Battery state of health online estimation method considering battery aging characteristics

Abstract

The application discloses a method for online estimation of state of health of storage batteries considering battery aging characteristics, and belongs to the technical field of online estimation of state of health of storage batteries. The method comprises the following steps: acquiring charging data of each single battery in a battery pack in real time, calculating a differential voltage curve, and constructing an inconsistency atlas containing a superimposed distribution map and a point-by-point differential matrix; inputting the inconsistency atlas into a double-task conditional generative adversarial network, and extracting three interpretable physical feature channels of peak value migration, peak value attenuation and shoulder broadening from a discriminator; inputting each single differential voltage curve into a pre-trained convolutional neural network to obtain deep features, and inputting the deep features into a multi-task prediction network after splicing with the interpretable physical feature channels, so as to simultaneously output a state of health value of the battery pack as a whole and a deviation degree of each single battery; and identifying a short-board single battery according to the deviation degree and giving a warning.

Description

Technical Field

[0001] This invention relates to the field of online estimation of battery health status, and more specifically to an online estimation method for battery health status that takes into account battery aging characteristics. Background Technology

[0002] As the core energy storage unit of electric vehicles and energy storage systems, the health status of batteries directly determines the driving range, safety and service life of the entire vehicle. State of Health (SOH) is a comprehensive evaluation index characterizing the degree of battery aging. Accurate online SOH estimation is of great significance for the optimized control of battery management systems and battery life management.

[0003] Currently, SOH estimation methods based on differential voltage curves or capacity increment curves are one of the research hotspots in this field. Existing technologies analyze the characteristics of capacity increment curves as the battery degrades, extract feature parameters, and establish a capacity estimation model using principal component regression, thereby achieving SOH estimation for individual cells. However, the aforementioned existing technologies still have the following shortcomings: Most existing methods are designed for individual cells and fail to effectively consider the impact of inconsistencies between multiple cells in a battery pack on SOH estimation. In practical applications, battery packs are composed of multiple cells connected in series or in parallel. Due to factors such as manufacturing process, temperature distribution, and charge / discharge rate differences, the aging speed and aging mode of each cell are significantly different. Existing methods treat the battery pack as a black box, ignoring the inconsistency information between cells, which limits the accuracy of SOH estimation and makes it impossible to identify the weakest cells in the pack. Existing methods rely on manually extracted health features, such as the peak height and peak position of the capacity increment curve. The design of these manually extracted features depends on expert experience and is difficult to fully characterize the multi-dimensional changes in electrochemical properties during battery aging. When the battery aging mode changes or the operating conditions change, the representativeness and robustness of the manually extracted features decrease, affecting the generalization ability of the model. The feature extraction process of existing deep learning methods lacks physical interpretability. Although some methods can learn features automatically, the deep features extracted are difficult to establish a correspondence with specific battery aging mechanisms such as loss of positive electrode active material and loss of available lithium ions. This makes it difficult for operation and maintenance personnel to understand and trust the output results of the model, thus limiting its application in battery life management decisions. Existing methods provide limited output information, typically only outputting the overall SOH value of the battery pack. They cannot provide information on the relative aging status of individual cells or the location of weak cells. In actual maintenance, maintenance personnel not only need to understand the overall health level of the battery pack, but also need to know which cells have experienced accelerated or abnormal aging in order to perform precise equalization maintenance or cell replacement. To address the aforementioned technical problems, this invention provides an online estimation method for the state of health (SOH) of a battery that takes into account battery aging characteristics. This method solves the technical problems in the prior art, such as the inability to estimate SOH using the inconsistency of individual cells within a battery pack, the lack of physical interpretability in feature extraction, and the inability to locate the weakest cells. Summary of the Invention

[0004] To overcome the problems of the prior art, the present invention discloses an online estimation method for battery health status that takes into account battery aging characteristics.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: An online method for estimating the state of health of a battery, taking into account battery aging characteristics, includes the following steps: S1, real-time acquisition of voltage and current data of each individual cell in the battery pack during the charging process, and calculation of the differential voltage curve of each individual cell within a preset voltage window. The differential voltage curve is the derivative curve of voltage with respect to capacity, and the capacity is obtained by time integration of the current data; S2, superimpose the differential voltage curves of all individual cells onto the same voltage coordinate system to construct an inconsistency map; The inconsistency map includes the superimposed distribution map of the differential voltage curves of each individual cell, as well as the point-by-point difference matrix of the differential voltage curve of each individual cell relative to the average differential voltage curve within the group. S3, input the inconsistency map into a pre-trained dual-task conditional generative adversarial network. The dual-task conditional generative adversarial network includes a generator and a discriminator. The discriminator simultaneously outputs a authenticity score and an aging score, and extracts multiple interpretable physical feature channels from the intermediate convolutional layers of the discriminator. The interpretable physical feature channels include at least: The peak migration characteristic channel characterizing the peak voltage offset of the differential voltage curve, the peak attenuation characteristic channel characterizing the peak amplitude attenuation of the differential voltage curve, and the shoulder broadening characteristic channel characterizing the peak broadening of the differential voltage curve. S4. Input the differential voltage curves of each individual cell into the pre-trained convolutional neural network to obtain deep features. The multiple interpretable physical feature channels extracted from the discriminator in step S3 are concatenated with the deep features. The concatenated features are then input into the multi-task prediction network, which simultaneously outputs the overall health status value of the battery pack and the degree of deviation of each individual cell relative to the average aging trend. S5. Identify the weakest cell in the group based on the deviation of each individual battery cell output in step S4, and output a warning signal when the deviation exceeds a preset threshold.

[0006] Preferably, the construction of the inconsistency map in step S2 specifically includes: The differential voltage curve of each individual cell is smoothed using Gaussian filtering. Calculate the arithmetic mean of the differential voltage curves of all individual cells in the group to obtain the average differential voltage curve of the group. The difference between the differential voltage curve of each individual cell and the average differential voltage curve within the group is calculated at each sampling point in the voltage coordinate system to obtain a point-by-point difference matrix. The rows of the point-by-point difference matrix correspond to each individual cell, and the columns correspond to each voltage sampling point. Each value in the point-by-point difference matrix represents the differential voltage deviation value of the corresponding cell at the corresponding voltage sampling point. The superimposed distribution map and the point-by-point difference matrix are used together as the output of the inconsistency map.

[0007] Preferably, the training method for the dual-task conditional generative adversarial network is as follows: The generator takes a random noise vector and conditional information as input to generate a virtual differential voltage curve; The discriminator takes the inconsistency map as input, and after processing by the convolutional layer, it outputs both the authenticity score and the aging score. During training, the generator and discriminator are updated alternately, and the features extracted by the discriminator have both authenticity and aging-related properties.

[0008] Preferably, the peak migration characteristic channel outputs a quantitative value that characterizes the degree to which the peak voltage position of the differential voltage curve migrates towards lower voltage as the battery ages. The quantitative value is positively correlated with the peak migration amount and is used to indicate the degree of loss of the positive electrode active material of the battery. The peak decay characteristic channel outputs a quantized value that characterizes the degree to which the peak amplitude of the differential voltage curve decreases with battery aging. The quantized value is positively correlated with the peak decay rate and is used to indicate the degree of loss of usable lithium ions in the battery. The shoulder broadening feature channel outputs a quantized value that characterizes the degree of peak broadening on both sides of the differential voltage curve. The quantized value is positively correlated with the degree of shoulder broadening of the curve and is used to indicate the degree of dispersion of battery aging mode.

