Lithium ion battery health state estimation method based on fragmented charging data

Abstract

The invention relates to a fragmented charging data-based lithium ion battery health state estimation method, which comprises the following steps of: acquiring voltage, current and timestamp original data of a constant current charging stage of a lithium ion battery, and intercepting a 100-600-second standardized fragmented charging fragment based on a window parameter consisting of an initial voltage threshold Vstart, a sampling interval delta t and a sampling point number n; the method comprises the following steps: calculating an accumulated charging capacity through a coulomb counting method, generating a voltage-capacity curve, deducing a capacity increment-voltage curve and a voltage increment-capacity curve, and carrying out smooth filtering, multi-modal data splicing and normalization processing to obtain a 3 * n multi-channel characteristic tensor; according to the method, complete charging and discharging data are not needed, different battery types and complex working conditions are adapted, the lowest SOH estimation root-mean-square error can reach 0.00355, the highest determination coefficient (R) can reach 0.99485, and the practicability and estimation efficiency in practical application are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a lithium ion battery state of health estimation method based on fragmented charging data, which is suitable for accurate SOH evaluation of lithium ion batteries in electric vehicles (EV), large-scale energy storage systems (ESS), smart grids and portable electronic devices, especially can solve the SOH estimation problem caused by fragmented charging data in practical applications, and provides core technical support for battery management system (BMS), belonging to the field of lithium ion battery state monitoring. BACKGROUND

[0002] Lithium ion batteries have become the core energy storage components in the fields of electric vehicles, energy storage systems, etc. due to their high energy density, long cycle life and environmental friendliness. The state of health (SOH) as a key indicator to measure the remaining capacity and performance degradation of the battery directly determines the safety, reliability and remaining useful life (RUL) of the battery, and is a key functional module of the battery management system (BMS), and its estimation accuracy directly affects the operating efficiency and safety performance of the equipment.

[0003] The existing SOH estimation methods for lithium ion batteries have many technical defects and cannot meet the actual application requirements:

[0004] 1. Data dependency limits practicality: In traditional direct measurement methods, full charge and discharge coulomb counting method requires complete charge and discharge cycles to realize SOH calibration, which is only suitable for laboratory scenarios. However, the charging behavior of electric vehicles and other equipment in actual applications is random, and the proportion of complete charge and discharge cycles is less than 5%. Fragmented charging fragments become the mainstream data form, which makes this method unable to be implemented. Electrochemical impedance spectroscopy (EIS) method relies on high-precision impedance test equipment, which not only has high cost, but also is difficult to integrate into embedded systems such as vehicle-mounted BMS, and has poor engineering applicability.

[0005] 2. Weak model generalization ability: The estimation methods based on equivalent circuit model or electrochemical model have high sensitivity to model parameters, are easily disturbed by factors such as temperature, charge and discharge rate, and cell consistency in actual working conditions, have poor adaptability in cross-battery type and cross-working condition scenarios, have large estimation error fluctuations, and cannot meet the diversified application requirements.

[0006] 3. Insufficient feature extraction and time series modeling: Traditional deep learning models (such as CNN, LSTM) have inherent limitations: the fixed receptive field of CNN cannot capture long-term temporal dependence, and cannot extract long-term aging trends hidden in fragmented data; LSTM has insufficient ability to extract multi-scale local attenuation features, and cannot accurately identify the detailed changes of voltage-capacity curve; and existing deep learning methods mostly rely on constant voltage (CV) stage or complete charging cycle data, and do not design special feature extraction and modeling mechanisms for fragmented charging data, which cannot fully extract aging information from short-term charging data.

[0007] 4. Fragmented data standardization missing: existing methods lack standardized processing mechanisms for fragmented charging data, and different cycles and different batteries have significant differences in starting voltage, sampling interval, etc., resulting in poor data consistency, affecting model training convergence efficiency, and making it difficult to ensure cross-scene estimation accuracy, limiting the practical deployment value of the method.

