Lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU

Abstract

The application provides a lithium battery residual life prediction method based on multiple health factor fusion and improved GRU, comprising the following steps: extracting multiple health factors strongly related to the actual capacity of a lithium battery from the charging and discharging process of the lithium battery; performing dimension reduction fusion on the screened factors by using principal component analysis to extract fused health factors; performing self-adaptive decomposition on the fused health factors by using continuous variational mode decomposition to obtain a plurality of subsequence components; constructing a gated recurrent unit network model with an introduced attention mechanism, predicting and integrating and reconstructing each subsequence component to obtain a fused health factor prediction value; and determining the residual service life of the lithium battery according to the prediction value and a preset failure threshold. The application effectively solves the problems of difficult prediction caused by the difficulty in online measurement of capacity and the capacity regeneration phenomenon, and significantly improves the prediction accuracy and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of lithium battery life prediction technology, specifically to a method for predicting the remaining life of lithium batteries based on the fusion of multiple health factors and an improved GRU. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, energy storage systems, and other fields. However, their performance degrades with increasing usage cycles. Accurately predicting their remaining useful life (RUL) is crucial for ensuring system safety and reducing operation and maintenance costs. Existing data-driven prediction methods have the following shortcomings: 1) Most methods directly use battery capacity as a health factor, but in practical applications, capacity is difficult to measure directly online; 2) Some methods extract indirect health factors from the charging and discharging process, but they often only use one or a few factors, resulting in incomplete information and failure to consider information redundancy between factors. 3) Lithium battery capacity data exhibits nonlinearity, non-stationarity, and capacity regeneration phenomena. A single prediction model cannot simultaneously capture long-term degradation trends and short-term random fluctuations, and the prediction accuracy and generalization need to be improved.

[0003] This reveals that existing methods for predicting the remaining life of lithium batteries suffer from several problems, including the selection of only one health factor, insufficient consideration of the charging and discharging process and temperature factors, and difficulties in prediction due to capacity regeneration.

[0004] Therefore, how to provide a method that helps improve the accuracy, robustness, and generalization ability of lithium battery remaining life prediction is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU, to overcome the shortcomings of the prior art in terms of incomplete health factor selection, difficulty in processing nonlinear non-stationary sequences, and low prediction accuracy of single model, and achieves high-precision and strong generalization prediction of lithium battery remaining life.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention discloses a method for predicting the remaining life of lithium batteries based on multi-health factor fusion and improved GRU, comprising the following steps: S1: Obtain multiple health factors of lithium batteries during the cycle charge and discharge process, and screen out the health factors that meet the preset conditions and are correlated with the actual capacity of the battery. S2: Principal component analysis is used to perform dimensionality reduction and fusion processing on the screened health factors, and the principal components are extracted as fused health factors. S3: Adaptively decompose the fused health factor using continuous variational mode decomposition to obtain multiple sub-sequence components with different frequency domain characteristics; S4: Construct a gated recurrent unit network model that incorporates an attention mechanism, use the gated recurrent unit network model to predict each of the subsequence components, and integrate and reconstruct the prediction results to obtain the fused health factor prediction value; S5: Based on the predicted value of the fused health factor and the preset failure threshold, determine the remaining service life of the lithium battery.

[0007] Preferably, the health factors include constant current charging time interval, constant voltage rise charging time interval, constant voltage drop discharging time interval, and maximum temperature rise time interval; wherein, The constant current charging time interval is the time taken for the lithium battery to reach the cutoff voltage during the constant current charging phase. The constant voltage rise charging time interval is the time it takes for the voltage of the lithium battery to rise from a first preset voltage value to a second preset voltage value during the charging process; The equal voltage drop discharge time interval is the time it takes for the voltage of the lithium battery to drop from the third preset voltage value to the fourth preset voltage value during the discharge process; The maximum temperature rise time interval is the time it takes for the internal temperature of the lithium battery to rise from the initial temperature to the highest temperature.

[0008] Preferably, the step of obtaining health factors in S1 that satisfy preset conditions in relation to the actual battery capacity includes: The correlation between each candidate factor and the actual battery capacity was calculated using Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient, respectively. Candidate factors whose three correlation coefficients all met the strong correlation threshold were selected as the health factors.