[0009] Preferably, the pre-trained convolutional neural network is pre-trained using the differential voltage curves of individual cells in the publicly available battery aging dataset as input. The pre-training task is multi-task learning, and it outputs the health status estimate and aging stage classification results of the individual cells. When migrating to the battery pack scenario, the weights of the first two convolutional layers of the pre-trained convolutional neural network are frozen, and the weights of the last two convolutional layers are fine-tuned only.

[0010] Preferably, the multi-task prediction network in step S4 is a three-layer fully connected network, whose input is a concatenation of a 512-dimensional convolutional neural network feature vector and a 16-dimensional inconsistent feature vector; The output layer of the multi-task prediction network consists of one first neuron and M second neurons, where M equals the total number of individual cells in the battery pack. The first neuron outputs the overall health status value of the battery pack; Each second neuron uniquely corresponds to a single cell and is used to output the degree of deviation of the corresponding single cell from the average aging trend. The degree of deviation is a normalized value between 0 and 1.

[0011] Preferably, the preset threshold in step S5 is set as follows: Based on the consistency level of the differential voltage curves of each individual cell in the battery pack at the time of manufacture, the deviation degree of the cell with the worst consistency is used as the basic threshold. When the real-time deviation degree of a certain cell exceeds 1.2 times the basic threshold, the cell is identified as the weakest cell and a warning signal is output.

[0012] Preferably, the method further includes a closed-loop verification step: After each charge is completed, the newly acquired inconsistency map is compared with the offline calibration data, and the decay slope between the overall health status value obtained in this charge and the overall health status value obtained in the previous charge is calculated. If the decay slope deviates from the preset exponential decay model and exceeds the preset range, the model retraining instruction is triggered.

[0013] Preferably, the preset voltage window is the complete range from the starting voltage to the charging cutoff voltage during the constant current charging stage, or a portion of the range where the voltage is between 2.8V and 3.6V during the constant current charging stage.

[0014] Preferably, the method is executed online in the battery management system (BMS), and the sampling frequency for acquiring voltage and current data in step S1 is not less than 1 Hz.

[0015] The beneficial effects of this invention are as follows: Compared with the prior art, the technical solution provided by the present invention has the following significant advantages: This invention constructs an inconsistency map, transforming the differences in differential voltage curves between multiple individual cells within a battery pack into quantifiable feature inputs. It transforms the individual cell inconsistency, which is considered an interfering factor in traditional methods, into a core diagnostic feature, effectively utilizing the collaborative aging information between individual cells within the battery pack and improving the accuracy of online SOH estimation for the battery pack. This invention extracts peak migration feature channel, peak decay feature channel, and shoulder broadening feature channel from the discriminator of a dual-task conditional generative adversarial network. These three feature channels have clear physical correspondences with the loss of positive electrode active material, the loss of available lithium ions, and the degree of dispersion of aging mode, respectively. This solves the problem of lack of interpretability in feature extraction of existing deep learning methods, and enables the output of the model to be understood and trusted by operation and maintenance personnel. This invention uses a multi-task prediction network to simultaneously output the overall health status value of the battery pack and the deviation of each individual cell from the average aging trend, thereby realizing online identification and early warning of weak cells. This provides a direct basis for precise balanced maintenance of the battery pack and replacement of individual cells, overcoming the shortcomings of existing technologies in outputting only single information. This invention combines a pre-trained convolutional neural network with a transfer learning strategy. When transferring to a battery pack scenario, the weights of the first two convolutional layers are frozen and the weights of the last two convolutional layers are fine-tuned. This not only preserves the general curve shape features learned by the pre-trained model from large-scale single-cell battery data, but also enables the model to adapt to the inconsistency of single cells in the battery pack scenario, thereby improving the model's generalization performance under small sample conditions. This invention is executed online in the battery management system with a sampling frequency of no less than 1 Hz. After each charge is completed, a closed-loop verification step is automatically executed. The model retraining instruction is triggered based on the degree to which the decay slope of the overall health status value deviates from the exponential decay model. This achieves adaptive monitoring and automatic updating of the estimation model, ensuring the long-term estimation accuracy of the model under different aging stages and different operating conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of an online battery health status estimation method that takes into account battery aging characteristics, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0018] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a” and “the” as used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0020] Example 1 Please refer to Figure 1This embodiment provides an online estimation method for battery health status that takes into account battery aging characteristics. This method is executed online in the battery management system (BMS). In step S1, the battery management system (BMS) collects the voltage and current data of each individual cell in the battery pack in real time during the charging process. The charging process is preferably a constant current charging stage. For each individual cell, the BMS calculates its differential voltage curve within a preset voltage window. The differential voltage curve is defined as the derivative curve of voltage with respect to capacity, where the capacity is obtained by time integration of current data. The current data is integrated over time to obtain the capacity value. The derivative of voltage with respect to capacity is calculated, and this derivative can be approximated by the difference between adjacent sampling points. The preset voltage window can be the complete range from the starting voltage to the charging cutoff voltage during the constant current charging stage, or it can be a partial range between 2.8 volts and 3.6 volts during the constant current charging stage, depending on the battery type and actual operating conditions.

[0021] In step S2, the BMS superimposes the differential voltage curves of all individual cells onto the same voltage coordinate system to construct an inconsistency map. This inconsistency map consists of two parts: The first part is a superimposed distribution diagram of the differential voltage curves of each individual cell, which is to draw the differential voltage curves of all cells on the same voltage axis and differential voltage value axis, and intuitively display the overall distribution of each curve; The second part is the point-by-point difference matrix of the differential voltage curve of each individual cell relative to the average differential voltage curve within the group. The average differential voltage curve within the group is obtained by calculating the arithmetic mean of the differential voltages of all individual cells at the same voltage sampling point; For each voltage sampling point, the differential voltage value of each individual cell is subtracted from the average value within the group to obtain the differential voltage deviation value of that individual cell. The deviation values ​​of all cells and all sampling points together constitute a point-by-point difference matrix, where the rows of the matrix correspond to each cell, the columns correspond to each voltage sampling point, and each value in the matrix represents the differential voltage deviation value of the corresponding cell at the corresponding voltage sampling point.

[0022] In step S3, BMS inputs the inconsistency graph constructed in step S2 into the pre-trained dual-task conditional generative adversarial network; This dual-task conditional generative adversarial network includes a generator and a discriminator; The generator takes a random noise vector and conditional information as input. The conditional information includes numerical features such as the peak position, peak height and curve skewness of the differential voltage curve extracted from the current charging segment. The generator generates a virtual differential voltage curve. The discriminator takes the inconsistency map as input, processes it through multiple convolutional layers, and outputs the results of two branches simultaneously: The first branch outputs a authenticity score, which is used to determine whether the input inconsistency map is a real measurement or generated by the generator; The second branch outputs an aging score, which is used to estimate the battery aging level corresponding to the inconsistency pattern. During training, the generator and discriminator are updated alternately, so that the features extracted by the discriminator have both authenticity and aging relevance. The specific method for extracting interpretable physical feature channels from the intermediate convolutional layers of the discriminator is as follows: The discriminator employs a one-dimensional convolutional neural network structure, and its global response value is the output feature map of the second convolutional layer:

[0023] Wherein, the feature map output by the second convolutional layer of the F discriminator, R is the global response value, C=64 is the number of channels, and L is the feature map length, which depends on the number of voltage sampling points in the input inconsistency map. For each channel c, its global response value Rc is defined as:

[0024] That is, the average absolute value response of the feature map of that channel.

