A lithium ion battery remaining useful life prediction method based on adaptive weighted fusion

Abstract

The application discloses a lithium ion battery remaining useful life prediction method based on adaptive weighted fusion. First, the battery aging cycle data is acquired and multi-dimensional health characteristics are extracted, the capacity degradation curve is fitted by using a Bacon-Watts model to locate the inflection point, and the whole life cycle is divided into a linear stage and a nonlinear stage. A two-stage collaborative predictor is constructed, an LSTM model is trained by using linear stage training set data, a BiGRU model is trained by using nonlinear stage training set data, and a cuckoo catfish optimizer (CCO) is used to optimize the BiGRU model parameters. A transformer gating network is established to mine the association between health characteristics and degradation stages and output stage probability weights. If the prediction accuracy of the two-stage collaborative predictor meets the requirements, the lithium ion battery remaining useful life RUL is output based on the weighted fusion result.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery remaining life prediction technology, and relates to a method for predicting the remaining life of lithium-ion batteries based on adaptive weighted fusion. Background Technology

[0002] With the widespread application of lithium-ion batteries in electric vehicles, energy storage systems, and portable electronic devices, their performance degradation has gradually attracted attention. During long-term charge-discharge cycles, the capacity and power performance of lithium-ion batteries will continuously decline due to factors such as electrode material aging, electrolyte decomposition, and changes in internal structure. Therefore, accurate prediction of the remaining service life (RUL) of lithium-ion batteries is of great significance for ensuring safe system operation, optimizing battery management strategies, and reducing maintenance costs.

[0003] However, current data-driven methods for predicting the remaining lifespan of lithium-ion batteries still have some shortcomings. In particular, the degradation process of lithium-ion batteries throughout their entire lifespan is often not a single trend, but rather exhibits a characteristic of gradual decline from a relatively stable initial stage to accelerated decline in the later stages. If only a single model is used for uniform prediction of the entire lifespan, it is often difficult to account for the characteristic differences of different degradation stages, thus affecting prediction accuracy and model generalization ability. Patent (CN114706006B) discloses a method for predicting the remaining lifespan of electric vehicle lithium-ion batteries based on an XGBoost-LSTM optimized model. This method extracts relevant feature parameters during the battery charging process and combines XGBoost and LSTM models to predict the remaining lifespan of lithium-ion batteries, improving prediction accuracy to some extent. However, this method mainly relies on a single prediction framework for modeling and does not adequately consider the stage-specific changes in the battery degradation process, making it difficult to establish more targeted prediction models for different degradation stages. Patent (CN115688415B) discloses a method and device for predicting the remaining lifespan of lithium-ion batteries that considers state switching. This method incorporates the switching process between the working state and storage state of the lithium-ion battery into the prediction of its remaining lifespan, and achieves battery life prediction through state transition and probabilistic models, providing some reference value for lifespan estimation under complex operating conditions. However, this method mainly focuses on the impact of the state switching process on lifespan evolution, and does not sufficiently explore the stage-based degradation characteristics reflected in the battery capacity degradation curve itself, and lacks targeted prediction structures for different degradation stages. This invention addresses the problems of insufficient consideration of the stage-based changes in battery capacity degradation and the inability of a single prediction model to adapt to the differences in characteristics of different degradation stages in existing lithium-ion battery remaining lifespan prediction methods, and proposes a lithium-ion battery remaining lifespan prediction method based on adaptive weighted fusion. Summary of the Invention

[0004] The problem solved by this invention is to provide a method for predicting the remaining service life of lithium-ion batteries based on adaptive weighted fusion, the main contents of which are as follows:

[0005] (1) The present invention constructs LSTM model and CCO-BiGRU model for different degradation stages, which can give full play to the modeling advantages of different models for different degradation laws and avoid the problem that a single model is difficult to take into account the prediction accuracy of the whole life cycle.