[0009] Preferably, S2 includes the following steps: S21: Standardize the selected health factor data and construct a covariance matrix; S22: Calculate the cumulative contribution rate of each principal component in the covariance matrix; S23: Select the first principal component whose cumulative contribution rate reaches the preset contribution rate threshold as the fusion health factor, and discard the remaining principal components.

[0010] Preferably, in S3, the subsequence components include at least a residual signal characterizing the global degradation trend, a low-frequency fluctuation signal characterizing the capacity regeneration phenomenon, and a high-frequency fluctuation signal characterizing random noise.

[0011] Preferably, S3 includes the following steps: An iterative decomposition method is adopted. After each decomposition, the correlation coefficient between the currently extracted modal component and the original fused health factor is calculated, and modal components with correlation coefficients higher than the first preset threshold are retained. Calculate the ratio of the energy of the current residual signal to the energy of the original fused health factor sequence; stop the decomposition when the energy ratio is lower than the second preset threshold.

[0012] Preferably, in the iterative decomposition process, after each extraction of a modal component, the current residual signal is subtracted from the extracted modal component to update the residual signal, and the next extraction of modal components is performed based on the updated residual signal.

[0013] Preferably, the gated recurrent unit network model incorporating an attention mechanism includes an update gate, a reset gate, a candidate hidden state generation unit, and an attention gate; The reset gate is used to control the degree of influence of the previous hidden state on the candidate hidden state; The candidate hidden state generation unit is used to generate candidate hidden states based on the output of the reset gate and the input at the current moment; The attention gate is used to generate an attention weight vector based on the current input and the hidden state of the previous time step. The network model performs element-wise multiplication of the attention weight vector with the candidate hidden state to obtain the attention-enhanced candidate state, and then uses the update gate to perform weighted fusion of the hidden state at the previous time step and the attention-enhanced candidate state to obtain the final hidden state at the current time step.

[0014] As can be seen from the above technical solution, compared with the prior art, the beneficial effects of the present invention include: To address the problems of limited selection of health factors, difficulty in online measurement, and insufficient prediction accuracy due to capacity regeneration in existing technologies, this invention provides a lithium battery remaining life prediction method based on multi-health factor fusion and improved gated cycling unit, aiming to comprehensively improve the accuracy, robustness, and generalization ability of the prediction.

[0015] Based on this, to address the problem that the fused health factors still have nonlinear and nonstationary characteristics, this invention uses continuous variational mode decomposition to adaptively decompose the signal, which can effectively separate the global degradation trend, capacity regeneration fluctuations and random noise. This not only avoids the mode mixing problem, but also provides more periodic and stable subsequence inputs for subsequent prediction models.

[0016] Furthermore, this invention introduces an attention mechanism into the gated recurrent unit network, enabling the model to adaptively focus on key degradation events such as capacity regeneration during training and prediction. Combined with sequence decomposition and ensemble prediction strategies, the output is reconstructed after modeling and predicting each sub-sequence component separately, thus significantly improving the prediction accuracy of remaining service life in complex degradation scenarios. Validation results on different types of lithium battery datasets demonstrate that this method possesses excellent generalization performance and transferability, meeting the requirements for prediction reliability and stability in practical engineering applications.

[0017] This method extracts multiple online measurable indirect health factors from three dimensions: charging, discharging, and temperature changes. It then uses principal component analysis to fuse and reduce the dimensionality of these factors, constructing a fusion health factor that comprehensively considers the effects of voltage, current, and temperature. This method effectively replaces capacity signals that are difficult to obtain directly online while eliminating information redundancy, thus laying a data foundation for subsequent modeling. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart of a lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of multiple health factors provided in an embodiment of the present invention; Figure 3 A comparison chart of the FHI and the normalized decay curves of the actual battery capacity provided in an embodiment of the present invention; Figure 4 A schematic diagram of the SVMD decomposition results of FHI constructed based on B05 cells provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the Attention-GRU network structure provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the prediction results of B05 in five scenarios provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention discloses a method for predicting the remaining life of lithium batteries based on multi-health factor fusion and improved GRU, such as... Figure 1 As shown, it mainly includes the following steps: S1: Obtain multiple health factors of the lithium battery that are correlated with the actual battery capacity and meet preset conditions during the cyclic charging and discharging process; S2: Principal component analysis is used to perform dimensionality reduction and fusion processing on the screened health factors, and the principal components are extracted as the fused health factors. S3: Adaptive decomposition of the fused health factor is performed using continuous variational mode decomposition to obtain multiple sub-sequence components with different frequency domain characteristics; S4: Construct a gated recurrent unit network model that incorporates an attention mechanism, use the gated recurrent unit network model to predict each subsequence component, and integrate and reconstruct the prediction results to obtain the fused health factor prediction value; S5: Determine the remaining lifespan of the lithium battery based on the predicted values ​​of the integrated health factors and the preset failure threshold.

