Lithium-ion battery aging mode estimation method based on domain knowledge guidance

Abstract

The application particularly relates to a lithium ion battery aging mode estimation method based on domain knowledge guidance, which comprises the following steps: S1, acquiring charging cycle data and low-rate pseudo open circuit voltage test data; S2, constructing a capacity-voltage sequence; S3, extracting an incremental capacity peak feature parameter of the capacity-voltage sequence of each cycle; S4, calculating positive electrode active material loss, negative electrode active material loss and circulatable lithium loss of each cycle to construct an aging mode label; S5, pre-training a shared feature extractor; S6, sequentially performing regressioner fine-tuning based on a frozen pre-training shared feature extractor and low learning rate global fine-tuning based on a thawed pre-training shared feature extractor to obtain a trained domain knowledge guided aging mode estimation model; and S7, inputting a capacity-voltage sequence of a battery to be estimated into the trained domain knowledge guided aging mode estimation model to output an aging mode estimation result. The application improves the accuracy and generalization of electrode-level aging mode estimation.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery health diagnosis and battery management technology, specifically to a method for estimating the aging mode of lithium-ion batteries based on domain knowledge guidance. Background Technology

[0002] Lithium-ion batteries inevitably experience capacity decay and internal resistance increase during long-term cycling and storage, which reduces the operational reliability of electric vehicles and energy storage systems and poses safety risks. The State of Health (SOH) of the battery can describe the overall degree of degradation, but it is difficult to explain the source of capacity loss at the electrode level. The mechanisms of solid electrolyte interfacial film growth, electrode cracking, electrolyte decomposition, lithium plating, and current collector corrosion within the battery can generally be summarized as the loss of capacitance (LAM) of the positive electrode active material. PE Loss of negative electrode active material LAM NE The three main aging modes are lithium loss (LLI) and recyclable lithium loss. Compared to a single SOH index, these aging modes provide more granular and mechanistically interpretable health information.

[0003] Existing aging mode labels are typically obtained through methods such as post-disassembly scanning or energy dispersive spectroscopy (EDS), low-rate pseudo-open-circuit voltage (POV) testing combined with half-cell model identification. Material-level testing is destructive and costly, and low-rate diagnostic testing is time-consuming; neither is suitable for online diagnostics or large-scale cycle-by-cycle labeling. Therefore, high-quality LAMs are needed. PE LAM NE LLI tags are usually very scarce. In contrast, conventional 1C charging data is easy to collect, and the capacity-voltage Q(V) curve also includes morphological changes related to aging. However, it is affected by ohmic voltage drop, polarization and mass transfer kinetics. Moreover, the positive and negative electrode reactions are coupled in the full cell curve. Different aging modes may produce similar or overlapping curve responses, making direct regression difficult.

[0004] Existing technologies primarily generate synthetic data through mechanistic models or perform end-to-end supervised learning using a small number of labels. The former relies on model accuracy and parameter coverage, and there may be discrepancies between the simulated distribution and the real battery. The latter requires the network to simultaneously extract local morphology from high-dimensional Q(V) sequences, form aging characterizations, and learn nonlinear mappings of multiple aging modes, which is prone to overfitting and characterization drift when the label ratio drops to 10% or even lower. Existing incremental capacity analysis can reflect peak area decay, peak position shift, and peak width changes, but direct numerical differentiation is sensitive to sampling noise and filtering parameters, and handcrafted features do not fully utilize the complete Q(V) waveform. Therefore, it is necessary to provide a technical solution that can extract stable domain knowledge from real conventional charging curves and complete electrode-level aging mode estimation with very few aging mode labels. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the technical problem this invention aims to solve is: how to provide a domain knowledge-guided method for estimating the aging modes of lithium-ion batteries. This method pre-trains a shared feature extractor by extracting the incremental capacity peak area, peak position, and peak width as domain knowledge supervision signals. It also incorporates a small number of labels generated by aligning low-rate pseudo-open-circuit voltage curves with positive and negative electrode potential curves to identify positive electrode active material loss, negative electrode active material loss, and cyclic lithium loss. The method performs regressor fine-tuning based on freezing the pre-trained shared feature extractor and low-learning-rate global fine-tuning based on unfreezing the shared feature extractor, achieving continuous estimation of the three aging modes. This reduces the reliance on low-rate diagnostic tests and high-quality aging mode labels, alleviates characterization drift and overfitting problems under label scarcity conditions, and improves the accuracy and generalization of electrode-level aging mode estimation.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] Domain knowledge-guided methods for estimating lithium-ion battery aging modes include:

[0008] S1: Obtain charging cycle data and low-rate pseudo-open circuit voltage test data for each cycle of lithium-ion battery during conventional 1C constant current charging.

[0009] S2: Construct the capacity-voltage sequence corresponding to each cycle based on the charging cycle data of each cycle;

[0010] S3: Extract the incremental capacity peak feature parameters of the capacity-voltage sequence of each cycle, and construct a domain knowledge dataset based on the capacity-voltage sequence of each cycle and the corresponding incremental capacity peak feature parameters;

[0011] S4: Alignment and identification are performed based on the low-rate pseudo-open-circuit voltage test data of each cycle and the corresponding positive and negative electrode potential curves. The loss of positive active material, negative active material and cyclic lithium of each cycle are calculated to construct aging mode labels. An aging mode dataset is constructed based on the capacity-voltage sequence of each cycle and the corresponding aging mode labels.

[0012] S5: Pre-train the shared feature extractor using a domain knowledge dataset to obtain the pre-trained shared feature extractor;

[0013] S6: By sequentially performing regressor fine-tuning based on frozen pre-trained shared feature extractor and low-learning-rate global fine-tuning based on unfrozen pre-trained shared feature extractor on the aging pattern dataset, a domain knowledge-guided aging pattern estimation model is obtained after training.

[0014] S7: Input the capacity-voltage sequence of the battery to be estimated into the trained domain knowledge to guide the aging mode estimation model. Through forward propagation calculation, output the aging mode estimation results including positive electrode active material loss, negative electrode active material loss and cyclic lithium loss.

[0015] Preferably, in step S2, the process of constructing the capacity-voltage sequence includes:

[0016] S201: Obtain the voltage at each time point in the corresponding loop. and current Charging cycle data;

[0017] S202: At the lower cutoff voltage Upper cutoff voltage Within the range, the capacity value corresponding to each voltage point is calculated using ampere-hour integration to construct an initial capacity-voltage curve;

[0018] Initial capacity-voltage curve The calculation is expressed as:

[0019] ;

[0020] S203: Discretize the voltage axis of the initial capacity-voltage curve using a preset voltage interval to form a uniform voltage sampling grid. ,in = , = The capacity values ​​on the initial capacity-voltage curve are mapped to a uniform voltage sampling grid using an interpolation algorithm, generating a capacity-voltage sequence with uniform dimensions that preserves local charging patterns. .