[0025] By employing a pre-designed supervised training strategy, different channels are forced to learn different aging-related features: When training a dual-task conditional generative adversarial network, in addition to the realism loss and aging loss, three auxiliary regression losses are added to supervise the mean square error between the response values ​​of the peak shift channel, peak decay channel and shoulder broadening channel and the actual physical quantities. For each training sample, i.e., the differential voltage curve for each battery aging cycle, three physical quantities are calculated offline: The peak mobility ΔVpeak is:

[0026] Wherein, Vinitial is the voltage value corresponding to the peak value of the differential voltage curve when the battery leaves the factory, and Vcurrent is the peak voltage value of the current cycle; The peak attenuation rate ΔHpeak is:

[0027] Where Initial is the initial peak height, and Hcurrent is the current peak height; The shoulder width (Wshoulder) is:

[0028] Where Vhigh and Vlow are the voltage boundaries corresponding to the half-peak widths on both sides of the peak, and Vrange is the width of the entire voltage window. Using these three physical quantities as auxiliary labels, during discriminator training, the global response values ​​Rc of three preset channels (channels 5, 12, and 23) are extracted from the output feature map of the second convolutional layer, respectively, and fitted to ΔVpeak, ΔHpeak, and Wshoulder. The loss function Lax is:

[0029] By backpropagation, the response values ​​of the three channels gradually approach the corresponding physical quantities, thus realizing supervised extraction of interpretable feature channels. In actual inference, i.e. online estimation, the global response values ​​R5, R12, and R23 of the three channels in the second convolutional layer of the discriminator are directly read and used as the output quantization values ​​of the peak shift feature channel, peak decay feature channel, and shoulder broadening feature channel, respectively. These quantitative values ​​are positively correlated with the degree of loss of positive electrode active material, the degree of loss of available lithium ions, and the degree of dispersion of aging mode, respectively, and are used to indicate the aging state of the battery.

[0030] In step S4, the BMS inputs the differential voltage curves of each individual cell into a pre-trained convolutional neural network to obtain deep features. The pre-trained convolutional neural network is pre-trained using the differential voltage curves of individual cells in the publicly available battery aging dataset as input. The pre-training task is multi-task learning, and it outputs the health status estimate of individual cells and the classification result of aging stage. When migrating to the battery pack scenario, the weights of the first two convolutional layers of the pre-trained convolutional neural network are frozen, and the weights of the last two convolutional layers are fine-tuned only. The structure of a pre-trained convolutional neural network is as follows: Input layer, 1×L one-dimensional differential voltage curve; Conv1, 32 3×1 convolutional kernels; Conv2, with 64 3×1 convolutional kernels; Conv3, with 128 3×1 convolutional kernels; Conv4, with 256 3×1 convolutional kernels; Global average pooling layer Fully connected layer, 512 dimensions; The output layer consists of two branches: SOH regression and aging classification.

[0031] Pre-training was performed on public datasets such as NASA and CALCE. The initial learning rate was set to 0.001, the optimizer was Adam, the batch size was 64, and the training lasted for 200 epochs. The early stopping condition was that the validation set loss did not decrease for 10 consecutive epochs. When migrating to the target battery pack scenario, perform the following fine-tuning steps: Load all weight parameters obtained from pre-training; Set the requiresgrad attribute of the weight parameters of Conv1 and Conv2 to False, which freezes these two layers and prevents gradient calculation and weight updates during subsequent training. The requiresgrad property of the Conv3 layer, Conv4 layer, global average pooling layer, fully connected layer, and output layer is set to True to allow updates. The learning rate during the fine-tuning phase is set to one-tenth of the initial learning rate during pre-training, i.e., 0.0001. The optimizer remains Adam, and the batch size is set to 16 because the amount of data in the target scenario is relatively small. The loss function in the fine-tuning stage is the same as that in the pre-training stage, which is the sum of the mean squared error loss of SOH regression and the cross-entropy loss of aging classification, with a weight ratio of 1:1. The differential voltage curves of each cell in the target battery pack during the first full charging cycle are used as training data. Typically, N=100~200 cells are used for training for 50 epochs. The early stopping condition is that the loss of the validation set does not decrease for 5 consecutive epochs. The validation set is randomly selected from 20% of the training data for validation. After fine-tuning, the frozen Conv1 and Conv2 layers remain unchanged, and feature extraction is performed only using the updated Conv3, Conv4 and subsequent layers.

[0032] Through the above freezing and fine-tuning strategies, the pre-trained model can adapt to the data distribution of the target battery pack while avoiding overfitting to small samples.

[0033] BMS concatenates the multiple interpretable physical feature channels extracted from the discriminator in step S3 with the aforementioned deep features; The concatenated feature vector is input into a multi-task prediction network, which is a three-layer fully connected network. Its input dimension is the sum of the 512 dimensions of the convolutional neural network feature vector and the 16 dimensions of the inconsistent feature vector, i.e., 528 dimensions. The output layer consists of one first neuron and M second neurons, where M equals the total number of individual cells in the battery pack; The first neuron outputs the overall health status value of the battery pack, which is a percentage between 0 and 100, representing the remaining capacity percentage of the battery pack. Each second neuron uniquely corresponds to a single cell. The i-th second neuron outputs the degree of deviation of the i-th single cell from the average aging trend. This degree of deviation is a normalized value between 0 and 1, where 0 indicates that the cell is completely consistent with the average aging level in the group, and 1 indicates a serious deviation. In step S5, the BMS identifies the weakest cell within the group based on the deviation of each individual cell output in step S4. The preset threshold is set as follows: Based on the consistency level of the differential voltage curves of each individual cell in the battery pack at the time of manufacture, the deviation degree of the cell with the worst consistency is used as the basic threshold. When the real-time deviation of a single cell exceeds 1.2 times the basic threshold, the BMS identifies the single cell as a weak link and outputs a warning signal. This warning signal can be used to trigger the active balancing strategy of the battery management system or to remind maintenance personnel to replace individual cells; The method in this embodiment also includes a closed-loop verification step after each charging is completed: The newly acquired inconsistency map is compared with the offline calibration data. The decay slope between the overall health status value obtained in this charge and the overall health status value obtained in the previous charge is calculated. If the decay slope deviates from the preset exponential decay model and exceeds the preset range, the model retraining instruction is triggered to achieve online adaptive update.

[0034] Specifically, this embodiment transforms the inconsistencies between individual cells within the battery pack into core diagnostic features by constructing an inconsistency map. It then uses a dual-task conditional generative adversarial network to extract interpretable feature channels with clear physical meanings. Combined with transfer learning and a multi-task prediction network, it simultaneously outputs the overall health status of the battery pack and the degree of deviation of each individual cell, thereby achieving online identification and early warning of weak cells and significantly improving the accuracy and interpretability of online estimation of battery health status.