[0022] In one embodiment, multiple health factors (HI) that can indirectly characterize battery capacity degradation are extracted from lithium battery cycle charge-discharge data, focusing on three aspects: charging stage, discharging stage, and battery temperature change. These health factors include constant current charging time interval (HI1), constant voltage rise charging time interval (HI2), constant voltage drop discharging time interval (HI3), and maximum temperature rise time interval (HI4). The constant current charging time interval (HI1) is the time taken for a lithium battery to reach its cutoff voltage during the constant current charging phase. For example... Figure 2 As shown in (a), this is a schematic diagram of the current curve during constant voltage charging of the B05 battery. During the charging stage of the lithium battery, it is often charged with a constant current of 1.5A until the voltage reaches 4.2V. This period is recorded as the constant current charging time interval. Constant current charging segments from the 5th, 50th, 100th, and 150th cycles of battery B05 were selected.

[0023] The constant voltage rise charging time interval (HI2) is the time required for the lithium battery voltage to rise from a first preset voltage value to a second preset voltage value during the charging process. For example... Figure 2As shown in (b), this is a schematic diagram of the voltage curve of the B05 battery during constant current charging, showing the time it takes for the B05 battery voltage to charge from 3.9V to 4.1V, and the time interval of constant voltage rise charging in the 5th cycle. Maximum, isobaric boost charging time interval during the 150th cycle Minimum, constant voltage rise charging time interval The decrease gradually increases with the number of charge-discharge cycles, following the same trend as capacity decay.

[0024] The equal voltage drop discharge time interval (HI3) is the time it takes for the voltage of a lithium battery to drop from a third preset voltage value to a fourth preset voltage value during discharge. In other words, it's the time interval between discharging from a certain starting voltage to another voltage (e.g., discharging from 4.0V to 3.9V, and from 3.9V to 3.6V). Figure 2 As shown in (c), this is a schematic diagram of the discharge voltage curve during constant current discharge of battery B05, and the equal voltage drop discharge time interval of the 5th charge-discharge cycle. Maximum, constant voltage drop discharge time interval with increasing charge / discharge cycles The constant voltage drop discharge time interval gradually decreases during the 150th charge-discharge cycle. Minimum. Overall, the equal voltage drop discharge time interval... The value decreases with increasing cycle count and is positively correlated with the battery capacity degradation trend.

[0025] The maximum temperature rise time interval (HI4) is the time it takes for the internal temperature of a lithium battery to rise from its initial temperature to its maximum temperature, i.e., the time interval between the battery's internal temperature rising from 24°C to its maximum temperature. Figure 2 As shown in (d), this is a schematic diagram of the temperature change curve of battery B05, showing the time interval of the maximum temperature rise during the 5th charge-discharge cycle. Maximum, the time interval of maximum temperature rise with increasing charge / discharge cycles. The time interval of maximum temperature rise during the 150th charge-discharge cycle gradually decreases. Minimum. Overall, the time interval between maximum temperature rises. The value decreases with increasing cycle count and is positively correlated with the battery capacity degradation trend.

[0026] In one embodiment, the step of obtaining health factors in S1 that satisfy preset conditions related to the actual battery capacity includes: The correlation between each candidate factor and the actual battery capacity was calculated using Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient respectively. Candidate factors that meet the strong correlation threshold of all three correlation coefficients were selected as healthy factors, and candidate factors with weak correlation were removed.

[0027] The correlation between the battery capacity of FHI and B05 batteries was verified, as shown in Table 1 below.

[0028] Table 1. Correlation coefficients between FHI and B05 battery capacities

[0029] Table 1 shows that the Pearson, Spearman, and Kendall correlation coefficients between FHI extracted by PCA and the capacity of B05 batteries are all greater than 0.97, very close to 1, indicating a strong correlation between FHI and B05 battery capacity, which can perfectly characterize the capacity degradation trend of B05 lithium batteries. Data normalization was performed on the FHI and B05 battery capacity data to address the impact of different units of measurement.