[0021] Preferably, in step S3, the processing steps for extracting the incremental capacity peak characteristic parameters of the capacity-voltage sequence include:

[0022] S301: Building based on A parameterized capacity-voltage model with superposition of arctangent terms;

[0023] The formula is expressed as:

[0024] ;

[0025] In the formula: , These are the capacity value and the voltage value, respectively. , and The first The peak area, peak center location, and peak width of each peak. For constant terms;

[0026] S302: Convert the corresponding cycle's capacity-voltage sequence Input a parameterized capacity-voltage model, and minimize the actual capacity-voltage sequence using a nonlinear least squares method. Fit the capacity-voltage curve to the model The error between them is used to identify the model parameters and obtain the peak area parameters corresponding to each arctangent term. Peak center position parameters Peak width Parameters and constants The peak area parameter, peak center position parameter, and peak width parameter are used as candidate incremental capacity peak feature parameters for the corresponding cycle.

[0027] The formula is expressed as:

[0028] ;

[0029] In the formula: , and They represent the first to the last. The peak area parameter, peak center position parameter, and peak width parameter corresponding to each arctangent term. The number of arctangent terms in the parameterized capacity-voltage model; This represents the constant term in the parameterized capacity-voltage model; This indicates the number of voltage sampling points in the capacity-voltage sequence; Indicates the first The actual capacity value corresponding to each voltage sampling point; The first one obtained based on the identification parameters The fitted capacity value corresponding to each voltage sampling point;

[0030] S303: Calculate the Pearson correlation coefficient between the characteristic parameters of each candidate incremental capacity peak and the corresponding losses of positive electrode active material, negative electrode active material, and cyclic lithium in the lithium-ion battery.

[0031] S304: Based on the Pearson correlation coefficient results, and combined with the stability of the candidate incremental capacity peak feature parameters during the cycling process and the complementarity of the representation information between different candidate incremental capacity peak feature parameters, the candidate incremental capacity peak feature parameters are screened to obtain incremental capacity peak feature parameters that can characterize the battery degradation state, and a domain knowledge dataset is constructed based on the incremental capacity peak feature parameters.

[0032] Preferably, in step S304, the domain knowledge dataset Represented as:

[0033] ;

[0034] In the formula: , The first Capacity-voltage sequence and incremental capacity peak characteristic parameters of each sample; This represents the number of samples in the domain knowledge dataset.

[0035] Preferably, in step S4, the process of constructing aging mode labels includes:

[0036] S401: Construct a low-rate pseudo-open-circuit voltage curve based on the low-rate pseudo-open-circuit voltage test data of the corresponding cycle;

[0037] S402: Introducing positive electrode capacity scaling parameters, negative electrode capacity scaling parameters, positive electrode capacity translation parameters, and negative electrode capacity translation parameters, the capacity coordinates of the standard positive and negative electrode potential curves are mapped to normalized coordinates of the two electrodes through scaling and translation. After subtraction, the open-circuit voltage curve of the full cell is reconstructed. ;

[0038] Full cell open circuit voltage curve The calculation formula is:

[0039] ;

[0040] ;

[0041] ;

[0042] In the formula: ; and These are the positive electrode potential curve and the negative electrode potential curve, respectively. , These are the positive electrode capacity scaling parameters and the negative electrode capacity scaling parameters, respectively. , These are the positive electrode capacity shift parameters and the negative electrode capacity shift parameters, respectively. Coordinates for total battery capacity; Capacity coordinates The normalized capacity coordinates of the positive and negative electrodes are obtained after scaling and translation mapping of the positive and negative electrode capacities;

[0043] S403: The low-rate pseudo-open-circuit voltage curve of the corresponding cycle. Discretized After sampling each capacity point, the open-circuit voltage curve of the reconstructed full cell is compared with... By performing point-by-point comparisons, the optimal alignment parameters for the corresponding loop are obtained by minimizing the mean square error between the measured and reconstructed values. ;

[0044] The formula is expressed as:

[0045] ;

[0046] In the formula: , The low-rate pseudo-open-circuit voltage curve and the full-cell open-circuit voltage curve are respectively shown in Capacity values ​​at each capacity sampling point;

[0047] S404: Obtain the positive electrode capacity scaling parameters obtained from the earliest characterization cycle identification of the corresponding lithium-ion battery. Negative electrode capacity scaling parameters Positive electrode capacity shift parameters and negative electrode capacity shift parameters This is the baseline state;

[0048] S405: Based on the baseline state, combined with the optimal alignment parameters Positive electrode capacity scaling parameters and negative electrode capacity scaling parameters Calculate the loss of positive electrode active material and loss of negative electrode active material ;

[0049] The calculation formula is:

[0050] ;

[0051] ;

[0052] S406: Based on the baseline state, combined with the optimal alignment parameters Positive electrode capacity scaling parameters Positive electrode capacity shift parameters and negative electrode capacity shift parameters Calculate the cyclic lithium capacity parameters of the baseline cycle And the current cycle's recyclable lithium quantity parameters And through the cyclic lithium quantity parameters of the benchmark cycle And the current cycle's recyclable lithium quantity parameters Calculate the loss of recyclable lithium ;

[0053] The calculation formula is:

[0054] ;

[0055] ;

[0056] ;

[0057] S407: Loss of positive electrode active material Loss of negative electrode active material and recyclable lithium loss As the aging mode label for the corresponding cycle.

[0058] Preferably, in step S4, the aging pattern dataset includes aging pattern labels. Represented as:

[0059] ;

[0060] In the formula: , The first Capacity-voltage sequence and aging pattern labeling for each sample; , , The first The loss of positive electrode active material, the loss of negative electrode active material, and the loss of cyclic lithium for each sample; This represents the number of samples in the aging pattern dataset.

[0061] Preferably, in step S5, the shared feature extractor comprises three cascaded one-dimensional convolutional blocks, a global average pooling layer, a fully connected layer, and a Dropout layer.

[0062] The first one-dimensional convolutional block contains two cascaded one-dimensional convolutional layers and a max pooling layer;

[0063] The second one-dimensional convolutional block contains two cascaded one-dimensional convolutional layers and a max pooling layer;

[0064] The third one-dimensional convolutional block contains two one-dimensional convolutional layers, each of which is connected sequentially to a layer normalization function and a ReLU activation function.

[0065] Preferably, in step S5, the pre-training process of the shared feature extractor and the incremental capacity peak feature regressor using the domain knowledge dataset includes:

[0066] S501: Obtain a domain knowledge dataset containing the capacity-voltage sequence and the corresponding incremental capacity peak feature parameters;

[0067] S502: Input the capacity-voltage sequence into the shared feature extractor to extract the high-level representation vector;

[0068] S503: Input the high-level representation vector into the incremental capacity peak feature regressor to obtain the predicted peak features;

[0069] S504: Calculate the predicted peak characteristics and corresponding incremental capacity peak characteristic parameters. Mean square error loss between Based on the mean square error loss, backpropagation is performed to jointly optimize all network parameters of the shared feature extractor and the incremental capacity peak feature regressor.

[0070] Mean square error loss The calculation formula is:

[0071] ;

[0072] S505: Repeat steps S501 to S504 for iterative training to obtain a pre-trained shared feature extractor with preliminary aging-related characterization capabilities.