[0035] Furthermore, in the aforementioned step S1, the battery management system (BMS) collects the voltage and current data of each individual cell in the battery pack during the charging process in real time. For each individual cell, the BMS calculates its differential voltage curve within a preset voltage window. The differential voltage curve is defined as the derivative curve of voltage with respect to capacity, where capacity is obtained by time integration of current data; The current data is integrated over time to obtain the capacity value. The derivative of voltage with respect to capacity is calculated, and this derivative can be approximated by the difference between adjacent sampling points. Specifically, the central difference method is used to improve accuracy. For the k-th sampling point, the differential voltage value D is:

[0036] Where Vk and Qk are the voltage and capacity values ​​at the kth sampling point, respectively; For the first and last endpoints, forward differential or backward differential is used. To reduce the impact of measurement noise on the derivative calculation, the voltage and capacity sequences are first subjected to Savitzky-Golay filtering before calculating the differential. The window length is set to 5 and the polynomial order is 2 before calculating the derivative. The preset voltage window can be the complete range from the starting voltage to the charging cutoff voltage during the constant current charging stage, or it can be a partial range between 2.8 volts and 3.6 volts during the constant current charging stage, depending on the battery type and actual operating conditions. In step S2 above, the BMS superimposes the differential voltage curves of all individual cells onto the same voltage coordinate system to construct an inconsistency map. The construction process of this inconsistency map specifically includes the following sub-steps: The differential voltage curve of each individual cell is smoothed by Gaussian filtering. The Gaussian filtering smoothing process uses the Gaussian function as the filtering kernel and performs a weighted average on each sampling point on the differential voltage curve and the data points in its neighborhood to filter out high-frequency noise and random interference introduced during the measurement process. The smoothed differential voltage curve retains the main waveform shape reflecting the battery aging characteristics, while reducing the impact of noise on subsequent calculations. Calculate the arithmetic mean of the differential voltage curves of all individual cells in the group to obtain the average differential voltage curve of the group. At each sampling point in the voltage coordinate system, the differential voltage values ​​of all individual cells after smoothing are added together and then divided by the total number of individual cells. The resulting value is the average differential voltage value within the group at that sampling point. Arrange the average differential voltage values ​​within the group of all sampling points in voltage order to obtain the average differential voltage curve within the group; The difference between the differential voltage curve of each individual cell and the average differential voltage curve within the group is calculated at each sampling point in the voltage coordinate system to obtain a point-by-point difference matrix. For the i-th cell and the k-th voltage sampling point, the differential voltage value of the cell at the sampling point is subtracted from the differential voltage value of the average differential voltage curve within the group at the sampling point. The resulting difference is the differential voltage deviation value of the cell at the sampling point. Perform the above difference calculation on each individual cell and each voltage sampling point to obtain a point-by-point difference matrix composed of all deviation values; The rows of the point-by-point difference matrix correspond to each individual cell, and the columns correspond to each voltage sampling point. Each value in the matrix represents the differential voltage deviation value of the corresponding cell at the corresponding voltage sampling point. This matrix quantifies the degree and pattern of deviation of the differential voltage curve of each individual cell from the average level within the group; The superimposed distribution map obtained in the preceding steps and the point-by-point difference matrix are used together as the output inconsistency map. The overlay distribution diagram is a graphic plot of the differential voltage curves of all individual cells on the same voltage axis and differential voltage value axis, which intuitively shows the overall distribution shape of each curve and the positional relationship between them. The point-by-point difference matrix quantifies the deviation of each individual curve in numerical form. Together, they form an inconsistency map, which provides both visualized curve distribution information and quantified deviation data, providing complete input features for the deep learning model in subsequent steps. In step S3 above, BMS inputs the constructed inconsistency map into the pre-trained dual-task conditional generative adversarial network to extract interpretable physical feature channels. In the aforementioned step S4, the BMS inputs the differential voltage curves of each individual cell into a pre-trained convolutional neural network to obtain deep features, and then concatenates the interpretable physical feature channels with the deep features and inputs them into a multi-task prediction network. At the same time, it outputs the overall health status value of the battery pack and the degree of deviation of each individual cell from the average aging trend. In the aforementioned step S5, the BMS identifies the weakest cell in the group based on the degree of deviation of each individual cell, and outputs a warning signal when the degree of deviation exceeds a preset threshold. Specifically, Gaussian filtering is used to smooth the differential voltage curve to reduce the impact of measurement noise. Inconsistency maps are constructed by calculating the average curve within the group and the point-by-point difference matrix. The intuitive visualization information of the superimposed distribution map is combined with the quantization deviation information of the point-by-point difference matrix, providing a complete feature input for subsequent deep learning models that reflects the aging inconsistency between individual cells within the battery pack.

[0037] Further, as described in steps S1 and S2 above, the BMS acquires the charging data of each individual cell in the battery pack in real time, calculates the differential voltage curve of each individual cell, and constructs an inconsistency map. The inconsistency map includes the superimposed distribution map of the differential voltage curves of each individual cell and the point-by-point difference matrix of the differential voltage curve of each individual cell relative to the average differential voltage curve in the group. In step S3, the BMS inputs the inconsistency map into the pre-trained dual-task conditional generative adversarial network, which is trained as follows: The generator takes a random noise vector and conditional information as input to generate a virtual differential voltage curve. The random noise vector is a sequence of random numbers sampled from a standard normal distribution to introduce diversity into the generation process. Conditional information is numerical features extracted from the inconsistency profile, including the peak position, peak height, and curve skewness of the differential voltage curve, which guides the generator to generate a virtual curve that matches a specific aging state. The discriminator takes the inconsistency map as input, and after processing by the convolutional layer, it outputs both the authenticity score and the aging score. The discriminator uses a one-dimensional convolutional neural network structure to extract features from the inconsistency map by passing it through multiple convolutional layers in sequence. Each convolutional layer contains several convolutional kernels. After processing by the convolutional layer, the discriminator splits into two branches: The first branch outputs an authenticity score, which indicates whether the input map is real measurement data or virtual data generated by the generator. The second branch outputs an aging score, which represents an estimate of the battery aging level corresponding to the input graph. During training, the generator and discriminator are updated alternately. The parameters of the generator are fixed, and the real inconsistency spectrum and the virtual differential voltage curve generated by the generator are respectively input into the discriminator. The loss function of the discriminator is calculated and its parameters are updated. With the discriminator parameters fixed, a random noise vector and conditional information are input into the generator to produce a virtual differential voltage curve. This virtual curve is then input into the discriminator to calculate the generator's loss function and update its parameters. This process is repeated until the network converges. Through the above-mentioned alternating update training method, the features extracted by the discriminator have both authenticity and aging correlation. The authenticity features enable the discriminator to accurately distinguish between real and virtual maps, while the aging correlation features enable the discriminator's intermediate convolutional layers to automatically learn feature representations that are closely related to the battery aging state. These features can be extracted from the intermediate convolutional layers of the discriminator as interpretable physical feature channels in subsequent steps; In the aforementioned step S4, the BMS inputs the differential voltage curves of each individual cell into a pre-trained convolutional neural network to obtain deep features, and concatenates multiple interpretable physical feature channels extracted from the discriminator with the deep features, inputs them into a multi-task prediction network, and outputs the overall health status value of the battery pack and the degree of deviation of each individual cell from the average aging trend. In the aforementioned step S5, the BMS identifies the weakest cell in the group based on the degree of deviation of each individual cell, and outputs a warning signal when the degree of deviation exceeds a preset threshold. Specifically, by alternating the training of the generator and the discriminator, the features extracted by the discriminator possess both authenticity and aging correlation, providing a high-quality feature foundation for subsequent extraction of interpretable physical feature channels and SOH multi-task prediction.