[0008] Therefore, developing a SOH estimation method based on fragmented charging data to achieve high precision and strong generalization, solving the problems of strong data dependence, poor adaptability, insufficient precision and other problems of existing technology, has become a technical problem urgently to be solved by the industry. SUMMARY

[0009] In order to overcome the defects of the existing lithium ion battery SOH estimation method, such as dependence on complete charging data, poor adaptability to fragmented data, insufficient model generalization ability and limited estimation accuracy, the present application provides a lithium ion battery health state estimation method based on fragmented charging data, which realizes high-precision SOH estimation using only 100-600 seconds of fragmented charging fragments through standardized data preprocessing and mixed deep learning model design, and meets the real-time monitoring needs of BMS in actual scenarios.

[0010] A lithium ion battery health state estimation method based on fragmented charging data, comprising the following steps:

[0011] Step 1: Obtain the original data of the lithium ion battery constant current charging (CC) stage, the original data including charging voltage, charging current and corresponding timestamp, and the charging protocol adopting the industry common constant current-constant voltage (CC-CV) mode; based on the preset window parameter, the charging fragments are extracted from the original CC charging data and standardized to form fragmented charging fragments, the extraction of the charging fragments based on the preset window parameter extracts short-period charging fragments from the original CC charging data, specifically including:

[0012] S1.1, the window parameters include initial voltage threshold Vstart, sampling interval Δt and sampling point number n;

[0013] S1.2, Vstart takes 3.6V, 3.7V, 3.8V and 3.9V respectively, the electrochemical characteristics of the battery in this voltage range change most significantly with aging, and the information density is the highest, Δt takes 5s and 10s, and n takes 16 and 32;

[0014] S1.3, when the voltage V(t) of the lithium ion battery constant current charging CC stage reaches or exceeds Vstart for the first time, sampling is started, n data points are collected at Δt intervals to form multiple fragmented charging fragments.

[0015] The window parameters include an initial voltage threshold Vstart, a sampling interval Δt, and a sampling point number n.

[0016] Step two: calculate the cumulative charge capacity of the fragmented charging segment obtained in step one by Coulomb counting method, generate standardized multi-dimensional feature information, and perform smoothing filtering, normalization preprocessing, and then convert into a multi-channel feature tensor, which specifically includes:

[0017] S2.1, calculate the cumulative charge capacity according to Coulomb's law by current integration, and generate a voltage-capacity V-Q curve;

[0018] S2.2, derive the V-Q curve to obtain the ICA curve, and inversely derive the V-Q curve to obtain the DVA curve;

[0019] S2.3, use a moving average filter to smooth the V-Q curve, ICA curve, and DVA curve to suppress current noise, voltage fluctuation high-frequency interference;

[0020] S2.4, splice the three modal data voltages V, capacity increments IC, and voltage increments DV corresponding to the smoothed V-Q curve, ICA curve, and DVA curve into a 3×n multi-channel feature tensor X, where n is the sequence length;

[0021] S2.5, independently perform minimum-maximum normalization on each channel of the multi-channel feature tensor X to eliminate dimensional differences;

[0022] Step three: construct an end-to-end hybrid model PTCA-Net composed of a parallel time convolution network (PTCN) and a Transformer attention encoder, and optimize the hyperparameters, the PTCA-Net includes a one-dimensional convolution layer, a PTCN module, a Transformer encoder module, a residual connection, and a fully connected prediction layer from input to output; use a tree structure Parzen estimator in the Bayesian optimization algorithm to optimize the network hyperparameters; the hyperparameters to be optimized in step three include the number of PTCN layers, the number of Transformer encoder layers, the initial learning rate, the dropout rate, the hidden size, the batch size, the number of MHA heads, and the compression scale of the final convolution layer;

[0023] Step four: input the normalized feature values of step S2.5 into the PTCA-Net model for training, extract multi-scale local attenuation features in the fragmented segment through the PTCN module, combine the global time sequence dependence captured by the Transformer encoder module, and output the SOH estimation value; determine the optimal window parameters according to the training results of different fragmented charging segments.