[0073] Preferably, in step S6, the regressor fine-tuning process based on the frozen pre-trained shared feature extractor includes:

[0074] S601: Obtain an aging pattern dataset containing capacity-voltage sequences and corresponding aging pattern labels;

[0075] S602: Remove the incremental capacity peak feature regressor and connect the newly initialized aging mode regressor after the pre-trained shared feature extractor;

[0076] S603: Completely freeze all network parameters of the pre-trained shared feature extractor, only map the capacity-voltage sequence to the frozen pre-trained shared feature extractor as a high-level representation, and then input it into the aging mode regressor to obtain the predicted aging mode;

[0077] S604: Calculate and predict aging modes Corresponding aging mode label Mean square error loss difference between And based solely on this mean square error loss Optimize the network parameters of the aging mode regressor and establish a preliminary mapping relationship from a fixed high-level representation to the three aging modes;

[0078] The formula for calculating the mean squared error loss is:

[0079] ;

[0080] S605: Repeat steps S601 to S604 for iterative training to obtain a preliminary network model for the aging mode estimation task.

[0081] Preferably, in step S6, the low-learning-rate global fine-tuning process based on unfreezing the pre-trained shared feature extractor includes:

[0082] S606: Based on the network model that has been initially adapted to the aging mode estimation task, unfreeze all network parameters of the pre-trained shared feature extractor and set a learning rate lower than that used for fine-tuning the pre-trained and frozen regressors.

[0083] S607: The capacity-voltage sequence is mapped through a shared feature extractor to obtain a high-level characterization, which is then input into an aging mode regressor to obtain a predicted aging mode;

[0084] S608: Calculate and predict aging modes Corresponding aging mode label Mean square error loss difference between And based solely on this mean square error loss Simultaneously and globally fine-tuning all network parameters of the shared feature extractor and aging pattern regressor;

[0085] S609: Repeat steps S606 to S608 for iterative training to obtain the trained domain knowledge-guided aging mode estimation model.

[0086] Compared with existing technologies, the domain knowledge-guided lithium-ion battery aging mode estimation method in this invention has the following advantages:

[0087] This invention employs a network structure consisting of a shared feature extractor and a task-specific regressor. Regular charging data is integrated in ampere-hours and sampled using a unified voltage grid to construct a Q(V) input sequence. Based on a parameterized Q(V) model with multiple superimposed arctangent terms, the area, position, and width of the corresponding incremental capacity peaks are identified through model parameter identification as candidate incremental capacity peak feature parameters. The domain knowledge supervision objective for the first stage is determined based on the Pearson correlation coefficient between the candidate incremental capacity peak feature parameters and the aging mode label, as well as the parameter evolution stability. The shared feature extractor and the incremental capacity peak feature regressor are pre-trained using Q(V) samples without aging mode labels.

[0088] For a small number of tagged samples, capacity scaling and translation alignment are performed using low-rate pseudo-open-circuit voltage curves and positive and negative electrode potential curves to calculate LAM. PE LAM NE And LLI. In the second stage, the incremental capacity peak feature regressor is removed, the aging mode regressor is connected, and the shared feature extractor is frozen, so that only the mapping from the fixed representation to the three aging modes is learned; then all network parameters are unfrozen and global fine-tuned at a low learning rate, so as to avoid a small number of labels from destroying the aging-related representations obtained by pre-training.

[0089] Compared with the prior art, the present invention has the following advantages:

[0090] (1) The conventional charging curves that do not require aging mode labels are transformed into domain knowledge supervision signals, enabling a large amount of unlabeled operation data to participate in aging characterization learning, reducing the reliance on low-rate pseudo-open circuit voltage diagnosis and manual labeling.

[0091] (2) The incremental capacity peak area, peak position and peak width can be obtained directly by analyzing the parameterized Q(V) model without performing explicit numerical differentiation on discrete measurement data, thus reducing the impact of sampling noise, filtering settings and adjacent peak overlap on feature stability.

[0092] (3) By using the training mechanism of “domain knowledge pre-training, regressor fine-tuning of frozen shared feature extractor, and global fine-tuning with low learning rate”, the effective degrees of freedom of end-to-end learning under the condition of scarce labels are reduced, and representation drift and overfitting are suppressed.

[0093] (4) It can simultaneously estimate LAM on public datasets with different material systems, capacity scales and aging conditions. PE LAM NE And LLI; with a 10% label ratio in dataset #1, the root mean square errors of the three modes are 0.79%, 1.67% and 0.78%, respectively, and they maintain low estimation errors on datasets with different material systems and aging conditions.

[0094] In summary, this invention pre-trains a shared feature extractor by extracting the incremental capacity peak area, peak position, and peak width as domain knowledge supervision signals. It also incorporates a small number of labels generated by aligning low-rate pseudo-open-circuit voltage curves with positive and negative electrode potential curves to identify positive electrode active material loss, negative electrode active material loss, and cyclic lithium loss. This is achieved by performing regressor fine-tuning based on freezing the pre-trained shared feature extractor and low-learning-rate global fine-tuning based on thawing the shared feature extractor. This enables continuous estimation of three aging modes, thereby reducing reliance on low-rate diagnostic tests and high-quality aging mode labels, alleviating characterization drift and overfitting problems under label scarcity conditions, and improving the accuracy and generalization of electrode-level aging mode estimation. Attached Figure Description

[0095] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0096] Figure 1 This is a logic block diagram of a domain knowledge-guided method for estimating the aging mode of lithium-ion batteries.

[0097] Figure 2 A schematic diagram of the overall framework for estimating a two-stage aging pattern guided by domain knowledge.

[0098] Figure 3 An evolution graph of aging pattern labels generated in dataset #1: Figure 3 (a) is LAM PE Evolution curve as the number of cycles changes, Figure 3 (b) is LAM NE Evolution curve as the number of cycles changes, Figure 3 (c) shows the evolution curve of LLI as the number of cycles changes.

[0099] Figure 4 Evolution of incremental capacity peak feature parameters in dataset #1: Figure 4 (a) to (m) represent A1 to A respectively. 4、 V 0,1 To V 0,3、 Evolution diagram from ω1 to ω3.

[0100] Figure 5 The graph shows a comparison of the regression results of DK-CNN and the baseline CNN for the three aging modes with a 10% label ratio: Figure 5 (a) is LAM PE Comparison chart of estimation results Figure 5 (b) is LAM NE Comparison chart of estimation results Figure 5 (c) is a comparison chart of LLI estimation results.

[0101] Figure 6 Box plots for estimating the absolute error of three aging modes under different label ratios: Figure 6 (a) is LAM PE Estimated absolute error distribution plot, Figure 6 (b) is LAM NE Estimated absolute error distribution plot, Figure 6 (c) is a plot of the absolute error distribution of LLI estimation.

[0102] Figure 7 The following is a graph showing the principal component analysis results of high-level features of DK-CNN and the baseline CNN with a 10% label ratio: Figure 7 (a) to Figure 7 (c) High-level features of the DK-CNN model in LAM PE LAM NE Principal component analysis results in the LLI direction, Figure 7 (d) to Figure 7 (f) represents the high-level features of the baseline CNN model in LAM. PE LAM NE Principal component analysis results in the LLI direction.

[0103] Figure 8 A graph comparing the errors of different estimation methods for dataset #1 at a 10% label ratio: Figure 8 (a) is LAM PEEstimated error comparison results Figure 8 (b) is LAM NE Estimated error comparison results Figure 8 (c) shows the comparison results of LLI estimation errors. Figure 8 (d) to Figure 8 (f) shows the comparison results of other evaluation indicators corresponding to the three aging modes.