[0038] Further, in step S3, BMS inputs the inconsistency map into the pre-trained dual-task conditional generative adversarial network and extracts multiple interpretable physical feature channels from the intermediate convolutional layer of the discriminator, including at least the peak transfer feature channel, the peak decay feature channel and the shoulder broadening feature channel. The peak migration characteristic channel output is a quantified value representing the degree to which the peak voltage position of the differential voltage curve migrates towards lower voltage as the battery ages. The quantization value is positively correlated with the peak shift, that is, the greater the distance the peak voltage position shifts towards the lower voltage direction, the higher the quantization value output by the peak shift characteristic channel. The peak migration characteristic channel is used to indicate the degree of loss of active material in the positive electrode of the battery; Loss of positive electrode active material is one of the main aging modes of lithium-ion batteries, which manifests as degradation of the positive electrode material structure, detachment of active particles from the conductive network or lattice distortion, resulting in a reduction in the number of lithium ions that can be inserted and extracted. The peak voltage position in the differential voltage curve corresponds to the voltage plateau of a specific electrochemical reaction inside the battery. When the loss of positive electrode active material intensifies, this voltage plateau shifts towards a lower voltage direction. The peak migration characteristic channel captures this migration trend and outputs a quantitative migration degree, providing an online diagnostic basis for the degree of loss of positive electrode active material in the battery. The peak attenuation characteristic channel output is a quantified value representing the degree to which the peak amplitude of the differential voltage curve decreases with battery aging; This quantization value is positively correlated with the peak attenuation rate; that is, the greater the reduction in peak amplitude, the higher the quantization value output by the peak attenuation characteristic channel. The peak decay characteristic channel is used to indicate the degree of loss of available lithium ions in the battery. The loss of available lithium ions is another key aging mode that leads to battery capacity decay, mainly caused by the continuous growth of the solid electrolyte interface film on the negative electrode surface and the consumption of active lithium by side reactions. The magnitude of the peak amplitude in the differential voltage curve is directly related to the amount of active lithium participating in the electrochemical reaction. When the loss of available lithium ions intensifies, the amount of active lithium participating in the reaction decreases, and the peak amplitude decreases accordingly. The peak decay feature channel quantifies the degree of reduction in peak amplitude and outputs an online estimate of the degree of available lithium-ion loss. The shoulder broadening characteristic channel outputs a quantized value that characterizes the degree of peak broadening on both sides of the peak of the differential voltage curve; The quantization value is positively correlated with the degree of shoulder widening of the curve; that is, the more flat and wide the peak shape on both sides of the peak is, the higher the quantization value output by the shoulder widening feature channel. The shoulder-widening characteristic channel is used to indicate the degree of dispersion of battery aging mode. When inconsistent aging phenomena occur inside the battery, such as local lithium plating, uneven loss of electrode materials, or uneven distribution of electrolyte, the peak shape of the differential voltage curve changes from sharp and concentrated to broad and diffuse, with a widening characteristic on both sides of the peak. The shoulder broadening feature channel captures the changing trend of this peak broadening and outputs a quantitative value of the dispersion of the aging mode, providing an online diagnostic basis for identifying whether the battery has undergone local abnormal aging. After the three interpretable physical feature channels mentioned above are extracted from the intermediate convolutional layer of the discriminator, they proceed to step S4 for further processing. In the aforementioned step S4, the BMS inputs the differential voltage curves of each individual cell into a pre-trained convolutional neural network to obtain deep features, and concatenates the peak migration feature channel, peak decay feature channel, and shoulder broadening feature channel extracted from the discriminator with the deep features. The concatenated features are then input into a multi-task prediction network, which simultaneously outputs the overall health status value of the battery pack and the degree of deviation of each individual cell from the average aging trend. Each of these three feature channels outputs a quantitative value with a clear physical meaning, corresponding to the degree of loss of positive electrode active material, the degree of loss of available lithium ions, and the degree of dispersion of aging mode, respectively. This provides interpretable aging mechanism feature inputs for the multi-task prediction network and enhances the model's comprehensive judgment ability on battery aging status. In the aforementioned step S5, the BMS identifies the weakest cell in the group based on the degree of deviation of each individual battery cell output in step S4, and outputs a warning signal when the degree of deviation exceeds a preset threshold. Specifically, this embodiment outputs a quantized value that is positively correlated with the peak migration amount through the peak migration feature channel and indicates the degree of loss of positive electrode active material; it outputs a quantized value that is positively correlated with the peak decay rate through the peak decay feature channel and indicates the degree of loss of usable lithium ions; and it outputs a quantized value that is positively correlated with the degree of curve shoulder broadening through the shoulder broadening feature channel and indicates the degree of dispersion of battery aging mode. This realizes the automatic extraction of interpretable physical features with clear aging mechanism correspondence from differential voltage curves, providing aging feature inputs with both interpretability and high discriminative power for subsequent multi-task prediction networks.

[0039] Furthermore, in step S4, BMS performs the training and transfer process of the pre-trained convolutional neural network, specifically as follows: Pre-training of a pre-trained convolutional neural network is performed, which is pre-trained with the differential voltage curves of individual cells in a publicly available battery aging dataset as input. The publicly available battery aging datasets are selected from one or more of the NASA lithium-ion battery aging dataset, the CALCE lithium-ion battery aging dataset, and the Oxford battery aging dataset. In these data sets, the differential voltage curve of each individual cell is obtained by calculating the voltage and current data during the constant current charging phase. Specifically, the calculation method is the derivative curve of voltage with respect to capacity, where the capacity is obtained by integrating the current data over time. The structure of the pre-trained convolutional neural network is a one-dimensional convolutional neural network, which contains multiple convolutional layers. Each convolutional layer contains several convolutional kernels, which are used to extract local feature patterns in the differential voltage curve. The pre-training task is multi-task learning, that is, the pre-trained convolutional neural network performs two output tasks simultaneously; The first output task is a regression task, which outputs the health status estimate of a single battery cell. This health status estimate is a percentage between 0 and 100, representing the remaining capacity percentage of the single battery cell. The mean squared error loss function is used for optimization. The second output task is a classification task, which outputs the classification results of the aging stage of a single battery cell. The aging stage is divided into three categories: early stage, middle stage and late stage. The cross-entropy loss function is used for optimization. The total loss function of the pre-training is the sum of the regression task loss and the classification task loss, or the weighted sum of the regression task loss and the classification task loss. The weights of all convolutional layers of the pre-trained convolutional neural network are iteratively updated through the backpropagation algorithm until the network converges. After pre-training is completed, the pre-trained convolutional neural network is transferred to the battery pack scenario. During the transfer process, the pre-trained convolutional neural network is fine-tuned. Freeze the weights of the first two convolutional layers of the pre-trained convolutional neural network. That is, in the first two convolutional layers, the weight parameters remain unchanged in the subsequent training process and do not participate in gradient updates. Only the weights of the last two convolutional layers are fine-tuned, meaning that the weight parameters of the last two convolutional layers continue to be updated and optimized during subsequent training. The reason for freezing the first two convolutional layers is that the shallow convolutional layers learn the basic shape features of the differential voltage curve, including the shape of the rising edge of the curve, the outline of the peak region, and the trend of the falling edge. These basic shape features are universal under different battery types, different aging levels, and different operating conditions, and do not need to be relearned for battery pack scenarios. The reason for fine-tuning the last two convolutional layers is that the deep convolutional layers learn more abstract aging pattern features, including peak migration patterns, peak decay trends, and peak broadening characteristics. These features are closely related to the differences in specific battery packs and operating conditions, and need to be adjusted specifically using actual data from battery pack scenarios to improve the model's representation ability and estimation accuracy in the target scenario.