[0024] The step two specifically includes:

[0025] S2.1, calculate the cumulative charging capacity of the lithium-ion battery charging stage according to Coulomb's law:

[0026]

[0027] wherein, is the cumulative charging capacity at the sampling point i, is the sampling starting point, and I(t) is the charging current;

[0028] S2.2, derive the ICA curve according to the formula

[0029]

[0030] wherein, represents the battery voltage;

[0031] S2.3, derive the DVA curve according to the formula

[0032]

[0033] S2.4, smooth the V-Q curve, the ICA curve and the DVA curve using a sliding average filter to suppress high-frequency noise;

[0034] S2.5, the smoothed curves are connected into a 3x n multi-channel feature tensor X through an input matrix, with three modalities of V, IC and DV:

[0035]

[0036] wherein n represents the final sequence length;

[0037] S2.6, eliminate the dimensional difference by minimum-maximum normalization:

[0038]

[0039] wherein, , are the minimum and maximum values of the corresponding channel respectively, is the normalized feature value.

[0040] The PTCA-Net in step three specifically comprises:

[0041] S3.1, sequentially comprises a one-dimensional convolution layer, a PTCA module, a Transformer encoder module, a residual connection and a fully connected prediction layer from input to output;

[0042] ​​S3.2, the one-dimensional convolution layer is used for linear transformation of the normalized feature value, and is mapped to a model adaptation dimension;

[0043] S3.3, the PTCT module comprises three parallel time convolution network (TCN) branches, each of which adopts a convolution kernel with a size of 3, 5 or 7, and extracts multi-scale local features through dilated causal convolution (DCC), each TCN branch comprises two time blocks, each of which integrates a DCC, and a GeLU activation function and group normalization are connected in sequence;

[0044] S3.4, the Transformer encoder module comprises a multi-head attention (MHA) layer and a feed-forward network (FFN), the MHA captures complex patterns through a plurality of groups of “query-key-value (Q-K-V)” calculations, and the FFN introduces a nonlinear transformation through two perceptrons and a GeLU activation function;

[0045] S3.5, the residual connection is applied to each time block inside the PTCT, around the MHA and FFN layers of the Transformer encoder, and the expression is:

[0046]

[0047] wherein x is an input, F(x) is an intra-block transformation, and y is an output;

[0048] S3.6, the fully connected prediction layer is used for mapping global features to one-dimensional SOH estimation values.

[0049] The mathematical expression of the dilated causal convolution (DCC) is:

[0050]

[0051] wherein K is a convolution kernel size, d is a dilation rate, is a convolution kernel weight, j is a current time step, x is an input feature, is an output feature.

[0052] The step three further comprises training the network, using an Adam optimizer and taking a mean square error (MSE) as a loss function. The expression of the loss function used in the model training in the step three is:

[0053]

[0054] wherein, is an actual SOH, is a predicted SOH; to avoid overfitting, if the loss of a verification set does not decrease for 20 consecutive epochs, the training is terminated.

[0055] The step four specifically comprises: taking the normalized feature value obtained in the step two as a model input; adopting a batch processing mode to train the model constructed in the step three, verifying the model performance through test set data, and calculating the mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), determination coefficient (R2) and other error indicators of the prediction result and the true SOH .

[0056] Compared with the prior art, the present application has the beneficial effects that:

[0057] The present application does not require complete charge-discharge cycle or CV stage data, but only needs 100-600 seconds of CC stage fragmented segments to realize SOH estimation, which is consistent with the charging behavior of "short-time energy supplement" in the actual scene, and solves the problem of "strong data dependence and poor practicability" of traditional methods. Through multi-channel feature fusion and hybrid model, multi-scale local features and global time sequence dependence are captured, and the accuracy is better than that of traditional CNN, LSTM and other models under different positive electrode material batteries and different charge-discharge rates. Through Bayesian optimization of hyperparameters and standardized data preprocessing, the model can adapt to different temperatures and different charge-discharge protocols, and the cross-scene estimation error fluctuation is very small. The method has excellent cross-protocol generalization ability, and even after training and testing on different charging protocols, it still maintains high accuracy, which proves its robustness in practical applications and is suitable for electric vehicles, battery energy storage and other fields.