[0104] Figure 9 A graph comparing the errors of different estimation methods for dataset #2 at a 10% label ratio: Figure 9 (a) is LAM PE Estimated error comparison results Figure 9 (b) is LAM NE Estimated error comparison results Figure 9 (c) shows the comparison results of LLI estimation errors. Figure 9 (d) to Figure 9 (f) shows the comparison results of other evaluation indicators corresponding to the three aging modes. Detailed Implementation

[0105] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0106] The following detailed explanation illustrates the specific implementation methods:

[0107] Example:

[0108] This embodiment discloses a method for estimating the aging mode of lithium-ion batteries based on domain knowledge.

[0109] like Figure 1 and Figure 2 As shown, the domain knowledge-guided method for estimating the aging mode of lithium-ion batteries includes:

[0110] S1: Obtain charging cycle data and low-rate pseudo-open circuit voltage test data for each cycle of lithium-ion battery during conventional 1C constant current charging.

[0111] S2: Construct the capacity-voltage sequence corresponding to each cycle based on the charging cycle data of each cycle;

[0112] S3: Extract the incremental capacity peak feature parameters of the capacity-voltage sequence of each cycle, and construct a domain knowledge dataset based on the capacity-voltage sequence of each cycle and the corresponding incremental capacity peak feature parameters;

[0113] S4: Alignment and identification are performed based on the low-rate pseudo-open-circuit voltage test data of each cycle and the corresponding positive and negative electrode potential curves. The loss of positive active material, negative active material and cyclic lithium of each cycle are calculated to construct aging mode labels. An aging mode dataset is constructed based on the capacity-voltage sequence of each cycle and the corresponding aging mode labels.

[0114] S5: Pre-train the shared feature extractor using a domain knowledge dataset to obtain the pre-trained shared feature extractor;

[0115] S6: By sequentially performing regressor fine-tuning based on frozen pre-trained shared feature extractor and low-learning-rate global fine-tuning based on unfrozen pre-trained shared feature extractor on the aging mode dataset, a domain knowledge-guided aging mode estimation model (DK-CNN network) is obtained after training.

[0116] S7: Input the capacity-voltage sequence of the battery to be estimated into the trained domain knowledge to guide the aging mode estimation model. Through forward propagation calculation, output the aging mode estimation results including positive electrode active material loss, negative electrode active material loss and cyclic lithium loss.

[0117] To better illustrate the technical solution of the present invention, this embodiment will be described in more detail through the following parts.

[0118] I. Constructing Capacity-Voltage Sequences

[0119] This invention uses the i-th characterizing cycle of the conventional 1C constant current charging Q(V) sequence x ᵢ As input, the corresponding loss of LAM in the positive electrode active material PE Loss of negative electrode active material LAM NE The output vector y is composed of the recyclable lithium loss LLI. ᵢ The aging pattern estimation is expressed as a multi-output regression problem from the Q(V) sequence to three consecutive degradation values:

[0120] .

[0121] Specifically, the process of constructing the capacity-voltage sequence includes:

[0122] S201: Obtain the voltage at each time point in the corresponding loop. and current Charging cycle data;

[0123] S202: At the lower cutoff voltage Upper cutoff voltage Within the range, the capacity value corresponding to each voltage point is calculated using ampere-hour integration to construct an initial capacity-voltage curve;

[0124] Initial capacity-voltage curve The calculation is expressed as:

[0125] ;

[0126] S203: Discretize the voltage axis of the initial capacity-voltage curve with a preset voltage interval (e.g., 10 mV) to form a uniform voltage sampling grid. ,in = , = (From the lower cutoff voltage to the upper cutoff voltage); The capacity values ​​on the initial capacity-voltage curve are mapped to a unified voltage sampling grid point using an interpolation algorithm, generating a capacity-voltage sequence with a unified dimension and preserving the local charging pattern:

[0127] .

[0128] II. Extracting the Capacity-Voltage Sequence

[0129] In the specific implementation process, the processing steps for extracting the incremental capacity peak characteristic parameters of the capacity-voltage sequence include:

[0130] S301: To avoid numerical differential amplification noise in traditional incremental capacity analysis, this invention directly adopts parameterization. The model is fitted. Let m be the number of identifiable incremental capacity peaks. The structure is constructed by superimposing m arctangent terms with a constant term. A parameterized capacity-voltage model with superposition of arctangent terms;

[0131] The formula is expressed as:

[0132] ;

[0133] In the formula: , These are the capacity value and the voltage value, respectively. , and The first The peak area, peak center location, and peak width of each peak. For constant terms;

[0134] S302: Convert the corresponding cycle's capacity-voltage sequence Input a parameterized capacity-voltage model, and minimize the actual capacity-voltage sequence using a nonlinear least squares method. Fit the capacity-voltage curve to the model The error between them is used to identify the model parameters and obtain the peak area parameters corresponding to each arctangent term. Peak center position parameters Peak width Parameters and constants The peak area parameter, peak center position parameter, and peak width parameter are used as candidate incremental capacity peak characteristic parameters for the corresponding cycle. The change in peak area reflects the effective capacity contribution of the corresponding phase transition interval, the peak position shift is related to impedance growth or polarization enhancement, and the change in peak width can reflect reaction non-uniformity and kinetic dispersion.

[0135] The formula is expressed as:

[0136] ;

[0137] In the formula: , and They represent the first to the last. The peak area parameter, peak center position parameter, and peak width parameter corresponding to each arctangent term. The number of arctangent terms in the parameterized capacity-voltage model; This represents the constant term in the parameterized capacity-voltage model; This indicates the number of voltage sampling points in the capacity-voltage sequence; Indicates the first The actual capacity value corresponding to each voltage sampling point; The first one obtained based on the identification parameters The fitted capacity value corresponding to each voltage sampling point;

[0138] S303: Calculate the Pearson correlation coefficient between the characteristic parameters of each candidate incremental capacity peak and the corresponding losses of positive electrode active material, negative electrode active material, and cyclic lithium in the lithium-ion battery.

[0139] S304: Based on the Pearson correlation coefficient results, and combined with the stability of the candidate incremental capacity peak feature parameters during the cycling process and the complementarity of the representation information between different candidate incremental capacity peak feature parameters, the candidate incremental capacity peak feature parameters are screened to obtain incremental capacity peak feature parameters that can characterize the battery degradation state, and a domain knowledge dataset is constructed based on the incremental capacity peak feature parameters.

[0140] In this embodiment, stability is achieved simply by the parameter evolution curve exhibiting strong monotonicity; complementarity refers to selecting parameters with different peak types as much as possible (because peak position, peak width, and peak area have different meanings regarding battery aging modes), and both of these points can be determined manually.

[0141] Specifically, domain knowledge datasets Represented as:

[0142] ;

[0143] In the formula: , The first Capacity-voltage sequence and incremental capacity peak characteristic parameters for each sample (cycle); This represents the number of samples in the domain knowledge dataset.

[0144] III. Constructing Aging Pattern Labels

[0145] A small number of high-quality aging mode labels were identified by low-rate pseudo-open-circuit voltage curves and publicly available positive and negative electrode potential curves.