[0006] (2) The present invention uses a Transformer-based gated network to adaptively weight and fuse the prediction results at different stages. It can dynamically allocate weights according to the current degradation state of the input sample, thereby improving the robustness and generalization ability of the model prediction results.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the remaining service life of lithium-ion batteries based on adaptive weighted fusion, the specific steps of which are as follows:

[0008] S1: Obtain battery experimental datasets, including battery voltage, battery current, charge / discharge time, and capacity during aging cycle charge / discharge. Extract health features reflecting battery degradation in intervals. The extracted health features include charging time and charging energy within the constant current stage charging voltage interval and the constant voltage stage charging current interval.

[0009] S2: The extracted health feature data, namely the charging time and charging energy within the constant current stage charging voltage range and the constant voltage stage charging current range, are standardized using min-max.

[0010] S3: Use Spearman correlation analysis (SCA) and grey relational analysis (GRA) to jointly screen and reduce the dimensionality of the selected health characteristics;

[0011] S4: Using battery capacity data as input, the Bacon-Watts model is used to perform offline fitting of the capacity degradation curve of the historical training set. The intersection of two intersecting lines is calculated as the capacity degradation inflection point. Using the inflection point as the boundary, a corresponding stage label is generated for each cycle time in the dataset. The dataset with initial stage labels is then integrated.

[0012] S5: Take the dataset with initial stage labels output from step S4 as input, pass it through the two-stage collaborative predictor, and output the first RUL prediction value and the second RUL prediction value of the lithium battery.

[0013] The dual-stage collaborative predictor is designed as follows: The dual-stage collaborative predictor consists of a linear stage prediction module and a nonlinear stage prediction module connected in parallel. (1) For the linear stage prediction module, the training set data of the linear stage is used to train the Long Short-Term Memory Network (LSTM) and construct the LSTM prediction model. The test set data of the linear stage is input into the LSTM model to verify whether the accuracy of the LSTM model meets the design requirements. If the requirements are not met, the model is reconstructed. If the requirements are met, the first RUL prediction value of the lithium battery is output. (2) For the nonlinear stage prediction module, the training set data of the nonlinear stage is used to train the Bidirectional Gated Recurrent Unit Network (BiGRU). At the same time, the Cuckoo Catfish Optimizer (CCO) is introduced to globally optimize the learning rate, number of hidden layer nodes and regularization coefficient of the BiGRU network and construct the CCO-BiGRU prediction model. The test set data of the nonlinear stage is input into the CCO-BiGRU model to verify whether the accuracy of the CCO-BiGRU model meets the design requirements. If the requirements are not met, the model is reconstructed. If the requirements are met, the second RUL prediction value of the lithium battery is output.

[0014] S6: Using the dimensionality-reduced health features from step S3 as input, construct a gated network based on the Transformer architecture. Employ a multi-head self-attention mechanism to mine the mapping relationship between feature data and battery degradation stages. Use the stage-labeled dataset generated in step S4 as supervision information to complete training, outputting the probability weights of the battery's current linear degradation stage. Probability weights for the nonlinear accelerated decay phase ,satisfy ;

[0015] S7: Obtain the first RUL prediction value, the second RUL prediction value output in step S5, and the stage probability weight at the current moment output in step S6; use the current stage probability weight to perform a weighted summation of the first RUL prediction value and the second RUL prediction value to obtain the final RUL prediction value of the lithium battery.

[0016] Preferably, the health characteristics mentioned in step S1 include charging time and charging energy within the charging voltage range of the constant current stage and the charging current range of the constant voltage stage; in the constant current stage, the charging voltage range is divided into different voltage intervals with a variation interval of 0.05V, and the charging time and charging energy within each voltage interval are extracted as health characteristics; in the constant voltage stage, the charging current range is divided into different current intervals with a variation interval of 0.02A, and the charging time and charging energy within each current interval are extracted as health characteristics; in the constant current discharge stage, the peak value of the capacity increment (IC) curve and the voltage at its corresponding position are extracted as health characteristics.

[0017] Preferably, the standardization process for the health characteristic data in step S2 uses the min-max standardization method, and the calculation formula is as follows: In the formula, For health characteristic data, For the standardized data, max(x) and min(x) are the maximum and minimum values ​​in the corresponding health characteristic data.