[0040] During the fine-tuning process, the differential voltage curve data of each individual battery cell collected in the battery pack scenario are used to train the weights of the last two convolutional layers. The fine-tuned learning rate is set to one-tenth of the initial learning rate in the pre-training stage. The weights of the last two convolutional layers are finely adjusted with a small step size to avoid destroying the general features already learned in the pre-training stage. Gradient calculation for the first two convolutional layers is disabled, and the weight parameters remain unchanged from their pre-training values. Through the aforementioned freezing and fine-tuning strategies, the pre-trained convolutional neural network, while maintaining its general feature extraction capabilities, adapts to the inconsistency of individual cells in the battery pack scenario, achieving effective transfer from individual cell data to the battery pack scenario.

[0041] After pre-training and transfer learning, the pre-trained convolutional neural network is used for deep feature extraction in step S4. The differential voltage curves of each individual cell are input into the convolutional neural network that has been pre-trained and fine-tuned as described above to obtain deep features. This deep feature is the feature vector output by the last convolutional layer or fully connected layer of the convolutional neural network, containing 512 feature values, which comprehensively represents the multi-dimensional information related to the battery health status in the differential voltage curve; In step S4 above, BMS also concatenates the multiple interpretable physical feature channels extracted from the discriminator in step S3 with the above deep features, and inputs the concatenated features into the multi-task prediction network. The multi-task prediction network simultaneously outputs the overall health status value of the battery pack and the degree of deviation of each individual cell relative to the average aging trend. In the aforementioned step S5, the BMS identifies the weakest cell in the group based on the degree of deviation of each individual battery cell output in step S4, and outputs a warning signal when the degree of deviation exceeds a preset threshold. Specifically, the convolutional neural network is pre-trained using the differential voltage curves of individual cells in a publicly available battery aging dataset as input. Multi-task learning is set up to simultaneously output health status estimates and aging stage classification results. When transferring to the battery pack scenario, the weights of the first two convolutional layers are frozen and only the weights of the last two convolutional layers are fine-tuned. This preserves the general curve shape features learned by the pre-trained model from large-scale individual cell data and enables the model to adapt to the inconsistency of individual cells in the battery pack scenario, thereby improving the model's generalization ability and estimation accuracy in small sample scenarios.

[0042] Furthermore, in step S4, the BMS inputs the differential voltage curves of each individual cell into a pre-trained convolutional neural network to obtain deep features. This deep feature is a 512-dimensional convolutional neural network feature vector, the dimension of which is determined by the structure of the pre-trained convolutional neural network; The pre-trained convolutional neural network is connected to a global average pooling layer or a flattening layer after the last convolutional layer, and outputs a feature vector of fixed length 512. This feature vector comprehensively represents the multi-dimensional deep information related to the battery health status in the differential voltage curve. BMS also transforms the multiple interpretable physical feature channels extracted from the discriminator in step S3 into a 16-dimensional inconsistency feature vector; The 16-dimensional inconsistency feature vector is composed of the output values ​​of the peak shift feature channel, the peak decay feature channel, and the shoulder widening feature channel, as well as other auxiliary features extracted from the intermediate convolutional layers of the discriminator; The peak migration feature channel, peak decay feature channel, and shoulder widening feature channel each output a quantized value. Together with 13 auxiliary feature values ​​extracted from other convolutional channels of the discriminator, they form a 16-dimensional inconsistency feature vector. This vector quantifies the degree and pattern of aging inconsistency between individual cells in the battery pack. BMS concatenates the 512-dimensional convolutional neural network feature vector with the 16-dimensional inconsistent feature vector to obtain a 528-dimensional concatenated feature vector. The concatenation operation is performed in the order of vector dimensions, that is, the 16-dimensional inconsistent feature vector is appended to the 512-dimensional convolutional neural network feature vector to form a new feature vector. The concatenated feature vector is input into a multi-task prediction network, which is a three-layer fully connected network. The first fully connected layer has an input dimension of 528 and an output dimension of 256. It is subjected to a linear transformation followed by a modified linear unit activation function. The second fully connected layer has an input dimension of 256 and an output dimension of 128. It also uses a linear transformation followed by a modified linear unit activation function. The third layer is the output layer, which has an input dimension of 128 and contains two neurons. It either does not use an activation function or uses a linear activation function. The first neuron in the output layer outputs the overall health status value of the battery pack, and the M second neurons output the deviation degree of each individual battery cell respectively. The overall health status value is a percentage between 0 and 100, representing the remaining capacity percentage of the battery pack. Its calculation formula is based on the nonlinear mapping relationship learned by the multi-task prediction network, which integrates the information of the 512-dimensional convolutional neural network feature vector and the 16-dimensional inconsistency feature vector. The deviation of each individual cell is calculated as follows: The multi-task prediction network outputs an original bias score vector of the same length as the number of individual batteries on each of its M second neurons. Each element represents the original bias score of the corresponding individual battery. The score is then mapped to the interval between 0 and 1 by the Sigmoid function to obtain the final normalized bias level. The output of the i-th second neuron is the normalized bias level of the i-th individual battery. A value of 0 indicates that the monomer is completely consistent with the average aging level within the group, a value of 1 indicates that the monomer deviates significantly from the average aging trend, and values ​​between 0 and 1 indicate different degrees of deviation. This degree of deviation is used to identify weak monomers in subsequent steps. In the aforementioned step S5, the BMS identifies the weakest cell in the group based on the degree of deviation of each individual battery cell output in step S4, and outputs a warning signal when the degree of deviation exceeds a preset threshold. Specifically, by concatenating the 512-dimensional convolutional neural network feature vector with the 16-dimensional inconsistency feature vector and inputting it into a three-layer fully connected network, the first neuron of the output layer outputs the overall health status value of the battery pack, and the M second neurons output the deviation degree of each individual battery cell respectively. The deviation degree is then normalized to a value between 0 and 1, thereby achieving a joint quantitative assessment of the overall aging level of the battery pack and the inconsistency of individual cells. This provides an accurate and comparable basis for the subsequent identification and early warning of weak cells.

[0043] Furthermore, in step S5, the BMS identifies the weakest cell in the group based on the deviation of each individual battery cell output in step S4, and outputs a warning signal when the deviation exceeds a preset threshold. The preset threshold is set as follows: To obtain the consistency level of the differential voltage curves of each individual cell in the battery pack at the time of leaving the factory, the battery pack is fully charged for the first time under standard operating conditions at the time of leaving the factory. The BMS collects the voltage and current data of each individual cell, calculates the differential voltage curve of each individual cell, and constructs the inconsistency map under the factory condition according to the method of step S2 mentioned above. Based on this inconsistency map, the initial deviation of each individual cell under factory conditions is output by the multi-task prediction network in step S4. This set of initial deviations reflects the inherent minor differences between individual cells in a brand-new battery pack, including consistency deviations caused by factors such as manufacturing process errors and material inhomogeneity. The deviation of the worst-performing cell is used as the base threshold. The worst-performing cell refers to the cell with the largest initial deviation in the factory state. The deviation value of this cell reflects the upper limit of the difference between cells when the battery pack leaves the factory. The deviation value corresponding to this cell is set as the base threshold, denoted as Tbase. During the subsequent use of the battery pack, the BMS calculates the real-time deviation of each individual cell. When the real-time deviation of a single cell exceeds 1.2 times the base threshold, that is, when the real-time deviation is greater than 1.2 times Tbase, the BMS determines that the single cell is a weak cell and outputs a warning signal. Warning signals can take the form of triggering the fault indicator light on the battery management system display, sending fault codes via the vehicle network, recording event logs, or sending notifications to the remote monitoring platform. Once a weak cell is identified, the BMS can take further proactive balancing measures or remind maintenance personnel to replace the cell.