[0030] Similarly, the correlation between FHI and the battery capacity of B06, B07, and B18 batteries was verified, as shown in Table 2 below.

[0031] Table 2 Correlation coefficients between FHI and NASA battery capacity

[0032] As shown in Table 2, in the correlation analysis between FHI and B06, all three correlation coefficients are greater than 0.9770; in the correlation analysis between FHI and B07, all three correlation coefficients are greater than 0.9650; and in the correlation analysis between FHI and B18, all three correlation coefficients are greater than 0.9570. This indicates that the FHI constructed by PCA in this invention has a strong correlation with the battery capacities of B06, B07, and B18. FHI can characterize the degradation state of lithium batteries and can replace battery capacity for subsequent battery RUL prediction.

[0033] In one embodiment, S2 includes the following steps: S21: Standardize the selected health factor data and construct a covariance matrix; S22: Calculate the cumulative contribution rate of each principal component in the covariance matrix; S23: Select the first principal component whose cumulative contribution rate reaches a preset contribution rate threshold (e.g., 95%) as the Fusion Health Factor (FHI), and discard the remaining principal components to reduce model complexity. This FHI comprehensively considers the effects of voltage, current, and temperature on battery degradation, and can fully characterize the capacity degradation trend of lithium batteries.

[0034] In this embodiment, the specific mathematical steps of Principal Component Analysis (PCA) are as follows: Feature data sample acquisition: Let the number of features be... m Each feature vector contains n Each element constructs a feature. Sample matrix.

[0035]

[0036] In the formula, For the first i The th eigenvector in the th eigenvector j Each element.

[0037] Data standardization: To eliminate the influence of different eigenvectors' dimensions, standardization is required. The formula is shown below:

[0038] In the formula, This is the original data; For data annotation; For the first j The mean of the column; For the first j Standard deviation of the column.

[0039] Calculate the normalized matrix covariance matrix As shown below:

[0040] Calculate the eigenvalues ​​of the covariance matrix With eigenvectors.

[0041] Calculate the contribution value and cumulative contribution rate of the eigenvalues. The contribution rate of this eigenvalue. Cumulative contribution rate:

[0042] .

[0043] In one embodiment, the FHI constructed based on PCA is highly correlated with capacity and can replace capacity for subsequent modeling and prediction. FHI contains most of the capacity information, including random fluctuations such as capacity regeneration. Therefore, FHI is still a nonlinear and non-stationary signal. To facilitate subsequent modeling and prediction, FHI data preprocessing is required. Continuous Variational Mode Decomposition (SVMD) is used to adaptively decompose the preprocessed fused health factor (FHI) sequence. When decomposing FHI using SVMD, SVMD dynamically determines the optimal number of mode decompositions K through iterative decomposition and correlation analysis, and dynamically adjusts the penalty factor according to the energy distribution of the residual signal to accurately separate long-term degradation trends, capacity regeneration fluctuations, and noise.

[0044] The preprocessing steps for FHI sequences are as follows: The FHI sequence obtained from S2 was normalized to obtain the normalized FHI sequence. Recorded as:

[0045] in, This is the original FHI sequence. The minimum value in the sequence. This represents the maximum value in the sequence.

[0046] In this embodiment, the normalized FHI sequence is decomposed using SVMD. It is decomposed into several modal components and a residual signal, represented as:

[0047] in, For the Kth modal component, For residual signals, K The number of modes obtained from the decomposition.

[0048] The variational model of SVMD can be expressed as:

[0049] The constraints are:

[0050] in, For the first k The center frequency of each modal component Let be the impulse function. j The imaginary unit, This represents the convolution operation.

[0051] like Figure 3 As shown, the black curve is the actual battery capacity decay curve, which is obtained by normalizing the calculation using the ampere-hour integration method based on the NASA dataset; the red curve is the battery capacity decay curve after normalization by incorporating health factors, which is obtained by normalizing the first principal component after dimensionality reduction using four health factors (HI) through PCA.