[0146] Specifically, the steps for constructing aging pattern labels include:

[0147] S401: Construct a low-rate pseudo-open-circuit voltage curve based on the low-rate pseudo-open-circuit voltage test data of the corresponding cycle;

[0148] S402: Introducing positive electrode capacity scaling parameters, negative electrode capacity scaling parameters, positive electrode capacity translation parameters, and negative electrode capacity translation parameters, the capacity coordinates of the standard positive and negative electrode potential curves are mapped to normalized coordinates of the two electrodes through scaling and translation. After subtraction, the open-circuit voltage curve of the full cell is reconstructed. ;

[0149] Full cell open circuit voltage curve The calculation formula is:

[0150] ;

[0151] ;

[0152] ;

[0153] In the formula: ; and These are the positive electrode potential curve and the negative electrode potential curve, respectively. , These are the positive electrode capacity scaling parameters and the negative electrode capacity scaling parameters, respectively. , These are the positive electrode capacity shift parameters and the negative electrode capacity shift parameters, respectively. Coordinates for total battery capacity; Capacity coordinates The normalized capacity coordinates of the positive and negative electrodes are obtained after scaling and translation mapping of the positive and negative electrode capacities;

[0154] S403: The low-rate pseudo-open-circuit voltage curve of the corresponding cycle. Discretized After sampling each capacity point, the open-circuit voltage curve of the reconstructed full cell is compared with that of the full cell. By performing point-by-point comparisons, the optimal alignment parameters for the corresponding loop are obtained by minimizing the mean square error between the measured and reconstructed values. ;

[0155] The formula is expressed as:

[0156] ;

[0157] In the formula: , The low-rate pseudo-open-circuit voltage curve and the full-cell open-circuit voltage curve are respectively shown in Capacity values ​​at each capacity sampling point;

[0158] S404: Obtain the positive electrode capacity scaling parameters obtained from the earliest characterization cycle identification of the corresponding lithium-ion battery. Negative electrode capacity scaling parameters Positive electrode capacity shift parameters and negative electrode capacity shift parameters This is the baseline state;

[0159] S405: Based on the baseline state, combined with the optimal alignment parameters Positive electrode capacity scaling parameters and negative electrode capacity scaling parameters Calculate the loss of positive electrode active material and loss of negative electrode active material ;

[0160] The calculation formula is:

[0161] ;

[0162] ;

[0163] S406: Based on the baseline state, combined with the optimal alignment parameters Positive electrode capacity scaling parameters Positive electrode capacity shift parameters and negative electrode capacity shift parameters Calculate the cyclic lithium capacity parameters of the baseline cycle And the current cycle's recyclable lithium quantity parameters And through the cyclic lithium quantity parameters of the benchmark cycle And the current cycle's recyclable lithium quantity parameters Calculate the loss of recyclable lithium ;

[0164] The calculation formula is:

[0165] ;

[0166] ;

[0167] ;

[0168] S407: Loss of positive electrode active material Loss of negative electrode active material and recyclable lithium loss This serves as the aging mode label for the corresponding cycle. This yields the LAM corresponding to a small number of characterization cycles. PE LAM NE And LLI supervised labels. The positive and negative potential curves are only used for label generation and are not used as input to DK-CNN. During the deployment stage, only the regular charging Q(V) curve is needed to output the aging mode estimate.

[0169] Specifically, aging pattern datasets labeled with aging patterns. Represented as:

[0170] ;

[0171] In the formula: , The first Capacity-voltage sequence and aging pattern label for each sample (cycle); , , The first Loss of positive electrode active material, loss of negative electrode active material, and loss of cyclic lithium in each sample (cycle); This represents the number of samples in the aging pattern dataset.

[0172] IV. DK-CNN Network Structure

[0173] The DK-CNN network adopts a structure of "shared feature extractor + task-specific regressor".

[0174] Shared feature extractor f E (·) Maps the Q(V) sequence to a higher-level representation, the aging pattern regressor r H(·) and incremental capacity peak feature regressor r C (·) Execute the final task and intermediate tasks separately:

[0175] ;

[0176] .

[0177] As shown in Table 1, the shared feature extractor consists of three cascaded one-dimensional convolutional blocks, a global average pooling layer, a 64-dimensional fully connected layer, and a Dropout layer with a dropout rate of 0.20.

[0178] The first one-dimensional convolutional block contains two cascaded one-dimensional convolutional layers with 16 filters and kernel sizes of 7 and 5, respectively, and a max pooling layer.

[0179] The second one-dimensional convolutional block contains two cascaded one-dimensional convolutional layers with 32 filters and kernel sizes of 5 and 3 respectively, and a max pooling layer;

[0180] The third one-dimensional convolutional block contains two one-dimensional convolutional layers with 64 filters and a kernel size of 3. Each one-dimensional convolutional layer is connected to a layer normalization function and a ReLU activation function in sequence.

[0181] Global average pooling aggregates convolutional features, and a 64-dimensional fully connected layer forms a high-level representation.

[0182] Dropout layers are used to suppress overfitting.

[0183] Table 1. Structural configuration of the DK-CNN shared feature extractor

[0184]

[0185] V. Model Training

[0186] This invention employs a two-stage training strategy to train the DK-CNN network. The first stage of training involves pre-training the shared feature extractor; the second stage includes fine-tuning the regressor based on freezing the pre-trained shared feature extractor and global fine-tuning with a low learning rate after unfreezing the pre-trained shared feature extractor.

[0187] The first stage of training, which involves pre-training the shared feature extractor and the incremental capacity peak feature regressor using a domain knowledge dataset, includes the following steps:

[0188] S501: Obtain a domain knowledge dataset containing the capacity-voltage sequence and the corresponding incremental capacity peak feature parameters;

[0189] S502: Input the capacity-voltage sequence into the shared feature extractor to extract the high-level representation vector;

[0190] S503: Input the high-level representation vector into the incremental capacity peak feature regressor (which can be a linear fully connected layer without activation function) to obtain the predicted peak features;

[0191] S504: Calculate the predicted peak characteristics and corresponding incremental capacity peak characteristic parameters. Mean square error loss between Based on the mean square error loss, backpropagation is performed to jointly optimize all network parameters of the shared feature extractor and the incremental capacity peak feature regressor.

[0192] Mean square error loss The calculation formula is:

[0193] ;

[0194] S505: Repeat steps S501 to S504 for iterative training to obtain a pre-trained shared feature extractor with preliminary aging-related characterization capabilities.

[0195] The steps for fine-tuning the regressor based on the frozen pre-trained shared feature extractor in the second stage of training include:

[0196] S601: Obtain an aging pattern dataset containing capacity-voltage sequences and corresponding aging pattern labels;

[0197] S602: Remove the incremental capacity peak feature regressor and connect a newly initialized aging mode regressor (which can be a linear fully connected layer without activation function) after the pre-trained shared feature extractor.

[0198] S603: Completely freeze all network parameters of the pre-trained shared feature extractor (keeping its weights unchanged during training), only map the frozen pre-trained shared feature extractor of the capacity-voltage sequence to a high-level representation, and then input it into the aging mode regressor to obtain the predicted aging mode.

[0199] S604: Calculate and predict aging modes Corresponding aging mode label Mean square error loss difference between And based solely on this mean square error loss Optimize the network parameters of the aging mode regressor and establish a preliminary mapping relationship from a fixed high-level representation to the three aging modes;

[0200] The formula for calculating the mean squared error loss is:

[0201] ;

[0202] S605: Repeat steps S601 to S604 for iterative training to obtain a preliminary network model for the aging mode estimation task.