[0018] Preferably, in step S3, Spearman correlation coefficient analysis (SCA) and grey relational analysis (GRA) are used to jointly screen and reduce the dimensionality of the health characteristic data. The specific steps are as follows: Given The total number of features of the observed samples, This represents the difference in rank between the i-th data pairs, and then the Spearman correlation coefficient is calculated. The formula is: After initial screening, the formula is then used to... ξi(k) is the grey relational coefficient between HF and SOH in the i-th cycle, k = 1, 2, ..., m, m is the number of battery cycles, y(k) is the SOH sequence, and xi(k) is the HF sequence. The resolution coefficient is set to 0.5. Then, by calculating the average value of the grey correlation coefficients, the grey correlation degree can be obtained, using the following formula: , The value range is [0,1]. The closer the value is to 1, the stronger the correlation between the reference sequence and the comparison sequence. Thus, the health features after secondary dimensionality reduction are obtained. Finally, the feature values ​​with Spearman correlation coefficient and grey relational coefficient both reaching 0.95 or higher are taken as the new health features.

[0019] Preferably, in step S4, the Bacon-Watts model is used to perform offline fitting on all cycle capacity data from the start of the first charge-discharge cycle to the end of the battery's lifespan (EOL) to accurately pinpoint the inflection point of capacity degradation. Assuming the number of charge-discharge cycles is x and the corresponding discharge capacity is y, the Bacon-Watts model introduces a hyperbolic tangent function tanh to smoothly connect the two linear stages. The fitting function is defined as: ,in This indicates the number of iterations at the inflection point. For smoothing parameters, For random error term, , , Regression coefficients. The slope of the left-hand linear gradual transition phase is... The slope of the nonlinear accelerated decay phase on the right is The optimal inflection point value is obtained by estimating the above parameters using the least squares method. Based on this inflection point value Generate stage labels for each time step t in the dataset. If the current loop count is... ,but This indicates that it belongs to the linear decay phase; if ,but This indicates that it belongs to the non-linear accelerated decay stage.

[0020] Preferably, in step S5, the two-stage collaborative predictor is designed as follows: For the linear stage prediction module, a Long Short-Term Memory (LSTM) network is used to process the linear decay stage data. The internal structure of the LSTM consists of three parts: a forget gate, an input gate, and an output gate, thereby achieving information protection and control. The specific steps of the LSTM are: The first step is to add a forget gate... The formula determines which information should be discarded from the cell state. In the formula, For the Sigmoid function, and Here are the weight matrix and bias terms for the forget gate. The output is the cell state at time t-1. The first step is to input the current cell state at time t; the second step is to decide which information to store in the internal state. This step consists of two parts, the first part being determined by the input gate. Decide which information needs to be updated, among which In the formula, and The second part consists of the weight matrix and bias terms of the input gate. Create a candidate vector ,in In the formula, and for The weight matrix and bias terms, and the current cell state. The updated formula is In the formula The cell state at time t-1; finally, the updated cell state and the output gate... The layer determines the output of the LSTM, as shown in the equation. , , The output of the LSTM at time t. The Sigmoid layer for the output gate. and The weight matrix and bias terms are for the output layer. Test set data from the linear stage are input into the LSTM model to verify if the LSTM model accuracy meets the design requirements. If it does not meet the requirements, the model is rebuilt; if it does, the first RUL prediction value of the lithium battery is output. For the nonlinear stage module, a bidirectional gated recurrent unit (BiGRU) is used to process the nonlinear accelerated decay stage data. The BiGRU consists of a forward GRU and a backward GRU, capable of simultaneously capturing past and future contextual information of the sequence data. The forward hidden state is calculated as follows: The backward hidden state is calculated as follows ,in and These represent the processing functions of the forward and backward GRU units, respectively. The input at time t, and These are the hidden states at the previous and next time steps, respectively. Finally, the output state at time t. It is composed of the forward and backward hidden states, that is, the output state is the concatenation of the two. To address the difficulty of parameter optimization in the later stages, a Cuckoo Catfish Optimizer (CCO) is introduced to optimize the learning rate of the BiGRU. Number of hidden layer nodes and regularization coefficient Optimization is performed to obtain the optimal CCO-BiGRU prediction model. Test set data from the nonlinear stage is input into the CCO-BiGRU model to verify whether the accuracy of the CCO-BiGRU model meets the design requirements. If it does not meet the requirements, the model is rebuilt. If it meets the requirements, the second RUL prediction value of the lithium battery is output.