[0052] In one embodiment, in S3, the subsequence components include at least a residual signal characterizing the global degradation trend, a low-frequency fluctuation signal characterizing the capacity regeneration phenomenon, and a high-frequency fluctuation signal characterizing sampling error, environmental disturbance, and other random noise. Figure 4 As shown, IMF1 represents the fluctuation signal of capacity regeneration, while IMF2 and IMF3 represent the implicit noise sequences in the lithium battery degradation process.

[0053] In one embodiment, S3 includes the following steps: An iterative decomposition method is adopted. After each decomposition, the correlation coefficient between the currently extracted modal component and the original fused health factor is calculated, and modal components with correlation coefficients higher than the first preset threshold are retained. Calculate the ratio of the energy of the current residual signal to the energy of the original fused health factor sequence; stop the decomposition when the energy ratio is lower than the second preset threshold.

[0054] The specific execution process of this embodiment is as follows: initialization k =1, perform SVMD decomposition on the current residual signal to obtain the th k Modal components .calculate Pearson correlation coefficient with the original FHI signal .like > If a preset threshold (typically 0.1) is found, the mode is considered a valid component and retained; otherwise, decomposition is stopped. The residual signal is then updated. ,make k = k +1, repeat the above process.

[0055] Calculate the Pearson correlation coefficients between each modal component and the original FHI sequence:

[0056] When the following conditions are met: If the modal component is positively correlated with the FHI degradation information, it is retained; otherwise, it is considered to contribute little to the FHI degradation information, belonging to spurious components or pure noise, and further decomposition is meaningless, so it is not retained. The threshold serves as a baseline for trend judgment, and prediction deviation affects the determination of the failure time. The preset correlation threshold is preferably set to 0.1.

[0057] During the iterative decomposition process, after each extraction of a modal component, the current residual signal is subtracted from the extracted modal component to update the residual signal. The next extraction of modal components is then performed based on the updated residual signal. The calculation process is as follows: Update the current residual signal:

[0058] And order k = k +1, repeat the above decomposition process until the ratio of the residual signal energy to the original FHI sequence energy is less than a preset threshold (less than 5% of the original signal energy) or the correlation coefficient is less than a threshold. At that point, the main component is considered to be residual noise, the decomposition is complete, and the decomposition is stopped.

[0059]

[0060]

[0061] in, The current residual signal energy, The energy of the original FHI sequence. The energy threshold is preferably set at 5%.

[0062] In one embodiment, a gated recurrent unit network (Attention-GRU) model incorporates an attention mechanism, integrating attention layers into the model to prioritize important elements in the input sequence. This simplifies the learning process and improves the overall model performance. The attention mechanism selects only the outputs of the previous layer that are crucial to the next stage of the model, allowing the network to selectively focus on specific pieces of information. The gated recurrent unit network model includes an update gate, a reset gate, a candidate hidden state generation unit, and an attention gate. The reset gate is used to control the degree of influence of the previous hidden state on the candidate hidden state; The candidate hidden state generation unit is used to generate candidate hidden states based on the output of the reset gate and the input at the current time. Attention gates are used to generate attention weight vectors based on the current input and the hidden state of the previous time step, assigning higher weights to important information in the input sequence to enhance the model's attention to key degradation features such as capacity regeneration. The network model performs element-wise multiplication of the attention weight vector with the candidate hidden state to obtain the attention-enhanced candidate state, and then uses the update gate to perform weighted fusion of the hidden state of the previous time step with the attention-enhanced candidate state to obtain the final hidden state of the current time step.

[0063] like Figure 5 As shown, the state update formula for the Attention-GRU network is as follows: Update Gate: ; Reset Door: ; Candidate hidden state: ; Attention gate: ; Final hidden state: .

[0064] in, for t Input at any time, This is the hidden state from the previous moment. To update the door status, To reset the door status, The candidate set state, For attention gate output, Hide the current state. , , For learning parameters, σ is the Sigmoid function, and ⊙ is the Hadamard product.

[0065] The Gated Recurrent Unit Network (Attention-GRU network) model provided in this embodiment is evaluated based on evaluation metrics: The root mean squared error (RMSE), mean absolute percentage error (MAPE), absolute error (AE), and relative error (RE) are used as evaluation metrics for this model. The smaller the evaluation metrics, the higher the accuracy of the prediction model. The formulas are as follows.

[0066]

[0067]

[0068]

[0069]

[0070] In the formula, This is the actual capacity value; To predict capacity values; This is the actual RUL value; This is the predicted RUL value; n To predict the number of cycle periods.