[0203] In the second stage of training, global fine-tuning with a low learning rate is performed based on the thawed pre-trained shared feature extractor:

[0204] S606: Based on the network model that has been initially adapted to the aging pattern estimation task, unfreeze all network parameters of the pre-trained shared feature extractor and set a learning rate lower than that used for pre-training and fine-tuning of the frozen regressor (i.e., a low learning rate).

[0205] S607: The capacity-voltage sequence is mapped through a shared feature extractor to obtain a high-level characterization, which is then input into an aging mode regressor to obtain a predicted aging mode;

[0206] S608: Calculate and predict aging modes Corresponding aging mode label Mean square error loss difference between And based solely on this mean square error loss Simultaneously, all network parameters of the shared feature extractor and aging mode regressor are finely tuned globally, so that the entire network can adapt to the final aging mode estimation task while retaining the aging-related representations learned in the pre-training stage to the maximum extent and suppressing overfitting.

[0207] S609: Repeat steps S606 to S608 for iterative training to obtain the trained domain knowledge-guided aging mode estimation model.

[0208] In this invention, the first stage of training jointly optimizes f on DC. E and r C The goal is not to directly predict aging patterns, but to constrain the feature space with incremental capacity peak parameters that more closely resemble the input curve shape and are related to aging evolution. The second stage of training removes r. C And connect to the newly initialized r H ; Frozen regressor fine-tuning f E Only a fixed mapping from pre-trained representations to three aging modes is established; low-learning-rate global fine-tuning freezes all parameters, and global fine-tuning with an even lower learning rate allows the network to adapt to the final task and retain as many pre-trained representations as possible.

[0209] VI. Experimental Instructions

[0210] To better illustrate the advantages of the technical solution of the present invention, the following experiment is disclosed in this embodiment.

[0211] This experiment uses two publicly available lithium-ion battery aging datasets. Dataset #1 is the Oxford battery aging dataset, containing eight Kokam pouch cells with a rated capacity of approximately 0.74 Ah. The positive electrode materials are lithium cobalt oxide and nickel cobalt oxide, and the negative electrode is graphite. The batteries are aged at 40 °C using 2C constant current and constant voltage charging and dynamic operating condition discharging, and 1C charge-discharge tests and 1 / 20C pseudo open circuit voltage tests are performed periodically. Dataset #2 is a graphite / NMC pouch cell aging dataset, containing batteries with a rated capacity of 5.0 Ah from the same batch. The positive electrode is NMC111, and the negative electrode is graphite. It covers different charge-discharge rates, depths of discharge, ambient temperatures, and external pressures. Thirteen batteries from this dataset were selected for additional verification in the example.

[0212] All input Q(V) sequences and output targets are Z-score standardized before training and inversely transformed to their original physical dimensions before evaluation. To avoid information leakage, the training and test sets are divided in a 3:1 ratio, with each battery's samples belonging to only one subset. Incremental capacity feature pre-training, validation, fine-tuning, standardized parameter estimation, model selection, and early termination all use only training batteries; test batteries are used only for final evaluation. 15% of the training set is randomly selected as the validation set. The Adam learning rate for Phase 1 is 1×10⁻⁶. -3 The learning rate for Phase 2, Step 1 is 1×10⁻⁶. -3 The learning rate for step 2 is 1×10. -4 The maximum number of training rounds is 800, the batch size is 4, and the program terminates early if the validation loss does not improve after 100 consecutive rounds.

[0213] LAM of dataset #1 PE And LLI as a whole increases with increasing circulation, LAM NE The early and mid-stage evolution of LAMNE is more complex, with some batteries showing weak growth, local fluctuations, or slight declines. This does not indicate a recovery in the negative electrode active material, but rather that commercial batteries typically have negative electrode capacity redundancy, and some early negative electrode losses are masked by this redundancy; as the negative electrode redundancy is gradually depleted, the upward trend of LAMNE becomes more pronounced.

[0214] Table 2 presents the Pearson correlation coefficients between candidate incremental capacity peak features and the labels of the three aging modes. Significant differences in correlation exist for different parameters, indicating the need to combine correlation strength and evolutionary stability when selecting intermediate supervision targets. (A1, A4, and V from dataset #1 are also mentioned.) 0,1 The three modes exhibit strong and consistent correlations; A3, A4 and ω3 of dataset #2 can balance correlation strength, stability and parameter type complementarity.

[0215] Table 2. Pearson correlation coefficients between incremental capacity peak characteristics and aging mode labels.

[0216]

[0217] Table 3 presents the incremental capacity feature extraction and open-circuit voltage reconstruction errors. The Q(V) fitting error is used to evaluate the quality of incremental capacity peak parameter identification, and the pseudo-open-circuit voltage reconstruction error is used to evaluate the quality of aging mode label generation. The open-circuit voltage reconstruction errors of both datasets are in the millivolt range; dataset #2 has a higher absolute Q(V) fitting error than dataset #1 due to its larger capacity and more complex peak structure. The purpose of parameterized fitting is to obtain stable, low-dimensional, and aging-sensitive peak features, rather than reproducing all local peak shape details. Figure 3 and Figure 4 The images show the evolution of aging mode labels and incremental capacity peak feature parameters identified using dataset #1.

[0218] Table 3. Incremental Capacity Feature Extraction and Open-Circuit Voltage Reconstruction Errors

[0219]

[0220] This experiment uses root mean square error (RMSE) and mean absolute error (MAE) to evaluate the estimation results of each aging mode:

[0221] ;

[0222] ;

[0223] Where N t This represents the number of test samples.

[0224] To verify the performance improvement of the domain knowledge-guided strategy on aging mode estimation under low-label conditions, the proposed DK-CNN method is compared with a benchmark CNN without domain knowledge pre-training under a 10% label ratio. Figure 5 The graph shows a comparison of the regression results of DK-CNN and the benchmark CNN for three aging modes with a 10% label ratio. Under the same training label ratio, the prediction results of DK-CNN are closer to the actual aging mode labels, indicating that using incremental capacity peak features as domain knowledge supervision signals can improve the learning ability of shared feature extractors to degradation-related features.

[0225] Table 4 shows the comparison results of dataset #1 under different label ratios. When the label ratio decreases, the overall error of both models increases, but the performance of DK-CNN decreases less. At a 10% label ratio, DK-CNN outperforms LAM... PE LAM NEThe RMSE of DK-CNN and LLI were 0.79%, 1.67%, and 0.78%, respectively, while those of directly supervised CNN were 2.11%, 4.92%, and 3.55%. Even with a 1% label ratio, DK-CNN still achieved RMSEs of 0.96%, 1.79%, and 1.39%, respectively.

[0226] Table 4. Comparison of RMSE (%) between DK-CNN and benchmark CNN under different label ratios

[0227]

[0228] Figure 6 The distribution of the absolute error estimation for the three aging modes under different label ratios is further illustrated. Figure 6 It can be seen that as the proportion of training labels decreases, the error distribution range of directly supervised CNNs expands significantly, while DK-CNN maintains a small error fluctuation range under different label proportions. This indicates that the domain knowledge supervision signal constructed based on incremental capacity peak features can help the network learn more stable degradation representations, reduce the model's dependence on the number of high-quality aging mode labels, and improve the estimation stability under conditions with few labels.