[0021] Preferably, in step S6, a Transformer gated network is constructed to output stage probability weights, and the network is trained using the stage labels generated in step S4 as supervision information. The multidimensional health feature sequence extracted in step S3 is then used. As input, the query matrix is ​​first generated through linear projection. Key matrix Sum matrix In the formula , , Let be a learnable projection weight matrix. A multi-head self-attention mechanism is used to mine deep correlations between features. The formula for calculating individual attention points is as follows: ,in The scaling factor is used. The outputs of all attention heads are concatenated and subjected to a linear transformation to obtain... Then, after layer normalization and a feedforward neural network layer, i.e. as well as Finally, the output is mapped to a probability distribution using a fully connected layer and a softmax activation function. , , These are the weight matrix and bias term of the output layer, respectively. Output result. , respectively, represent the probability weights of the battery being in the linear degradation phase at the current moment. and probability weights in the nonlinear accelerated decay phase And satisfy Calculate the cross-entropy loss function to update the gradient.

[0022] Preferably, in step S7, the first initial value of the RUL prediction output in step S5 is obtained. Second RUL prediction initial value And the current time-stage probability weights output by S6. The two initial RUL prediction values ​​are adaptively weighted and summed using the aforementioned stage probability weights. The specific calculation formula is as follows: This yields the final predicted value of the remaining lifespan of the lithium-ion battery.

[0023] Compared with existing technologies, this invention has the following technical advantages: It accurately captures the capacity degradation inflection point using the Bacon-Watts model, scientifically dividing the entire lifespan into linear and nonlinear stages, thus overcoming the limitation of traditional models in failing to account for the differences in battery physical degradation characteristics. By introducing a Transformer gated network and utilizing a self-attention mechanism to achieve online adaptive perception of the degradation stage, and by adaptively fusing the dual-branch model through output dynamic probability weights, it overcomes the technical pain point of abrupt transitions at switching points in traditional segmented prediction methods, improving the smoothness and accuracy of the entire lifespan prediction. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0025] In the attached diagram:

[0026] Figure 1 This is a flowchart of the method for predicting the remaining service life of lithium-ion batteries according to the present invention;

[0027] Figure 2 This is a schematic diagram of the LSTM structure in this invention;

[0028] Figure 3 This is a schematic diagram of the BiGRU structure in this invention. Detailed Implementation

[0029] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0030] Example: Figure 1 As shown, a method for predicting the remaining lifespan of lithium-ion batteries based on adaptive weighted fusion is described, with the following specific steps:

[0031] S1: Obtain voltage, current, and time data from the battery experiment dataset during the aging cycle charge-discharge process. Extract health features reflecting battery degradation by segment. The battery charge-discharge experiment process is as follows: First, charge with a constant current (CC) of 1.5A until the voltage rises to 4.3V, then charge with a constant voltage (CV) until the current drops to 15mA. During the discharge process, keep the current constant at 2A. Discharge ends when the battery voltage drops from 4.3V to the cutoff voltage. The extracted health features are the charging time and charging energy within the CC stage charging voltage range of 3.7V to 4.3V, denoted as HF1 and HF2, and the charging time and charging energy within the CV stage charging current range of 1.5A to 0.5A. Divide the CC stage charging voltage range into different voltage segments with a variation interval of 0.05V, and divide the CV stage charging current range into different current segments with a variation interval of 0.02A. Extract the charging time and charging energy within each voltage and current segment as health features, denoted as HF3 and HF4. During the CC phase, the charging range is from 3.5V to 4.3V, with a variation interval of 0.01V. The peak value of the capacity increment (IC) curve and the voltage at the corresponding position are extracted as health characteristics, denoted as HF5 and HF6.

