling calculations separately:
[0037] The average pooling and max pooling results are fed into two independent linear mapping layers, respectively. The outputs of the two linear mapping layers are summed and then activated by the Sigmoid function to generate recalibrated weights, which are used as the channel gating weights.
[0038] in, and These are the weight matrices for two independent linear mapping layers. and These are the corresponding bias terms, Sigmoid is the activation function, and G is the channel gating weight.
[0039] After broadcasting the channel gating weight G to the variable dimension N, it is multiplied element-wise with the variable-level embedding representation Z to obtain the channel premodulated feature representation:
[0040] in, This represents element-wise multiplication. This represents the characteristics of the channel after premodulation.
[0041] Adaptive gating fusion is performed on the second set of learnable channel embeddings and the channel premodulated feature representations based on the channel premodulated feature representations to obtain the fused channel enhancement information.
[0042] In a more specific embodiment provided in this application, the obtained channel premodulation feature representation is... Input to the first gated mapping layer to premodulate the feature representation of the channel. A nonlinear transformation is performed to extract higher-level channel semantic information, resulting in the semantic representation of the intermediate channels.
[0043]
[0044] in, and Let represent the weight matrix and bias term of the first gated mapping layer, respectively. ReLU is the activation function, and Dropout is the random deactivation operation. This represents the semantics of the intermediate channel. The semantic representation of the intermediate channel The input to the second gating mapping layer is activated by the Sigmoid function to generate dynamic gating weights. :
[0045] in, and These represent the weights and biases of the linear transformation component, respectively. The value range is [0,1], which can be regarded as a dynamic selection coefficient for different channels and different hidden dimensions.
[0046] After obtaining the dynamic gating weight The two sets of channel representations are then weighted and fused. The first part consists of the semantic representation of the intermediate channel. and Element-wise multiplication yields the first weighted channel representation:
[0047] The second part is embedded by the second set of learnable channels. With the dynamic gating weight Element-wise multiplication yields the second weighted channel representation:
[0048] The first weighted channel representation and the second weighted channel representation are fused to obtain the channel enhancement information:
[0049] The fused channel enhancement information is superimposed onto the variable-level embedding representation in a residual manner to obtain the channel enhancement representation.
[0050] In a more specific embodiment provided in this application, the obtained fused channel enhancement information The residuals are superimposed onto the variable-level embedding representation Z to obtain the final output of the dual-source channel enhancement module, i.e., the channel enhancement representation.
[0051] This residual connection method preserves the basic degradation information in the original input Z, avoiding the loss of original information due to channel enhancement operations. Simultaneously, it explicitly injects the semantic priors of variables obtained from channel embedding and gated fusion, thereby improving the model's ability to perceive the differences in contributions from different health factors. The resulting channel-enhanced representation Z′ will then serve as the encoder input.
[0052] S3, the channel enhancement representation is input into the encoder, processed by a multi-head attention mechanism, to establish the global dependency relationship between variables in the multivariate degenerate time series data, and to obtain the encoded feature representation; In some embodiments of the present invention, in this step, the channel enhancement representation obtained in S2 is used to input the encoder.
[0053] The encoder includes multiple encoder layers, each encoder layer including a multi-head attention sublayer, a first residual connection and layer normalization sublayer, a feedforward network sublayer, and a second residual connection and layer normalization sublayer.
[0054] In the multi-head attention sublayer, single-head attention is calculated as follows:
[0055] Where Q is the query vector, K is the keyword vector, V is the value vector, and softmax is the normalization function. Let K be the dimension of the keyword vector.
[0056] Multi-head attention is composed of multiple single-head attention sequences, the first of which is the first single-head attention sequence. i The size is represented as:
[0057] Multi-head attention can then be represented as:
[0058] in, h For the number of attention heads, To output the projection weight matrix, This indicates a splicing operation.
[0059] In the layer normalization sublayer, the formula for calculating layer normalization is:
[0060] Where H is the matrix to be normalized. The first element to be normalized in matrix H n There are N sequences, where Mean() is the mean function and Var() is the variance function. N is the number of sequences in matrix H that need to be normalized.
[0061] The feedforward sublayer consists of two linear transformation units and a nonlinear activation function located between the two linear transformation units. The input to the feedforward sublayer is:
[0062] The nonlinear calculation process of the feedforward network sublayer is as follows:
[0063] in, W r1 , W r2 It is the weight matrix that needs to be learned. b 1. b 2 is the bias vector.
[0064] After residual connections and layer normalization, the output of each encoder layer is:
[0065] After being processed by stacking multiple encoder layers, the encoded feature representation is obtained.