[0044] The aforementioned coefficient of 1.2 is an empirical value obtained through statistical analysis of multiple sets of battery aging test data; During the normal aging process of the battery pack, the degree of deviation of each cell will slowly increase with the number of cycles, but it usually will not exceed 1.2 times the factory basic threshold. When the deviation of a certain monomer exceeds this threshold, it indicates that the monomer has undergone accelerated aging or abnormal aging, which requires attention. If a cell is identified as a weak cell, the BMS continuously monitors the trend of the deviation of that cell. If the deviation of that cell falls back to less than 1.2 times the base threshold during subsequent charging, the warning status is lifted. If the threshold is exceeded for three consecutive charging cycles, a persistent warning signal will be output, indicating that the cell has entered the accelerated aging stage and timely maintenance is recommended.

[0045] After step S5 mentioned above, the BMS stores the results of the faulty single-cell determination and the warning information in the local memory, and uploads them to the cloud server through the communication interface for the whole life cycle health management of the battery pack. Specifically, by obtaining the consistency level of the differential voltage curves of each individual cell in the battery pack at the time of manufacture, the deviation of the cell with the worst consistency is used as the basic threshold. When the real-time deviation of a certain cell exceeds 1.2 times the basic threshold, it is identified as a weak cell and an early warning signal is output. This achieves adaptive identification of abnormally aged cells in the battery pack. This threshold setting method fully considers the individual differences in the consistency of different battery packs at the time of manufacture, and avoids misjudgment or omission caused by using a fixed threshold.

[0046] Furthermore, in the above online estimation process, the method also includes a closed-loop verification step, specifically as follows: After each charge is completed, the BMS performs closed-loop verification, comparing the newly acquired inconsistency map during the charging process with the offline calibration data; Offline calibration data is an inconsistency map collected and saved under standard operating conditions during a full charge before the battery pack leaves the factory or when it is first put into use. This map represents the baseline characteristics of the battery pack in its initial healthy state. The comparison is made by calculating the similarity or difference between the newly acquired inconsistency map and the offline calibration data. Specifically, the root mean square error of the point-by-point difference matrix or the structural similarity index can be used for quantification. If the difference between the two exceeds the preset similarity threshold, it indicates that the internal characteristics of the battery pack have changed significantly, and the reliability of the estimation results needs to be further verified. Calculate the decay slope between the overall health status value obtained in this charge and the overall health status value obtained in the previous charge; The overall health status value is output by the multi-task prediction network in step S4 and stored in the BMS's storage unit. It is updated after each charge. The formula for calculating the decay slope is: The difference between the current overall health status value and the previous overall health status value is divided by the time interval or the difference in the number of cycles between the two charging sessions. The unit of this degradation slope can be a percentage per day or a percentage per cycle, indicating the rate at which the battery pack's health changes over time or the number of uses.

[0047] The calculated attenuation slope is compared with the preset exponential attenuation model, which is obtained in advance by fitting the historical aging data of the same type of battery. Its mathematical form is that the overall health status value decreases exponentially with time or number of cycles. The model provides a theoretical range of the decay slope under normal aging conditions. The preset range is an interval formed by extending a certain tolerance above and below the decay slope predicted by the exponential decay model. The size of the tolerance is determined based on the statistical distribution of historical data, for example, taking twice the standard deviation. If the calculated decay slope deviates from the preset exponential decay model by more than the preset range, that is, the decay slope is less than the lowest value predicted by the model or greater than the highest value predicted by the model, then the model retraining instruction is triggered. A low decay slope indicates that the battery is aging slower than expected, which may be due to overestimation error in the model. A high decay slope indicates that the battery is aging faster than expected, which may be due to underestimation error in the model or abnormal accelerated aging of the battery. After triggering the model retraining command, the BMS will package all the data collected during this charging process, including inconsistency maps, differential voltage curves, overall health status values ​​and the degree of deviation of each individual unit, and send them to the cloud server or local training platform through the communication interface. The platform uses the newly accumulated data to incrementally train or fine-tune the pre-trained convolutional neural network and multi-task prediction network. After updating the model parameters, the new model is sent to BMS for subsequent online estimation. If the decay slope does not exceed the preset range, the model retraining instruction will not be triggered, and the BMS will continue to use the current model for online estimation. At the same time, the data from this charging will be included in the historical database for subsequent adjustment of the parameters and preset range of the exponential decay model. This closed-loop verification step is executed automatically after each charge, requiring no manual intervention.

[0048] Specifically, by comparing the newly acquired inconsistency map with the offline calibration data after each charging, the decay slope of the overall health status value is calculated, and a model retraining instruction is triggered when the slope deviates from the preset exponential decay model beyond the preset range. This achieves adaptive monitoring and automatic updating of the online estimation model, ensuring the long-term estimation accuracy of the model under different aging stages and different operating conditions.

[0049] In this embodiment, the selection method of the preset voltage window has been fully defined in step S1. The preset voltage window adopts one of the following two methods: In the first method, the preset voltage window is the complete range of voltage rise from the starting voltage to the charging cutoff voltage during the constant current charging stage. The starting voltage is the initial voltage value when the battery pack starts constant current charging, and the charging cutoff voltage is the upper limit of voltage when the battery pack reaches full charge. Using the complete range as the voltage window, we can obtain all the morphological information of the differential voltage curve throughout the entire charging process, including the plateau region of the low voltage region, the peak region of the medium voltage region, and the cutoff region of the high voltage region, providing the most comprehensive data foundation for the subsequent construction of inconsistency maps and feature extraction.

[0050] In the second method, the preset voltage window is a portion of the voltage range between 2.8 volts and 3.6 volts during the constant current charging stage. This voltage range corresponds to the main discharge platform region of ternary lithium-ion batteries and lithium iron phosphate batteries. Within this region, the peak characteristics of the differential voltage curve are most significant and the variation pattern is most stable. Selecting this range as the voltage window can shorten the data acquisition time and reduce the computational complexity while ensuring the acquisition of key aging characteristics. It is particularly suitable for online application scenarios where the charging process is incomplete or the charging time is limited.

[0051] In practical applications, either of the two voltage windows can be flexibly selected based on battery type, charging conditions, and computing resources.