[0032] S2: The min-max standardization method is used, and the calculation formula is as follows: In the formula, For health characteristic data, For the standardized data, X max and X min These represent the maximum and minimum values ​​in the corresponding health characteristic data.

[0033] S3: The Spearman correlation coefficient analysis (SCA) and grey relational analysis (GRA) methods are used to jointly screen and reduce the dimensionality of health characteristic data. The specific steps are as follows: Given... The total number of features of the observed samples, This represents the difference in rank between the i-th data pairs, and then the Spearman correlation coefficient is calculated. The formula is: After initial screening, the formula is then used to... , where ξ iy(k) is the grey relational coefficient between HF and SOH in the i-th cycle, k = 1, 2, ..., m, where m is the number of battery cycles, y(k) is the SOH sequence, and x... i (k) is an HF sequence. The resolution coefficient is set to 0.5. Then, by calculating the average value of the grey correlation coefficients, the grey correlation degree can be obtained, using the following formula: , The value range is [0,1]. The closer the value is to 1, the stronger the correlation between the reference sequence and the comparison sequence. Thus, the health features after secondary dimensionality reduction are obtained. Finally, the feature values ​​with Spearman correlation coefficient and grey relational coefficient both reaching 0.95 or higher are taken as the new health features.

[0034] S4: Use the Bacon-Watts model to fit the capacity decay curve to locate the inflection point. The specific steps are as follows:

[0035] (1) Construct a capacity degradation dataset, with the input being a sequence of cycle counts for the entire battery lifespan. and the corresponding discharge capacity sequence in, This indicates the total number of cycles required for the battery to degrade to failure.

[0036] (2) A Bacon-Watts piecewise linear regression mathematical model was established. In order to scientifically identify the abrupt change in the battery from linear stable degradation to nonlinear accelerated degradation, the following mathematical expression was used to fit the capacity sequence:

[0037]

[0038] in, This represents the predicted capacity value for the k-th iteration. The capacity intercept at the inflection point; The linear slope represents the initial stage of degradation. This represents the change in slope during the later stages of degradation. This is the capacity decay inflection point (Knee-point) to be solved. To control the parameters for a smooth transition, they are usually taken as very small positive numbers to simulate step characteristics; For fitting residuals;

[0039] (3) Parameter estimation and optimization: Non-linear least squares is used as the optimization objective, and the objective function is minimized. To obtain the optimal parameter set

[0040]

[0041] In the process of seeking the best The value range of is limited to [1, n]. The output optimal fitting parameters This refers to the physical segmentation points of the battery's entire lifespan;

[0042] (4) Stage division and label generation based on inflection points, based on the solved inflection points The entire lifecycle data is divided into two intervals, and supervision labels are generated for the subsequent training of the Transformer gated network. If the current loop count is... ,but This indicates that it belongs to the linear decay phase; if ,but This indicates that it belongs to the non-linear accelerated decay stage.

[0043] S5: The two-stage collaborative predictor is designed as follows:

[0044] For the linear stage prediction module, a Long Short-Term Memory (LSTM) network is used to process the linear decay stage data. The internal structure of the LSTM consists of three parts: a forget gate, an input gate, and an output gate, which achieves information protection and control. The specific steps of the LSTM are as follows: The first step is to add a forget gate... The formula determines which information should be discarded from the cell state. In the formula, For the Sigmoid function, and Here are the weight matrix and bias terms for the forget gate. The output is the cell state at time t-1. The first step is to input the current cell state at time t; the second step is to decide which information to store in the internal state. This step consists of two parts, the first part being determined by the input gate. Decide which information needs to be updated, among which In the formula, and The second part consists of the weight matrix and bias terms of the input gate. Create a candidate vector ,in In the formula, and for The weight matrix and bias terms, and the current cell state. The updated formula is In the formula The cell state at time t-1; finally, the updated cell state and the output gate... The layer determines the output of the LSTM, as shown in the equation. , , The output of the LSTM at time t. The Sigmoid layer for the output gate. and Here are the weight matrix and bias terms for the output layer. The test set data from the linear stage is input into the LSTM model. Mean absolute error (MAE) and root mean square error (RMSE) are selected as evaluation metrics to assess the prediction accuracy of the LSTM model, thereby determining whether the expected results are met. The specific formula is as follows: , In the formula, N represents the estimated number of experiments. This is the actual RUL value. This is the estimated RUL value.