[0229] LAM NE The improvement under some higher label proportions is not as good as LAM. PE The stability of DK-CNN compared to LLI is due to factors such as negative electrode capacity redundancy, electrode balance changes, and non-monotonic evolution caused by differences between cells; however, at low label ratios of 10% and 1%, DK-CNN still exhibits less performance degradation than directly supervised CNN.

[0230] With a 10% label ratio, 3D principal component analysis was performed on the output of the 64-dimensional fully connected layer. The DK-CNN features formed relatively continuous degradation trajectories, with the first principal component and LAM... PE LAM NE The absolute values ​​of the Pearson correlation coefficients for LLI were 0.970, 0.946, and 0.981, respectively, and the absolute values ​​of the Spearman correlation coefficients were 0.991, 0.933, and 0.996, respectively. All of these values ​​were higher than the baseline CNN, indicating that incremental capacity peak pre-training makes the high-level representation more consistent with aging evolution. The principal component analysis results are shown in Table 5.

[0231] Figure 7 This further demonstrates the distribution of high-level features in the three-dimensional principal component space of DK-CNN and benchmark CNNs. Figure 7It can be seen that the feature points extracted by DK-CNN can form a more continuous and smooth evolution trajectory along the battery degradation process, while the feature distribution of the benchmark CNN is relatively discrete and it is difficult to reflect obvious degradation evolution laws. The above results show that the domain knowledge supervision signal constructed based on incremental capacity peak features can guide the shared feature extractor to learn high-level feature representations related to battery aging mechanisms, thereby improving the model's ability to represent changes in electrode-level aging modes.

[0232] Table 5. Correlation between the first principal component and the aging mode label at a 10% labeling percentage.

[0233]

[0234] With 10% label ratio in dataset #1, DK-CNN is compared with DK-LSTM, DK-GRU, CNN, LSTM, GRU, Support Vector Regression (SVR), and Random Forest (RF), and the results and errors are shown in Table 6. DK-CNN shows the most balanced performance across the three modes, with the lowest errors in LAMPE and LLI; DK-GRU has the lowest error in LAMNE. Convolutional layers can extract local morphologies such as peak attenuation, peak shift, and capacity redistribution in specific voltage ranges along the voltage axis, thus making them suitable for Q(V) input and incremental capacity peak pre-training tasks. Figure 8 This further demonstrates the error comparison results of different methods on dataset #1. Figure 8 As can be seen, the overall error level of DK-CNN is lower than that of other comparative methods, which verifies the effectiveness of combining the domain knowledge-guided strategy with the convolutional feature extraction structure.

[0235] Table 6. Error comparison of different methods for dataset #1 at a 10% label ratio (%)

[0236]

[0237] On dataset #2, which has different material systems, rated capacities, and aging conditions, the LAM of DK-CNN is measured with a 10% label ratio. PE LAM NE The RMSEs of LLI were 0.46%, 1.13%, and 1.05%, respectively, and the MAEs were 0.37%, 0.92%, and 0.89%, respectively, all of which were superior to the directly supervised model and the traditional machine learning model. The results and errors are shown in Table 7. The results indicate that the scheme of extracting domain knowledge from the real conventional Q(V) curve and performing two-stage training can improve the applicability of few-label aging mode estimation under different data distributions. Figure 9 This further demonstrates the error comparison results of different methods on dataset #2. Figure 9As can be seen, DK-CNN maintains a low estimation error under different battery systems and aging conditions, verifying the generalization ability of the proposed method under cross-dataset conditions.

[0238] Table 7. Error comparison of different methods for dataset #1 at a 10% label ratio (%)

[0239]

[0240] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for estimating the aging mode of lithium-ion batteries based on domain knowledge, characterized in that, include: S1: Obtain charging cycle data and low-rate pseudo-open circuit voltage test data for each cycle of lithium-ion battery during conventional 1C constant current charging. S2: Construct the capacity-voltage sequence corresponding to each cycle based on the charging cycle data of each cycle; S3: Extract the incremental capacity peak feature parameters of the capacity-voltage sequence of each cycle, and construct a domain knowledge dataset based on the capacity-voltage sequence of each cycle and the corresponding incremental capacity peak feature parameters; S4: Alignment and identification are performed based on the low-rate pseudo-open-circuit voltage test data of each cycle and the corresponding positive and negative electrode potential curves. The loss of positive active material, negative active material and cyclic lithium of each cycle are calculated to construct aging mode labels. An aging mode dataset is constructed based on the capacity-voltage sequence of each cycle and the corresponding aging mode labels. S5: Pre-train the shared feature extractor using a domain knowledge dataset to obtain the pre-trained shared feature extractor; S6: By sequentially performing regressor fine-tuning based on frozen pre-trained shared feature extractor and low-learning-rate global fine-tuning based on unfrozen pre-trained shared feature extractor on the aging pattern dataset, a domain knowledge-guided aging pattern estimation model is obtained after training. S7: Input the capacity-voltage sequence of the battery to be estimated into the trained domain knowledge to guide the aging mode estimation model. Through forward propagation calculation, output the aging mode estimation results including positive electrode active material loss, negative electrode active material loss and cyclic lithium loss.

2. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 1, characterized in that: In step S2, the process of constructing the capacity-voltage sequence includes: S201: Obtain the voltage at each time point in the corresponding loop. and current Charging cycle data; S202: At the lower cutoff voltage Upper cutoff voltage Within the range, the capacity value corresponding to each voltage point is calculated using ampere-hour integration to construct an initial capacity-voltage curve; Initial capacity-voltage curve The calculation is expressed as: ； S203: Discretize the voltage axis of the initial capacity-voltage curve using a preset voltage interval to form a uniform voltage sampling grid. ,in = , = The capacity values ​​on the initial capacity-voltage curve are mapped to a uniform voltage sampling grid using an interpolation algorithm, generating a capacity-voltage sequence with uniform dimensions that preserves local charging patterns. .

3. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 2, characterized in that: Step S3, the processing steps for extracting the incremental capacity peak characteristic parameters of the capacity-voltage sequence include: S301: Building based on A parameterized capacity-voltage model with superposition of arctangent terms; The formula is expressed as: ； In the formula: , These are the capacity value and the voltage value, respectively. , and The first The peak area, peak center location, and peak width of each peak. For constant terms; S302: Convert the corresponding cycle's capacity-voltage sequence Input a parameterized capacity-voltage model, and minimize the actual capacity-voltage sequence using a nonlinear least squares method. Fit the capacity-voltage curve to the model The error between them is used to identify the model parameters and obtain the peak area parameters corresponding to each arctangent term. Peak center position parameters Peak width Parameters and constants The peak area parameter, peak center position parameter, and peak width parameter are used as candidate incremental capacity peak feature parameters for the corresponding cycle. The formula is expressed as: ； In the formula: , and They represent the first to the last. The peak area parameter, peak center position parameter, and peak width parameter corresponding to each arctangent term. The number of arctangent terms in the parameterized capacity-voltage model; This represents the constant term in the parameterized capacity-voltage model; This indicates the number of voltage sampling points in the capacity-voltage sequence; Indicates the first The actual capacity value corresponding to each voltage sampling point; The first one obtained based on the identification parameters The fitted capacity value corresponding to each voltage sampling point; S303: Calculate the Pearson correlation coefficient between the characteristic parameters of each candidate incremental capacity peak and the corresponding losses of positive electrode active material, negative electrode active material, and cyclic lithium in the lithium-ion battery. S304: Based on the Pearson correlation coefficient, candidate incremental capacity peak feature parameters are screened to obtain incremental capacity peak feature parameters that can characterize the battery degradation state, and a domain knowledge dataset is constructed based on the incremental capacity peak feature parameters.

4. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 3, characterized in that: In step S304, the domain knowledge dataset Represented as: ； In the formula: , The first Capacity-voltage sequence and incremental capacity peak characteristic parameters of each sample; This represents the number of samples in the domain knowledge dataset.

5. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 1, characterized in that: In step S4, the processing steps for constructing aging mode labels include: S401: Construct a low-rate pseudo-open-circuit voltage curve based on the low-rate pseudo-open-circuit voltage test data of the corresponding cycle; S402: Introducing positive electrode capacity scaling parameters, negative electrode capacity scaling parameters, positive electrode capacity translation parameters, and negative electrode capacity translation parameters, the capacity coordinates of the standard positive and negative electrode potential curves are mapped to normalized coordinates of the two electrodes through scaling and translation. After subtraction, the open-circuit voltage curve of the full cell is reconstructed. ; Full cell open circuit voltage curve The calculation formula is: ； ； ； In the formula: ; and These are the positive electrode potential curve and the negative electrode potential curve, respectively. , These are the positive electrode capacity scaling parameters and the negative electrode capacity scaling parameters, respectively. , These are the positive electrode capacity shift parameters and the negative electrode capacity shift parameters, respectively. Coordinates for total battery capacity; Capacity coordinates The normalized capacity coordinates of the positive and negative electrodes are obtained after scaling and translation mapping of the positive and negative electrode capacities; S403: The low-rate pseudo-open-circuit voltage curve of the corresponding cycle. Discretized After sampling each capacity point, the open-circuit voltage curve of the reconstructed full cell is compared with... By performing point-by-point comparisons, the optimal alignment parameters for the corresponding loop are obtained by minimizing the mean square error between the measured and reconstructed values. ; The formula is expressed as: ； In the formula: , The low-rate pseudo-open-circuit voltage curve and the full-cell open-circuit voltage curve are respectively shown in Capacity values ​​at each capacity sampling point; S404: Obtain the positive electrode capacity scaling parameters obtained from the earliest characterization cycle identification of the corresponding lithium-ion battery. Negative electrode capacity scaling parameters Positive electrode capacity shift parameters and negative electrode capacity shift parameters This is the baseline state; S405: Based on the baseline state, combined with the optimal alignment parameters Positive electrode capacity scaling parameters and negative electrode capacity scaling parameters Calculate the loss of positive electrode active material and loss of negative electrode active material ; The calculation formula is: ； ； S406: Based on the baseline state, combined with the optimal alignment parameters Positive electrode capacity scaling parameters Positive electrode capacity shift parameters and negative electrode capacity shift parameters Calculate the cyclic lithium capacity parameters of the baseline cycle And the current cycle's recyclable lithium quantity parameters And through the cyclic lithium quantity parameters of the benchmark cycle And the current cycle's recyclable lithium quantity parameters Calculate the loss of recyclable lithium ; The calculation formula is: ； ； ； S407: Loss of positive electrode active material Loss of negative electrode active material and recyclable lithium loss As the aging mode label for the corresponding cycle.

6. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 5, characterized in that: In step S4, the aging pattern dataset with aging pattern labels Represented as: ； In the formula: , The first Capacity-voltage sequence and aging pattern labeling for each sample; , , The first The loss of positive electrode active material, the loss of negative electrode active material, and the loss of cyclic lithium for each sample; This represents the number of samples in the aging pattern dataset.

7. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 1, characterized in that: In step S5, the shared feature extractor consists of three cascaded one-dimensional convolutional blocks, a global average pooling layer, a fully connected layer, and a Dropout layer. The first one-dimensional convolutional block contains two cascaded one-dimensional convolutional layers and a max pooling layer; The second one-dimensional convolutional block contains two cascaded one-dimensional convolutional layers and a max pooling layer; The third one-dimensional convolutional block contains two one-dimensional convolutional layers, each of which is connected sequentially to a layer normalization function and a ReLU activation function.

8. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 1, characterized in that: Step S5, the pre-training process of the shared feature extractor and the incremental capacity peak feature regressor using the domain knowledge dataset, includes: S501: Obtain a domain knowledge dataset containing the capacity-voltage sequence and the corresponding incremental capacity peak feature parameters; S502: Input the capacity-voltage sequence into the shared feature extractor to extract the high-level representation vector; S503: Input the high-level representation vector into the incremental capacity peak feature regressor to obtain the predicted peak features; S504: Calculate the mean squared error loss between the predicted peak feature and the corresponding incremental capacity peak feature parameter, and backpropagate based on the mean squared error loss to jointly optimize all network parameters of the shared feature extractor and the incremental capacity peak feature regressor. S505: Repeat steps S501 to S504 for iterative training to obtain a pre-trained shared feature extractor with preliminary aging-related characterization capabilities.

9. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 8, characterized in that: In step S6, the regressor fine-tuning process based on the frozen pre-trained shared feature extractor includes: S601: Obtain an aging pattern dataset containing capacity-voltage sequences and corresponding aging pattern labels; S602: Remove the incremental capacity peak feature regressor and connect the newly initialized aging mode regressor after the pre-trained shared feature extractor; S603: Completely freeze all network parameters of the pre-trained shared feature extractor, only map the capacity-voltage sequence to the frozen pre-trained shared feature extractor as a high-level representation, and then input it into the aging mode regressor to obtain the predicted aging mode; S604: Calculate the mean squared error loss difference between the predicted aging mode and the corresponding aging mode label, and optimize the network parameters of the aging mode regressor based solely on this mean squared error loss to establish a preliminary mapping relationship from a fixed high-level representation to the three aging modes. S605: Repeat steps S601 to S604 for iterative training to obtain a preliminary network model for the aging mode estimation task.

10. The lithium-ion battery aging mode estimation method based on domain knowledge as described in claim 9, characterized in that: In step S6, the processing steps for low-learning-rate global fine-tuning based on unfreezing the pre-trained shared feature extractor include: S606: Based on the network model that has been initially adapted to the aging mode estimation task, unfreeze all network parameters of the pre-trained shared feature extractor and set a learning rate lower than that used for fine-tuning the pre-trained and frozen regressors. S607: The capacity-voltage sequence is mapped through a shared feature extractor to obtain a high-level characterization, which is then input into an aging mode regressor to obtain a predicted aging mode; S608: Calculate the mean squared error loss difference between the predicted aging mode and the corresponding aging mode label, and fine-tune all network parameters of the shared feature extractor and the aging mode regressor simultaneously and globally based on this mean squared error loss. S609: Repeat steps S606 to S608 for iterative training to obtain the trained domain knowledge-guided aging mode estimation model.

Images