[0052] Furthermore, the online estimation method for battery health status that takes into account battery aging characteristics provided in this embodiment is executed online in the battery management system (BMS). The BMS is the core management unit of the battery pack and has functions such as data acquisition, status calculation, communication interaction, and protection control. In step S1, the BMS acquires the voltage and current data of each individual cell in the battery pack during the charging process in real time. The sampling frequency for acquiring voltage and current data should not be less than 1 Hz. The specific setting of the sampling frequency can be determined according to the hardware performance of the BMS and the actual accuracy requirements. It is usually set to 1 Hz, 2 Hz, 5 Hz or 10 Hz. A sampling frequency of not less than 1 Hz ensures that the number of data points collected in each charging cycle is sufficient to support the smooth calculation of the differential voltage curve and the accurate extraction of peak features. If the sampling frequency is below 1 Hz, the voltage change between adjacent sampling points may be too large, resulting in severe step-like distortion of the differential voltage curve, which affects the accuracy of subsequent inconsistency map construction and the extraction effect of interpretable physical feature channels. The Battery Management System (BMS) collects the terminal voltage of each individual battery cell using voltage sensors and the charging and discharging current of the battery pack using current sensors, according to a set sampling frequency. The collected data is stored in the BMS's internal memory in time-series format. During the data acquisition process, the BMS simultaneously records the corresponding timestamps and temperature data for subsequent capacity integration calculations and temperature compensation. After step S1, BMS continues to execute steps S2 to S5, including constructing an inconsistency map, extracting interpretable physical feature channels, obtaining deep features, multi-task prediction, and identifying and warning of weak points. The entire online estimation process is completed inside BMS without relying on external computing devices, realizing real-time health status estimation in an embedded environment. The BMS repeats the above steps during each charging process and stores the overall health status value, the degree of deviation of each individual unit, and the warning signal obtained each time in non-volatile memory. At the same time, it uploads them to the vehicle controller or cloud monitoring platform through the controller local area network bus or wireless communication module. Specifically, by executing the above method online in the BMS and setting the sampling frequency in step S1 to no less than 1 Hz, the real-time data acquisition and the calculation accuracy of the differential voltage curve are guaranteed, thus meeting the real-time requirements for online estimation of battery health status in vehicle and energy storage scenarios.

[0053] Although alternative embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make further changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0054] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this invention.

Claims

1. A method for online estimation of battery health status considering battery aging characteristics, characterized in that, Includes the following steps: S1, real-time acquisition of voltage and current data of each individual cell in the battery pack during the charging process, and calculation of the differential voltage curve of each individual cell within a preset voltage window. The differential voltage curve is the derivative curve of voltage with respect to capacity, and the capacity is obtained by time integration of the current data; S2, superimpose the differential voltage curves of all individual cells onto the same voltage coordinate system to construct an inconsistency map; The inconsistency map includes a superimposed distribution map of the differential voltage curves of each individual cell, and a point-by-point difference matrix of the differential voltage curve of each individual cell relative to the average differential voltage curve within the group. S3, input the inconsistency map into a pre-trained dual-task conditional generative adversarial network, the dual-task conditional generative adversarial network including a generator and a discriminator, the discriminator simultaneously outputting a authenticity score and an aging score, and extracting multiple interpretable physical feature channels from the intermediate convolutional layers of the discriminator, the interpretable physical feature channels including at least: The peak migration characteristic channel characterizing the peak voltage offset of the differential voltage curve, the peak attenuation characteristic channel characterizing the peak amplitude attenuation of the differential voltage curve, and the shoulder broadening characteristic channel characterizing the peak broadening of the differential voltage curve. S4. Input the differential voltage curves of each individual cell into the pre-trained convolutional neural network to obtain deep features. The multiple interpretable physical feature channels extracted from the discriminator in step S3 are concatenated with the deep features. The concatenated features are then input into a multi-task prediction network, which simultaneously outputs the overall health status value of the battery pack and the degree of deviation of each individual battery cell relative to the average aging trend. S5. Identify the weakest cell in the group based on the deviation degree of each individual battery cell output in step S4, and output a warning signal when the deviation degree exceeds a preset threshold.

2. The method according to claim 1, characterized in that, The construction of the inconsistency map in step S2 specifically includes: The differential voltage curve of each individual cell is smoothed using Gaussian filtering. Calculate the arithmetic mean of the differential voltage curves of all individual cells in the group to obtain the average differential voltage curve of the group. The difference between the differential voltage curve of each individual cell and the average differential voltage curve within the group is calculated at each sampling point in the voltage coordinate system to obtain a point-by-point difference matrix. The rows of the point-by-point difference matrix correspond to each individual cell, and the columns correspond to each voltage sampling point. Each value in the point-by-point difference matrix represents the differential voltage deviation value of the corresponding cell at the corresponding voltage sampling point. The superimposed distribution map and the point-by-point difference matrix are used together as the output of the inconsistency map.

3. The method according to claim 1, characterized in that, The training method for the dual-task conditional generative adversarial network is as follows: The generator takes a random noise vector and conditional information as input to generate a virtual differential voltage curve; The discriminator takes the inconsistency map as input, processes it through a convolutional layer, and outputs both an authenticity score and an aging score. During training, the generator and the discriminator are updated alternately, and the features extracted by the discriminator have both authenticity and aging correlation.

4. The method according to claim 1, characterized in that, The peak migration feature channel outputs a quantized value that characterizes the degree to which the peak voltage position of the differential voltage curve migrates towards lower voltage as the battery ages. The quantized value is positively correlated with the peak migration amount and is used to indicate the degree of loss of the positive electrode active material of the battery. The peak decay characteristic channel outputs a quantized value that characterizes the degree to which the peak amplitude of the differential voltage curve decreases with battery aging. The quantized value is positively correlated with the peak decay rate and is used to indicate the degree of loss of usable lithium ions in the battery. The shoulder broadening feature channel outputs a quantized value that characterizes the degree of peak broadening on both sides of the differential voltage curve. The quantized value is positively correlated with the degree of shoulder broadening of the curve and is used to indicate the degree of dispersion of the battery aging mode.

5. The method according to claim 1, characterized in that, The pre-trained convolutional neural network is pre-trained using the differential voltage curves of individual cells in the publicly available battery aging dataset as input. The pre-training task is multi-task learning, and it outputs the health status estimate and aging stage classification results of the individual cells. When migrating to the battery pack scenario, the weights of the first two convolutional layers of the pre-trained convolutional neural network are frozen, and the weights of the last two convolutional layers are fine-tuned only.

6. The method according to claim 1, characterized in that, In step S4, the multi-task prediction network is a three-layer fully connected network, whose input is a concatenation of a 512-dimensional convolutional neural network feature vector and a 16-dimensional inconsistent feature vector. The output layer of the multi-task prediction network includes one first neuron and M second neurons, where M is equal to the total number of individual cells in the battery pack. The first neuron outputs the overall health status value of the battery pack; Each second neuron uniquely corresponds to a single cell and is used to output the degree of deviation of the corresponding single cell from the average aging trend, wherein the degree of deviation is a normalized value between 0 and 1.

7. The method according to claim 1, characterized in that, The preset threshold in step S5 is set as follows: Based on the consistency level of the differential voltage curves of each individual cell in the battery pack at the time of manufacture, and using the deviation of the cell with the worst consistency as the basic threshold, when the real-time deviation of a certain cell exceeds 1.2 times the basic threshold, the cell is identified as a weak cell and a warning signal is output.

8. The method according to claim 1, characterized in that, The method also includes a closed-loop verification step: After each charge is completed, the newly acquired inconsistency map is compared with the offline calibration data, and the decay slope between the overall health status value obtained in this charge and the overall health status value obtained in the previous charge is calculated. If the decay slope deviates from the preset exponential decay model and exceeds the preset range, a model retraining instruction is triggered.

9. The method according to any one of claims 1 to 8, characterized in that, The preset voltage window is the complete range from the starting voltage to the charging cutoff voltage during the constant current charging stage, or a partial range between 2.8V and 3.6V during the constant current charging stage.

10. The method according to any one of claims 1 to 8, characterized in that, The method is executed online in the battery management system (BMS), and the sampling frequency for acquiring voltage and current data in step S1 is no less than 1Hz.

Images