[0045] For the nonlinear phase prediction module, a bidirectional gated recurrent unit (BiGRU) is used to process the nonlinear accelerated decay phase data. The BiGRU consists of a forward GRU and a backward GRU, capable of simultaneously capturing past and future contextual information of the sequence data. The forward hidden state is calculated as follows: The backward hidden state is calculated as follows ,in and These represent the processing functions of the forward and backward GRU units, respectively. The input at time t, and These are the hidden states at the previous and next time steps, respectively. Finally, the output state at time t. It is composed of the forward and backward hidden states, that is, the output state is the concatenation of the two. To address the difficulty of parameter optimization in the later stages, a Cuckoo Catfish Optimizer (CCO) is introduced to optimize the learning rate of the BiGRU. Number of hidden layer nodes and regularization coefficient Optimization is performed to obtain the optimal CCO-BiGRU prediction model. The CCO algorithm iteratively optimizes the model by simulating the search, encirclement, and chaotic predation behavior of the cuckoo catfish. During the exploration phase, a multi-dimensional encirclement search and space compression strategy is used to effectively limit the search space. The shrinkage efficiency of the encirclement is controlled by a nonlinear shrinkage factor T, which is calculated using the following formula: ,in This represents the current iteration number. This factor represents the maximum number of iterations. It ensures slower contraction in the early stages of the algorithm (exploration phase), maintaining global search capability, while accelerating contraction in the later stages (development phase), improving local optimization accuracy. Test set data from the nonlinear phase is input into the CCO-BiGRU model. Mean absolute error (MAE) and root mean square error (RMSE) are selected as evaluation metrics to assess the prediction accuracy of the CCO-BiGRU model, thereby determining whether the expected results are met. The specific formula is as follows: , In the formula, N represents the estimated number of experiments. This is the actual RUL value. This is the estimated RUL value.

[0046] S6: Construct a Transformer gated network to output stage probability weights, and train the network using the stage labels generated in step S4 as supervision information. The multidimensional health feature sequence extracted in step S3... As input, the query matrix is ​​first generated through linear projection. Key matrix Sum matrix In the formula , , Let be a learnable projection weight matrix. A multi-head self-attention mechanism is used to mine deep correlations between features. The formula for calculating individual attention points is as follows: ,in The scaling factor is used. The outputs of all attention heads are concatenated and subjected to a linear transformation to obtain... Then, after layer normalization and a feedforward neural network layer, i.e. as well as Finally, the output is mapped to a probability distribution using a fully connected layer and a softmax activation function. , , These are the weight matrix and bias term of the output layer, respectively. Output result. , respectively, represent the probability weights of the battery being in the linear degradation phase at the current moment. and probability weights in the nonlinear accelerated decay phase And satisfy Calculate the cross-entropy loss function to update the gradient.

[0047] S7: Obtain the first initial value of the RUL prediction output in step S5. Second RUL prediction initial value And the current time-stage probability weights output by S6. The two initial RUL prediction values ​​are adaptively weighted and summed using the aforementioned stage probability weights. The specific calculation formula is as follows: This yields the final predicted value of the remaining lifespan of the lithium-ion battery.