[0066] The channel enhancement representation Z′, after passing through the encoder, yields the encoded feature as follows:
[0067] S4. The encoded feature representation is input into the nonlinear prediction head, and after nonlinear mapping processing, the SOH prediction value of the lithium battery within the future prediction window is obtained.
[0068] In some embodiments of the present invention, in this step, the encoded feature representation obtained in S3 is input into the nonlinear prediction head. The nonlinear prediction head includes a first linear transform layer, a first GELU activation function, a second linear transform layer, a second GELU activation function, and an output linear layer connected in sequence.
[0069] In the first linear transform layer and the first GELU activation function, the encoded feature representation is input into the first linear transform layer, and then nonlinearly mapped by the first GELU activation function to obtain the first intermediate prediction feature:
[0070] in, and These represent the weight matrix and bias term of the first-layer prediction mapping, respectively, with GELU being the activation function. This is the first intermediate prediction feature.
[0071] In the second linear transformation layer and the second GELU activation function, the first intermediate predicted features are... The input is processed through the second linear transformation layer, followed by a nonlinear mapping using the second GELU activation function, to obtain the second intermediate predicted feature:
[0072] in, and These represent the weight matrix and bias term of the second-layer mapping, respectively. This is the second intermediate prediction feature.
[0073] In the output linear layer, the second intermediate prediction feature is... The input-output linear layer is mapped to the target prediction step space, and then subjected to a Permute dimension transformation operation to obtain the SOH prediction value of the lithium battery within the future prediction window:
[0074] in, and These represent the weight matrix and bias term of the third-layer prediction mapping, respectively. P This indicates the prediction step size, and Y is the predicted SOH value of the lithium battery within the future prediction window.
[0075] This application proposes the DVCEformer model, which is a lithium battery SOH prediction model that introduces a dual-source adaptive variable channel enhancement (DSAVCE) mechanism and a nonlinear prediction head on the basis of iTransformer.
[0076] To verify the effectiveness of the DVCEformer model proposed in this invention, comparative and ablation experiments were conducted on a public dataset, and the experimental results were analyzed in detail.
[0077] This embodiment uses the lithium-ion battery aging dataset from NASA's Ames Center of Excellence for Prediction (PCoE) and the battery dataset from the University of Maryland's Center for Advanced Life Cycle Engineering (CALCE), referred to as the NASA dataset and the CALCE dataset, respectively. Both datasets record relatively complete information on the degradation process and are therefore widely used for training and validating lithium battery life prediction models.
[0078] In the NASA dataset, four typical battery samples—B0005, B0006, B0007, and B0018—were selected as experimental subjects; in the CALCE dataset, battery samples CS2_35, CS2_36, CS2_37, and CS2_38 were selected for experiments. The discharge cycles of each battery were chosen as the basic research unit for degradation state analysis and health state estimation.
[0079] like Figure 4As shown, using battery B0005 as an example, the changes in battery information over time and cycles are illustrated. The two datasets contain data on voltage, current, and capacity changes during multiple charge-discharge cycles. In the experiment, discharge cycle information will be selected as the raw data for calculating the battery's state of harm (SOH).
[0080] To visually demonstrate the basic parameters and test conditions of the selected dataset samples, the information such as the sample number, ambient temperature, rated capacity, battery type, charge / discharge method, and number of discharge cycles for each battery sample is compiled as shown in Table 1: Table 1. Basic information about the NASA and CALCE datasets.
[0081] To comprehensively evaluate the predictive performance of the proposed model in the lithium battery lifetime prediction task, this embodiment selects the coefficient of determination (R²). 2 The mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) are used as model evaluation metrics. These metrics comprehensively measure the model's prediction results from different perspectives, including goodness of fit, squared error loss, absolute bias, error dispersion, and relative error.
[0082] To comprehensively evaluate the predictive performance of the proposed method, comparative and ablation experiments were conducted on a unified experimental platform. The hardware equipment and software environment configuration used in the experiments are shown in Table 2.
[0083] Table 2 Hardware and Software Environment Description
[0084] To better simulate real-world applications and more realistically evaluate the model's generalization performance on unknown batteries, this invention employs a Leave-One-Battery-Out (LOBO) strategy for experiments. Specifically, this strategy involves selecting one battery sample as the test set each time, and using the remaining battery samples as the training set. This process is repeated until all batteries have been evaluated as the test set once. The final result is the average of all battery test results. This strategy effectively evaluates the model's generalization ability on new batteries that were not used in training, more closely reflecting real-world scenarios where the battery type and operating conditions are unknown.
[0085] The training parameters were set as shown in Table 3 during the experiment: Table 3 Parameter Description
[0086] Here, the input sequence length `seq_len` represents the number of historical cycles used for prediction, and the prediction length `pred_len` represents the number of future cycles to be predicted. Because the NASA dataset and the CALCE dataset have different degradation rates, different input sequence length and prediction length settings are used.