[0071] In one embodiment, to overcome the accumulated error introduced by traditional direct superposition reconstruction, step S4 adopts an adaptive weighted integration strategy based on prediction error feedback, the specific steps of which are as follows: S41: Remove the end of the training set... m Each sample is divided into a validation set. The root mean square error of prediction for each subsequence component (including residual components and K modal components) obtained from step S3 is calculated on the validation set. RMSE k ( k =1,2,…, K )and RMSE r ; S42: Set initial ensemble weights: , As can be seen from this formula, the larger the verification error, the smaller its weight in the final reconstruction; S43: Introduce a sliding window mechanism (window width L=10). Whenever a new cycle of measured data (or high-confidence prediction data) is added, the RMSE of each component within the window is recalculated and the weights are updated. This enables the weights to adaptively track the time-varying characteristics of the battery degradation process; S44: The prediction results of each sub-sequence component are weighted and summed using the updated dynamic weights, and the final predicted value of the fused health factor is obtained by the following formula. : .

[0072] in, For the Kth modal component in the prediction result, This represents the residual signal in the prediction results.

[0073] In one embodiment, step S5, the strategy for determining the remaining lifespan of the lithium battery based on the fused health factor prediction value and a preset failure threshold includes: S51: Set Failure Threshold: Set the capacity failure threshold based on the battery's rated capacity. Th (e.g., 1.4Ah), and convert it to the equivalent failure threshold of FHI; S52: Failure initiation time determination: ; S53: Subcyclic precision linear interpolation correction: ; S54: Calculate Remaining Service Life (RUL): ; in, Th The failure threshold is set to 0.7 after normalization, and L is the confirmation window width (preferably 3). This represents the number of charge / discharge cycles at the current predicted starting point. To integrate predicted sequences of health factors, .

[0074] To verify the effectiveness of the FHI constructed based on PCA in predicting lithium battery life, the following five health factor scenarios were set up for comparative prediction analysis: Scenario 1: Using the constant current charging time interval HI1 as model input to predict lithium battery life; Scenario 2: Using the constant voltage rise charging time interval HI2 as model input for lithium battery life prediction; Scenario 3: Using the equal voltage drop discharge time interval HI3 as model input to predict lithium battery life; Scenario 4: Using the maximum temperature rise time interval HI4 as model input for lithium battery life prediction; Scenario 5: Using the health factor FHI as model input for lithium battery life prediction.

[0075] The experimental subject was the B05 battery. First, the HI dataset, characterizing the capacity degradation of the B05 battery, was acquired. Then, SVMD multimodal decomposition was performed on the HI dataset to obtain residual components and IMF1~IMF3 components, totaling 168 data sets. Training and test sets were then created: the training set used data from the first 105 charge-discharge cycles, and the test set used data from the last 63 cycles. Next, an Attention-GRU neural network model was used to predict the above scenarios. Finally, the prediction results were analyzed. Figure 1 The prediction results of B05 battery in five scenarios are presented. The failure threshold of B05 battery is 70% of rated capacity, i.e., 1.4Ah. The model evaluation metrics are RMSE, MAPE and AE.

[0076] Depend on Figure 6 It can be seen that the B05 battery achieves good prediction results in all five scenarios, indicating that the health factors HI1, HI2, HI3, HI4, and FHI extracted in this invention can all replace battery capacity for lifespan prediction. The evaluation indicators obtained from the model predictions in each scenario are shown in Table 1 below. The actual RUL value of battery B05 is 19.

[0077] Table 1 Comparison of Evaluation Indicators

[0078] As shown in Table 1, the absolute error of remaining lifespan (AE) under the five scenarios is controlled within 2, MAPE is controlled within 1%, and RMSE is controlled within 0.015Ah. This indicates that the constant current charging time interval HI1, constant voltage rise charging time interval HI2, constant voltage drop discharging time interval HI3, and maximum temperature rise time interval HI4, as indirect health factors characterizing lithium battery capacity degradation, have good predictive performance in battery life prediction. However, the above individual health factors do not consider as many factors as FHI, which comprehensively considers the effects of voltage, current, and temperature. Therefore, FHI is more reasonable as the input for the life prediction model in this invention.