[0048] Finally, it should be noted that the above descriptions are merely examples of the present invention and are not intended to limit the invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining useful life (RUL) of a lithium-ion battery based on adaptive weighted fusion, the RUL method comprising the following steps: S1: Obtain battery experimental datasets, including battery voltage, battery current, charge / discharge time, and capacity during aging cycle charge / discharge. Extract health features reflecting battery degradation in intervals. The extracted health features include charging time and charging energy within the constant current stage charging voltage interval and the constant voltage stage charging current interval. S2: The extracted health feature data, including the charging time and charging energy within the constant current stage charging voltage range and the constant voltage stage charging current range, are standardized using min-max. S3: Use Spearman correlation analysis (SCA) and grey relational analysis (GRA) to jointly screen and reduce the dimensionality of the selected health characteristics; S4: Using battery capacity data as input, the Bacon-Watts model is used to perform offline fitting of the capacity degradation curve of the historical training set. The intersection of two intersecting lines is calculated as the capacity degradation inflection point. Using the inflection point as the boundary, a corresponding stage label is generated for each cycle time in the dataset. The dataset with initial stage labels is then integrated. S5: Take the dataset with initial stage labels output from step S4 as input, pass it through the two-stage collaborative predictor, and output the first RUL prediction value and the second RUL prediction value of the lithium battery. The design of the two-stage collaborative predictor is as follows: The two-stage collaborative predictor consists of a linear stage prediction module and a nonlinear stage prediction module connected in parallel. (1) For the linear stage prediction module, the training set data of the linear stage is used to train the long short-term memory network (LSTM) and construct the LSTM prediction model; the test set data of the linear stage is input into the LSTM model to verify whether the accuracy of the LSTM model meets the design requirements. If the requirements are not met, the model is rebuilt. If the requirements are met, the first RUL prediction value of the lithium battery is output. (2) For the nonlinear stage prediction module, the training set data of the nonlinear stage is used to train the bidirectional gated recurrent unit network (BiGRU). At the same time, the cuckoo catfish optimizer (CCO) is introduced to globally optimize the learning rate, number of hidden layer nodes and regularization coefficient of the BiGRU network and build the CCO-BiGRU prediction model. The test set data of the nonlinear stage is input into the CCO-BiGRU model to verify whether the accuracy of the CCO-BiGRU model meets the design requirements. If the requirements are not met, the model will be rebuilt; if the requirements are met, the second RUL prediction value of the lithium battery will be output. S6: Using the dimensionality-reduced health features from step S3 as input, construct a gated network based on the Transformer architecture. Employ a multi-head self-attention mechanism to mine the mapping relationship between feature data and battery degradation stages. Use the stage-labeled dataset generated in step S4 as supervision information to complete training, outputting the probability weights of the battery's current linear degradation stage. Probability weights for the nonlinear accelerated decay phase ,satisfy ; S7: Obtain the first RUL prediction value, the second RUL prediction value output in step S5, and the stage probability weight at the current moment output in step S6; use the current stage probability weight to perform a weighted summation of the first RUL prediction value and the second RUL prediction value to obtain the final RUL prediction value of the lithium battery.

2. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, characterized in that, In step S3, Spearman correlation coefficient analysis (SCA) and grey relational analysis (GRA) are used to jointly screen and reduce the dimensionality of health characteristic data. The specific steps are as follows: Given The total number of features of the observed samples, This represents the difference in rank between the i-th data pairs, and then the Spearman correlation coefficient is calculated. The formula is: After initial screening, the formula is then used to further refine the selection. ξi(k) is the grey relational coefficient between HF and SOH in the i-th cycle, k = 1, 2, ..., m, m is the number of battery cycles, y(k) is the SOH sequence, and xi(k) is the HF sequence. The resolution coefficient is set to 0.

5. Then, by calculating the average value of the grey correlation coefficients, the grey correlation degree can be obtained, using the following formula: , The value range is [0,1]. The closer the value is to 1, the stronger the correlation between the reference sequence and the comparison sequence. Thus, the health features after secondary dimensionality reduction are obtained. Finally, the feature values ​​with Spearman correlation coefficient and grey relational coefficient both reaching 0.95 or higher are taken as the new health features.

3. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, characterized in that, In step S4, the Bacon-Watts model is used to perform offline fitting of the capacity degradation curves in the historical training set to accurately locate the inflection point of capacity degradation. Assuming the number of charge-discharge cycles of the battery is x, and the corresponding discharge capacity is y, the Bacon-Watts model introduces the hyperbolic tangent function tanh to smoothly connect the two linear stages. The fitting function is defined as: ,in This indicates the number of iterations at the inflection point. For smoothing parameters, For random error term, , , The regression coefficients, with the slope of the left-hand linear gradual phase being... The slope of the nonlinear accelerated decay phase on the right is The optimal inflection point value is obtained by estimating the above parameters using the least squares method. Based on this inflection point value Generate stage labels for each time step t in the dataset. If the current loop count ,but This indicates that it belongs to the linear decay phase; if ,but This indicates that it belongs to the non-linear accelerated decay stage.

4. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, characterized in that, In step S5, the two-stage collaborative predictor is designed as follows: For the linear stage prediction module, a Long Short-Term Memory (LSTM) network is used to process the linear stage data. The internal structure of LSTM consists of three parts: a forget gate, an input gate, and an output gate. The specific steps of LSTM are as follows: The first step is to add a forget gate... The formula determines which information should be discarded from the cell state. In the formula, For the Sigmoid function, and Here are the weight matrix and bias terms for the forget gate. The output is the cell state at time t-1. The first step is to input the current cell state at time t; the second step is to decide which information to store in the internal state. This step consists of two parts, the first part being determined by the input gate. Decide which information needs to be updated, among which In the formula, and The second part consists of the weight matrix and bias terms of the input gate. Create a candidate vector ,in In the formula, and for The weight matrix and bias terms, and the current cell state. The updated formula is In the formula The cell state at time t-1; finally, the updated cell state and the output gate... The layer determines the output of the LSTM, as shown in the equation. , , The output of the LSTM at time t. The Sigmoid layer for the output gate. and The weight matrix and bias terms of the output layer are defined; the test set data of the linear stage is input into the LSTM model to verify whether the accuracy of the LSTM model meets the design requirements. If the requirements are not met, the model will be rebuilt; if the requirements are met, the first RUL prediction value of the lithium battery will be output. For the nonlinear stage prediction module, a bidirectional gated recurrent unit network (BiGRU) is used to process the nonlinear accelerated decay stage data. The BiGRU consists of a forward GRU and a backward GRU, capable of simultaneously capturing past and future contextual information of the sequence data. The forward hidden state is calculated as follows: The backward hidden state is calculated as follows ,in and These represent the processing functions of the forward and backward GRU units, respectively. The input at time t, and These are the hidden states at the previous and next time steps, respectively. Finally, the output state at time t. It is composed of the forward and backward hidden states, that is, the output state is the concatenation of the two. To address the difficulty of parameter optimization in the later stages, a Cuckoo Catfish Optimizer (CCO) is introduced to optimize the learning rate of the BiGRU. Number of hidden layer nodes and regularization coefficient Optimize the model to obtain the optimal CCO-BiGRU prediction model; input the test set data of the nonlinear stage into the CCO-BiGRU model to verify whether the accuracy of the CCO-BiGRU model meets the design requirements. If the requirements are not met, the model will be rebuilt; if the requirements are met, the second RUL prediction value for the lithium battery will be output.

5. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, characterized in that, In step S6, a Transformer gated network is constructed to output stage probability weights. Training is completed using the stage-labeled dataset generated in step S4 as supervision information, and the multidimensional health feature sequence extracted in step S3 is used. As input, the query matrix is ​​first generated through linear projection. Key matrix Sum matrix In the formula , , The projected weight matrix is ​​learnable; a multi-head self-attention mechanism is used to mine deep correlations between features, the th... The formula for calculating individual attention points is as follows: ,in The scaling factor is used; the outputs of all attention heads are concatenated and linearly transformed to obtain... Then, after layer normalization and feedforward neural network layers, i.e. as well as Finally, the output is mapped to a probability distribution using a fully connected layer and a softmax activation function. , , These are the weight matrix and bias term of the output layer, respectively, and the output result is... , respectively, represent the probability weights of the battery being in the linear degradation phase at the current moment. and probability weights in the nonlinear accelerated decay phase And satisfy Calculate the cross-entropy loss function to update the gradient.

6. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, characterized in that, In step S7, the first RUL prediction value output in step S5 is obtained. Second RUL prediction value and the current time stage probability weight output in step S6 The two initial RUL prediction values ​​are weighted and summed using the aforementioned stage probability weights. The specific calculation formula is as follows: This yields the final predicted value of the remaining lifespan of the lithium-ion battery.

Images