[0058] 1. Adapt to fragmented charging scenes, and the practicability is greatly improved: The present application does not require complete charge-discharge cycle data or CV stage data, but only needs to collect 100-600 seconds of CC stage fragmented charging segments to realize accurate SOH estimation, which is completely consistent with the charging behavior of "short-time energy supplement" in the actual scene, breaks the dependence of traditional methods on complete charging data, enables SOH estimation to be integrated into practical application systems such as vehicle-mounted BMS, and solves the core pain points of traditional methods such as "strong data dependence and poor practicability".

[0059] 2. High estimation accuracy and excellent error indicators: Through multi-channel feature fusion (V-Q, ICA, DVA three modal data) and PTCA-Net hybrid model design, the PTCA module accurately extracts multi-scale local attenuation features (such as local trend of V-Q curve, peak value of ICA curve, fluctuation details of DVA curve) through parallel TCN branches with 3, 5 and 7 size convolution kernels, and the Transformer encoder module captures global time sequence dependence through multi-head attention mechanism, which realizes comprehensive mining of battery aging information. According to the test results in Table 1, the root mean square error (RMSE) of SOH estimation of the present application is as low as 0.00355, the mean absolute error (MAE) is as low as 0.00279, and the determination coefficient (R2) is as high as 0.99485, which is significantly better than traditional CNN, LSTM and other models.

[0060] 3. Strong generalization ability and good cross-scene adaptability: The dimension difference and noise interference of fragmented data from different sources are eliminated through standardized data preprocessing (window parameter standardization, smoothing filter, minimum-maximum normalization), and the precise optimization of model hyperparameters is realized by combining the Bayesian optimization algorithm (TPE), so that the model can adapt to different positive electrode material batteries, different charge-discharge rates, different temperatures, and different charge-discharge protocols, and the cross-scene estimation error fluctuation is small; even if training and testing are performed under different charge-discharge protocols, high accuracy can still be maintained, and the robustness is outstanding, which is suitable for various application scenarios such as electric vehicles, energy storage systems, smart grids, and portable electronic devices.

[0061] 4. Flexible deployment and high engineering realizability: The present application designs a window parameter optimization mechanism, which can determine the voltage range (optimal Vstart) with the most significant change in electrochemical characteristics with aging and the matching sampling interval Δt and sampling point number n in the target battery system through a one-time characterization stage, without the need for complex equipment modification or additional data acquisition costs; the residual connection is introduced in the model training process to solve the gradient vanishing problem of the deep network, and the Adam optimizer and early stopping mechanism are used to avoid overfitting, which has high training efficiency and is easy to deploy and implement. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0063] Figure 1 FIG. 1 is a schematic diagram of the architecture of the PTCA-Net model in the lithium-ion battery state of health estimation method based on fragmented charging data according to the present application. The figure shows the functions of each module from input to output of the model, including one-dimensional convolution layer, PTCA module (including 3 parallel TCN branches), Transformer encoder module (including MHA layer and FFN layer), residual connection and full connection prediction layer, clearly presenting the processing flow and feature extraction of multi-channel feature tensor, and the core mechanism of time series modeling. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] As Figure 1 shown, a lithium-ion battery state of health estimation method based on fragmented charging data includes the following steps:

[0066] Step one: Obtain the original data of the lithium-ion battery constant current (CC) stage, including the charging voltage, charging current and corresponding time stamp, and the charging protocol adopts the industry common constant current-constant voltage (CC-CV) mode; Based on the preset window parameters, the charging fragments are intercepted from the original CC charging data and standardized to form fragmented charging fragments, and the window parameters include the initial voltage threshold Vstart, the sampling interval Δt, and the sampling point number n;

[0067] S1.1, the window parameters include the initial voltage threshold Vstart, the sampling interval Δt, and the sampling point number n;

[0068] S1.2, Vstart takes 3.6V, 3.7V, 3.8V, 3.9V respectively, the electrochemical characteristics of the battery in this voltage interval change most significantly with aging, and the information density is the highest, Δt takes 5s, 10s, and n takes 16, 32;

[0069] S1.3, when the CC stage voltage V(t) first reaches or exceeds Vstart, sampling is started, n data points are collected at Δt intervals to form multiple fragmented charging fragments, and the fragment data of different cycles and different batteries are ensured to have comparability.