[0079] As shown in Table 2 below, the actual RUL of battery B05 is 19, that of battery B06 is 3, that of battery B07 is 20, and that of battery B18 is 14.

[0080] Table 2 shows the predicted remaining lifespan of the remaining batteries.

[0081] The RMSE predicted by the Attention-GRU neural network is 0.0091 Ah, and the root mean square error is less than 0.01 Ah.

[0082] The above provides a detailed description of the lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU provided by the present invention. Specific examples are used in this embodiment to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in these embodiments may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the remaining life of lithium batteries based on multi-health factor fusion and improved GRU, characterized in that, Includes the following steps: S1: Obtain multiple health factors of the lithium battery that are correlated with the actual battery capacity and meet preset conditions during the cyclic charging and discharging process; S2: Principal component analysis is used to perform dimensionality reduction and fusion processing on the screened health factors, and the principal components are extracted as fused health factors. S3: Adaptively decompose the fused health factor using continuous variational mode decomposition to obtain multiple sub-sequence components with different frequency domain characteristics; S4: Construct a gated recurrent unit network model that incorporates an attention mechanism, use the gated recurrent unit network model to predict each of the subsequence components, and integrate and reconstruct the prediction results to obtain the fused health factor prediction value; S5: Based on the predicted value of the fused health factor and the preset failure threshold, determine the remaining service life of the lithium battery.

2. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 1, characterized in that, The health factors include constant current charging time interval, constant voltage rise charging time interval, constant voltage drop discharging time interval, and maximum temperature rise time interval; wherein... The constant current charging time interval is the time taken for the lithium battery to reach the cutoff voltage during the constant current charging phase. The constant voltage rise charging time interval is the time it takes for the voltage of the lithium battery to rise from a first preset voltage value to a second preset voltage value during the charging process; The equal voltage drop discharge time interval is the time it takes for the voltage of the lithium battery to drop from the third preset voltage value to the fourth preset voltage value during the discharge process; The maximum temperature rise time interval is the time it takes for the internal temperature of the lithium battery to rise from the initial temperature to the highest temperature.

3. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 1, characterized in that, The steps in S1 to obtain health factors that meet preset conditions and are correlated with the actual battery capacity include: The correlation between each candidate factor and the actual battery capacity was calculated using Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient, respectively. Candidate factors whose three correlation coefficients all met the strong correlation threshold were selected as the health factors.

4. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 1, characterized in that, S2 includes the following steps: S21: Standardize the selected health factor data and construct a covariance matrix; S22: Calculate the cumulative contribution rate of each principal component in the covariance matrix; S23: Select the first principal component whose cumulative contribution rate reaches the preset contribution rate threshold as the fusion health factor, and discard the remaining principal components.

5. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 1, characterized in that, In S3, the subsequence components include at least a residual signal characterizing the global degradation trend, a low-frequency fluctuation signal characterizing the capacity regeneration phenomenon, and a high-frequency fluctuation signal characterizing random noise.

6. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 5, characterized in that, S3 includes the following steps: An iterative decomposition method is adopted, and the correlation coefficient between the currently extracted modal component and the original fused health factor is calculated after each decomposition. Modal components with correlation coefficients higher than a first preset threshold are retained. Calculate the ratio of the energy of the current residual signal to the energy of the original fused health factor sequence; Decomposition stops when the energy ratio falls below the second preset threshold.

7. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 6, characterized in that, In the iterative decomposition process, after each extraction of a modal component, the current residual signal is subtracted from the extracted modal component to update the residual signal, and the next extraction of modal components is performed based on the updated residual signal.

8. The lithium battery remaining life prediction method based on multi-health factor fusion and improved GRU according to claim 1, characterized in that, The gated recurrent unit network model that incorporates an attention mechanism includes an update gate, a reset gate, a candidate hidden state generation unit, and an attention gate. The reset gate is used to control the degree of influence of the previous hidden state on the candidate hidden state; The candidate hidden state generation unit is used to generate candidate hidden states based on the output of the reset gate and the input at the current moment; The attention gate is used to generate an attention weight vector based on the current input and the hidden state of the previous time step; The network model performs element-wise multiplication of the attention weight vector with the candidate hidden state to obtain the attention-enhanced candidate state, and then uses the update gate to perform weighted fusion of the hidden state at the previous time step and the attention-enhanced candidate state to obtain the final hidden state at the current time step.

Images