[0087] To verify the effectiveness of the proposed DVCEformer model, several advanced models for lithium battery lifetime prediction were selected for comparative experiments, including GRU, LSTM, Transformer, and MSSformer. All comparative models were trained and tested on the same dataset, employing the same leave-one-battery strategy.
[0088] The results of the comparative experiments are shown in Table 4: Table 4. Experimental results comparing the model in this paper with other models.
[0089] As shown in Table 4, the DVCEformer model proposed in this invention achieves optimal prediction results on both the NASA and CALCE datasets, demonstrating high fitting accuracy and low prediction error. Specifically, on the NASA dataset, the R-value of the model in this invention is [missing information]. 2 The R score reached 0.990, with MAE and RMSE decreasing to 0.440% and 0.861%, respectively. On the CALCE dataset, R... 2 The accuracy reached 0.998, with MAE and RMSE of only 0.242% and 0.376%, respectively, demonstrating significantly better overall performance than other comparative models. This indicates that the method proposed in this invention has stronger accuracy and generalization ability in the lithium battery SOH prediction task.
[0090] To more intuitively demonstrate the ability of each model to track battery degradation trends, local fluctuation characteristics, and changes in the end-of-life range, a set of experiments was conducted using the B0018 battery from the NASA dataset and the CS2_38 battery from the CALCE dataset, with visualizations provided for each. For example... Figure 5 As shown, when using the B0018 battery as the test object in the NASA dataset, the predicted curve of the model of this invention is closest to the actual SOH degradation curve, especially showing higher prediction accuracy in the critical interval near the end of life; as Figure 6 As shown, when the CS2_38 battery is used as the test object in the CALCE dataset, the model of this invention also exhibits the best degradation trend tracking ability and local fluctuation fitting ability, further verifying the generalization performance of the model of this invention in different datasets and different degradation modes.
[0091] To verify the effectiveness of each component of the dual-source channel enhancement module in this invention, an ablation experiment was designed to verify each component individually. The predictive performance differences between the variant models and the complete model were compared by removing the variable channel enhancement module, the adaptive gated fusion strategy, the residual connection, and replacing the nonlinear prediction head with a single-layer linear projection layer. Experimental results show that removing the variable channel enhancement module resulted in the most significant performance decrease, indicating that this module is the primary source of performance improvement. Removing the adaptive gated fusion strategy led to an increase in prediction error, indicating that the gated fusion strategy further enhanced the selection ability of variable information. Removing the nonlinear prediction head reduced the model's fitting ability in the complex degradation stage, indicating that the nonlinear prediction head plays a supplementary optimization role in the complete model.
[0092] In a more specific embodiment provided in this application, a lithium battery SOH prediction system based on dual-source channel enhancement is provided, comprising the following modules: The inversion embedding module is used to acquire multivariate degradation time series data of lithium batteries within a historical time window, input the multivariate degradation time series data into the inversion embedding layer, and independently embed the historical time series corresponding to each variable in the multivariate degradation time series data into a variable-level representation to obtain a variable-level embedding representation; A dual-source channel enhancement module is used to input the variable-level embedding representation into the dual-source channel enhancement module, and perform gated fusion enhancement on the variable-level embedding representation through two sets of learnable channel embeddings to obtain a channel-enhanced representation. The encoder module is used to input the channel enhancement representation into the encoder, process it through a multi-head attention mechanism, establish the global dependency relationship between variables in the multivariate degenerate time series data, and obtain the encoded feature representation. The nonlinear prediction head module is used to input the encoded feature representation into the nonlinear prediction head, and after nonlinear mapping processing, obtain the SOH prediction value of the lithium battery within the future prediction window.
[0093] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.
Claims
1. A method for predicting the state of harm (SOH) of a lithium battery based on dual-source channel enhancement, characterized in that, Includes the following steps: S1, acquire multivariate degradation time series data of lithium battery within a historical time window, input the multivariate degradation time series data into the inverted embedding layer, and independently embed the historical time series corresponding to each variable in the multivariate degradation time series data into a variable-level representation to obtain a variable-level embedding representation; S2, the variable-level embedding representation is input into the dual-source channel enhancement module, and the variable-level embedding representation is enhanced by gating fusion through two sets of learnable channel embeddings to obtain the channel-enhanced representation; S3, the channel enhancement representation is input into the encoder, processed by a multi-head attention mechanism, to establish the global dependency relationship between variables in the multivariate degenerate time series data, and to obtain the encoded feature representation; S4. The encoded feature representation is input into the nonlinear prediction head, and after nonlinear mapping processing, the SOH prediction value of the lithium battery within the future prediction window is obtained.
2. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 1, characterized in that, The multivariate degradation time series data in S1 includes the monitored values of at least two of the following variables at each historical time step: voltage, current, temperature, internal resistance, and capacity.
3. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 1, characterized in that, In step S1, the process of independently embedding the historical time series corresponding to each variable in the multivariate degraded time series data into a variable-level representation to obtain the variable-level embedded representation specifically includes: Each variable in the multivariate degraded time series data is treated as an independent variable-level token, resulting in N variable-level tokens, where N is the number of variables in the multivariate degraded time series data. Each of the variable-level tokens is projected onto a D-dimensional latent space through a linear mapping to obtain the variable-level embedding representation.
4. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 1, characterized in that, The process in S2 specifically includes: Channel gating weights are generated by the first set of learnable channel embeddings, and the variable-level embedding representation is pre-modulated using the channel gating weights to obtain the channel-premodulated feature representation. Adaptive gating fusion is performed on the second set of learnable channel embeddings and the channel premodulated feature representations based on the channel premodulated feature representations to obtain the fused channel enhancement information; The fused channel enhancement information is superimposed onto the variable-level embedding representation in a residual manner to obtain the channel enhancement representation.
5. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 4, characterized in that, The generation of channel gating weights through the embedding of the first set of learnable channels specifically includes: Obtain the first set of learnable channel embedding parameters, and perform max pooling and average pooling on the first set of learnable channel embedding parameters in the variable dimension to obtain the global max pooling result and the global average pooling result. The global max pooling result and the global average pooling result are respectively input into independent linear mapping layers. The outputs of the two linear mapping layers are summed and then activated by the Sigmoid function to generate recalibrated weights as the channel gating weights.
6. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 4, characterized in that, The step of adaptively gating and fusing the second set of learnable channel embeddings with the channel premodulated feature representations to obtain fused channel enhancement information specifically includes: The pre-modulated feature representation of the channel is input into the first gated mapping layer for nonlinear transformation to obtain the semantic representation of the intermediate channel. The semantic representation of the intermediate channel is input into the second gating mapping layer and activated by the Sigmoid function to generate dynamic gating weights. The channel information after nonlinear mapping is multiplied by the dynamic gating weight to obtain the first weighted channel representation; The second set of learnable channels is embedded and multiplied by the dynamic gating weights to obtain the second weighted channel representation; The first weighted channel representation is added to the second weighted channel representation to obtain the fused channel enhancement information.
7. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 1, characterized in that, The encoder in S3 includes multiple encoder layers, each encoder layer including a multi-head attention sublayer, a first residual connection and layer normalization sublayer, a feedforward network sublayer, and a second residual connection and layer normalization sublayer.
8. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 1, characterized in that, In S4, the nonlinear prediction head includes a first linear transformation layer, a first GELU activation function, a second linear transformation layer, a second GELU activation function, and an output linear layer connected in sequence. The nonlinear mapping process yields the predicted SOH value of the lithium battery within the future prediction window, specifically including: The encoded feature representation is input into the first linear transform layer and nonlinearly mapped with the first GELU activation function to obtain the first intermediate prediction feature; The first intermediate prediction feature is input into the second linear transformation layer and nonlinearly mapped with the second GELU activation function to obtain the second intermediate prediction feature; The second intermediate prediction feature is input into the output linear layer and mapped to the target prediction step space to obtain the SOH prediction value.
9. The lithium battery SOH prediction method based on dual-source channel enhancement according to claim 7, characterized in that, The feedforward network sublayer includes two linear transformation units and a nonlinear activation function located between the two linear transformation units.
10. A lithium battery SOH prediction system based on dual-source channel enhancement, characterized in that, include: The inversion embedding module is used to acquire multivariate degradation time series data of lithium batteries within a historical time window, input the multivariate degradation time series data into the inversion embedding layer, and independently embed the historical time series corresponding to each variable in the multivariate degradation time series data into a variable-level representation to obtain a variable-level embedding representation; A dual-source channel enhancement module is used to input the variable-level embedding representation into the dual-source channel enhancement module, and perform gated fusion enhancement on the variable-level embedding representation through two sets of learnable channel embeddings to obtain a channel-enhanced representation. The encoder module is used to input the channel enhancement representation into the encoder, process it through a multi-head attention mechanism, establish the global dependency relationship between variables in the multivariate degenerate time series data, and obtain the encoded feature representation. The nonlinear prediction head module is used to input the encoded feature representation into the nonlinear prediction head, and after nonlinear mapping processing, obtain the SOH prediction value of the lithium battery within the future prediction window.