[0070] Step two: Calculate the cumulative charging capacity of the fragmented charging fragments obtained in step one by coulomb counting method, generate standardized multi-dimensional feature information, and after pretreatment such as smoothing filtering and normalization, convert it into a multi-channel feature tensor to ensure the consistency and availability of the data, which includes:

[0071] S2.1, according to the Coulomb law, the cumulative charging capacity in the sampling period is calculated by current integration, and the formula is:

[0072]

[0073] Where, is the sampling start time stamp, is the time stamp of the i-th sampling point (i=1, 2,..., n), is the sampling start point, I(t) is the charging current at time t, is the cumulative charging capacity at time t; Through this calculation, the voltage-capacity (V-Q) curve can be generated;

[0074] S2.2, the ICA curve is obtained by deriving the V-Q curve, which reflects the relationship between capacity increment and voltage,

[0075]

[0076] wherein, represents the battery voltage at sampling point i;

[0077] S2.3, the DVA curve reflecting the relationship between voltage increment and capacity is obtained by reverse derivation of the V-Q curve:

[0078]

[0079] S2.4, a moving average filter is used to smooth the V-Q, ICA, and DVA curves to suppress current noise, voltage fluctuations, and other high-frequency interference, forming the smoothed curves;

[0080] S2.5, the smoothed curves of S2.4 are connected into a 3*n multi-channel feature tensor X through an input matrix, connecting the three modal data (V, IC, DV) into a 3*n multi-channel feature tensor X:

[0081]

[0082] wherein n represents the final sequence length;

[0083] S2.6, the minimum-maximum normalization is independently performed on each channel of the multi-channel feature tensor X to eliminate the dimensional difference and ensure the convergence efficiency of model training:

[0084]

[0085] wherein, , min and max are the minimum and maximum values of the corresponding channel, is the normalized feature value;

[0086] Step three: construct an end-to-end hybrid model PTCA-Net composed of parallel time series convolution network (PTCN) and Transformer attention encoder and optimize the hyperparameters, wherein the PTCA-Net sequentially includes one-dimensional convolution layer, PTCN module, Transformer encoder module, residual connection, and full connection prediction layer from input to output; the tree structure Parzen estimator in the Bayesian optimization algorithm is used to optimize the network hyperparameters;

[0087] S3.1, the PTCA-Net sequentially includes one-dimensional convolution layer, PTCN module, Transformer encoder module, residual connection, and full connection prediction layer from input to output;

[0088] S3.1.1, one-dimensional convolution layer: the normalized feature value obtained in S2.6 Linear transformation is performed to map the channel dimension to the model adaptation dimension, preparing for subsequent feature extraction;

[0089] S3.1.2, PTCN module: contains 3 parallel time convolution network (TCN) branches, respectively using 3, 5, 7 convolution kernels of different sizes, small size convolution kernel matches the local trend feature of V-Q curve, captures the gentle change of voltage with capacity in short window, medium size convolution kernel can effectively extract the peak value feature of ICA curve, large size convolution kernel captures the fluctuation details of DVA curve, each branch extracts multi-scale local features through dilated causal convolution (DCC), the mathematical expression is:

[0090]

[0091] where K is the size of the convolution kernel, d is the dilation rate, is the convolution kernel weight, j is the current time step, x is the input feature, is the output feature;

[0092] Each TCN branch contains two time blocks, each time block integrates a DCC, followed by a GeLU activation function and group normalization, forming a time convolution network;

[0093] S3.1.3, Transformer encoder module: contains multi-head attention (MHA) layer and feed-forward network (FFN), used to model global temporal dependencies, MHA calculates through multiple sets of "query-key-value (Q-K-V)", enhancing the model's ability to capture complex patterns through parallel attention heads, FFN introduces nonlinear transformation to enhance the model's expression ability, two perceptrons apply GeLU activation function between linear transformations, enabling the model to learn complex nonlinear relationships in data;

[0094] S3.1.4, residual connection: to solve the problem of gradient disappearance of deep network and accelerate the convergence of training, residual connection is introduced inside each time block of PTCN, around MHA and FFN layers of Transformer encoder:

[0095]

[0096] where x is the input, F(x) is the block transformation, y is the output, i.e. the output of the Transformer encoder;

[0097] S3.1.5, fully connected prediction layer: maps the global features output by the Transformer encoder to a one-dimensional scalar, which is the SOH estimate value;

[0098] S3.2, the tree structure Parzen estimator (TPE) in the Bayesian optimization algorithm is used to optimize the model hyperparameters in the predefined search space, and the optimization goal is to minimize the SOH estimation error (RMSE) of the validation set;

[0099] S3.3, the Adam optimizer is used with the mean square error (MSE) as the loss function:

[0100]

[0101] wherein, is the actual SOH, is the predicted SOH; to avoid overfitting, if the validation set loss does not decrease for 20 consecutive epochs, the training is terminated;

[0102] Step four: input the normalized feature values of step S2.5 into the PTCA-Net model for training, extract multi-scale local attenuation features in the fragmented segment through the PTCA module, combine the global time sequence dependence captured by the Transformer encoder module, and output the SOH estimation value; determine the optimal window parameter according to the training results of different fragmented charging segments.

[0103] The two obtained normalized feature values are used as model inputs; the model constructed in S3 is trained in batch mode, the model performance is verified by test set data, and the average absolute error (MAE), average absolute percentage error (MAPE), root mean square error (RMSE), and determination coefficient (R2) of the predicted results and the true SOH are calculated. ).

[0104] The test set data is input to verify the network performance, and the relevant model parameters and test results are shown in Table 1:

[0105] Table 1

[0106]

[0107] As shown in Table 1, the initial voltage threshold Vstart is the most critical parameter affecting the accuracy of the model. As Vstart rises from 3.6V to 3.9V, RMSE and MAE continue to rise, while sharp decline, indicating that the estimation error gradually increases. This indicates that a higher starting voltage will result in a decrease in the amount of data window information, which in turn affects the performance of the model. The influence of the sampling interval Δt on the performance of the model is secondary. For a fixed Vstart, extending the sampling interval usually reduces the estimation error and improves The fundamental reason is that longer Δt can suppress high-frequency noise during data acquisition. This improves the signal-to-noise ratio, enabling the model to better identify key degradation patterns. The impact of the number of sampling points n is minimal. While the data in the table suggest that increasing n can reduce error, this improvement is generally limited. Increasing n from 16 to 32 when Vstart= 3.9 V and Δt = 5 seconds reduces RMSE by 11.1%, which is less than the improvement obtained by extending Δt. This suggests that larger n can increase computational complexity, but accuracy can not scale proportionally, potentially by introducing redundant information.

[0108] These results validate that the standardized window mechanism can effectively extract sufficient diagnostic information from short-term CC charging segments, making it unnecessary to rely on impractical full charging cycles. When deployed in practice on new battery systems, the voltage range where electrochemical characteristics change most significantly with aging can be determined through a one-time characterization phase, thereby determining the optimal parameters.

[0109] The embodiments of the present application are described in detail with reference to the accompanying drawings, but the present application is not limited to the described embodiments. For those skilled in the art, various changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the present application, and still fall within the scope of the present application.

Claims

1. A method for estimating the state of health of a lithium-ion battery based on fragmented charging data, characterized in that: Includes the following steps: Step 1: Obtain the raw data of the constant current charging (CC) stage of the lithium-ion battery. The raw data includes the charging voltage, charging current and corresponding timestamps. The charging protocol adopts the industry-standard constant current-constant voltage CC-CV mode. Based on preset window parameters, extract charging segments from the raw CC charging data and standardize them to form fragmented charging segments. Step 2: Calculate the cumulative charging capacity of the fragmented charging segments obtained in Step 1 using the Coulomb counting method, generate standardized multidimensional feature information, perform smoothing filtering and normalization preprocessing, and then convert it into a multi-channel feature tensor. Step 3: Construct an end-to-end hybrid model PTCA-Net and optimize its hyperparameters. The PTCA-Net, from input to output, includes a one-dimensional convolutional layer, a PTCN module, a Transformer encoder module, residual connections, and a fully connected prediction layer. The tree-structured Parzen estimator in the Bayesian optimization algorithm is used to optimize the network hyperparameters. Step 4: Input the normalized feature values ​​from Step 2 into the PTCA-Net model for training. Extract multi-scale local decay features from fragmented segments using the PTCN module. Combine this with the global temporal dependencies captured by the Transformer encoder module to output SOH estimates. Determine the optimal window parameters based on the training results of different fragmented charging segments.

2. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 1, characterized in that: Step one specifically includes: S1.1, Window parameters include initial voltage threshold Vstart, sampling interval Δt, and number of sampling points n; S1.2 and Vstart are set to 3.6V, 3.7V, 3.8V and 3.9V respectively. Within this voltage range, the electrochemical characteristics of the battery change most significantly with aging and the information density is the highest. Δt is set to 5s and 10s, and n is set to 16 and 32 respectively. S1.3 When the constant current charging CC stage voltage V(t) of the lithium-ion battery first reaches or exceeds Vstart, sampling is started, and n data points are collected at Δt intervals to form multiple fragmented charging segments.

3. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 1, characterized in that: Step two specifically includes: S2.1 Calculate the cumulative charging capacity by integrating the current according to Coulomb's law, and generate the voltage-capacity VQ curve; S2.

2. Differentiate the VQ curve to obtain the ICA curve, and inversely differentiate the VQ curve to obtain the DVA curve; S2.

3. A moving average filter is used to smooth and filter the VQ curve, ICA curve, and DVA curve to suppress current noise and high-frequency interference from voltage fluctuations. S2.

4. The three modal data corresponding to the smoothed VQ curve, ICA curve, and DVA curve, namely voltage V, capacity increment IC, and voltage increment DV, are concatenated into a 3×n multi-channel feature tensor X, where n is the sequence length. S2.

5. Perform min-max normalization independently on each channel of the multi-channel feature tensor X to eliminate dimensional differences.

4. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 1, characterized in that: Step three of PTCA-Net specifically includes: S3.1, from input to output, includes a one-dimensional convolutional layer, a PTCN module, a Transformer encoder module, a residual connection, and a fully connected prediction layer; S3.2 The one-dimensional convolutional layer is used to perform a linear transformation on the normalized feature values ​​and map them to the model adaptation dimension; S3.3 The PTCN module contains three parallel temporal convolutional network (TCN) branches, which use convolutional kernels of sizes 3, 5, and 7 respectively. Multi-scale local features are extracted by dilated causal convolution DCC. Each TCN branch contains two time blocks, and each time block integrates a DCC. Subsequently, GeLU activation function and group normalization are concatenated. S3.4 The Transformer encoder module includes a multi-head attention MHA layer and a feedforward network FFN. The MHA captures complex patterns through multiple sets of "query-key-value QKV" calculations, and the FFN introduces nonlinear transformations through two layers of perceptrons and the GeLU activation function. S3.5 The residual connection is applied within each time block of the PTCN and around the MHA and FFN layers of the Transformer encoder; S3.6 The fully connected prediction layer is used to map global features to one-dimensional SOH estimates.

5. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 1, characterized in that: The extraction of the charging segment is based on preset window parameters, which extract short-term charging segments from the original CC charging data.

6. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 3, characterized in that: The hyperparameters to be optimized in step three include the number of PTCN layers, the number of Transformer encoder layers, the initial learning rate, the dropout rate, the hidden size, the batch size, the number of MHA heads, and the compression scale of the final convolutional layers.

7. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 1, characterized in that: Step three also includes training the network using the Adam optimizer with mean squared error (MSE) as the loss function.

8. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 7, characterized in that: The loss function MSE in step three is derived from the actual SOH and the predicted SOH.

9. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 4, characterized in that: The dilated causal convolution DCC is derived based on the kernel size, dilation rate, kernel weights, current time step, and input features.

10. The method for estimating the state of health of a lithium-ion battery based on fragmented charging data according to claim 1, characterized in that: Step four specifically includes: using the normalized feature values ​​obtained in step two as model input; training the model constructed in step three using batch processing; verifying the model performance using test set data; and calculating the mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination between the predicted results and the true SOH. .

Images