Battery capacity evaluation method and device, computer device, and storage medium

Abstract

The application relates to the field of vehicle battery capacity and discloses a battery capacity evaluation method and device, a computer device and a storage medium, which comprise the following steps: obtaining original feature data of a target battery cell; performing feature screening processing on a rapid charging event based on the original feature data to obtain target feature data; wherein the target feature data comprises first charging features and second charging features derived from the same rapid charging event; predicting a battery capacity deviation of the target battery cell based on the target feature data by using a target prediction model to obtain a consistency evaluation result of the target battery cell; wherein the target prediction model is a double-tower model comprising a first encoder and a second encoder; the first encoder is used for processing the first charging features; and the second encoder is used for processing the second charging features. The method has the beneficial effect that the rapid charging event is screened, the interference of charging pile interaction noise on prediction is reduced from the data level, and the evaluation efficiency and coverage range are effectively improved.

Description

Technical Field

[0001] This application relates to the field of vehicle battery capacity technology, and in particular to a method, apparatus, computer equipment, and storage medium for evaluating battery capacity. Background Technology

[0002] With the rapid development of the new energy vehicle market, the capacity consistency between cells in a power battery system has a significant impact on the performance, driving range, and cycle life of the battery pack. The capacity deviation between cells (i.e., Ah deviation) is a core indicator for measuring battery consistency. Excessive capacity deviation can accelerate the overall aging and performance degradation of the battery pack, and even lead to safety risks.

[0003] In related technologies, the accurate calculation of cell-level capacity deviation mainly relies on the analysis of complete charging data curves under slow charging conditions. However, in practical applications, the acquisition cycle of slow charging data is long and the coverage is limited. In some cases, there is almost no slow charging condition for some vehicles, making it difficult to achieve efficient monitoring of vehicle battery consistency.

[0004] Therefore, there is a need for a battery capacity assessment method that can effectively predict cell-level capacity deviation (Ah deviation) with high accuracy. Summary of the Invention

[0005] This application aims to at least partially solve one of the technical problems in related technologies. To this end, this application proposes a method, apparatus, computer device, storage medium, and program product for evaluating battery capacity. The main technical solutions adopted in this application include: In a first aspect, this application provides a method for evaluating battery capacity. The method includes: acquiring raw feature data of a target battery cell; performing feature filtering processing on fast charging events based on the raw feature data to obtain target feature data; wherein the target feature data includes a first charging feature and a second charging feature from the same fast charging event; using a target prediction model to predict the battery capacity deviation of the target battery cell based on the target feature data to obtain a consistency evaluation result of the target battery cell; wherein the target prediction model is a dual-tower model including a first encoder and a second encoder; the first encoder is used to process the first charging feature; the second encoder is used to process the second charging feature; and the parameters of the first encoder and the second encoder are independent.

[0006] This method filters fast charging events, proactively eliminating noisy data to reduce the interference of charging pile interaction noise on predictions at the data level. Simultaneously, it employs a parameter-independent dual-tower coding structure to extract targeted features from two different physical meanings of feature data, fully utilizing the electrochemical information of fast charging events. This method does not rely on complete slow charging data; it can achieve high-precision prediction of cell capacity deviation using only fast charging data. It can support large-scale battery consistency monitoring of vehicles in operation, effectively improving evaluation efficiency and coverage.

[0007] Optionally, feature filtering processing of fast charging events is performed based on the original feature data to obtain target feature data, including: fast charging event filtering processing based on the original feature data to obtain intermediate feature data; and splitting and filtering processing of charging periods based on the intermediate feature data to obtain target feature data.

[0008] By employing a two-step feature filtering process involving event integrity and time segmentation, the system first ensures that the input event contains complete and valid information segments from the beginning and end. Then, it further eliminates invalid noise data in the middle segment. This process not only ensures the integrity of the valid information but also achieves efficient noise reduction and purification of fast charging data, providing a high-quality input foundation for the accurate prediction of the subsequent dual-tower model.

[0009] Optionally, the charging period includes a first period, a second period, and a third period; the charging period is split and filtered based on intermediate feature data to obtain target feature data, including: Based on the charging period, the intermediate feature data is split into features to obtain the first charging feature corresponding to the first period, the second charging feature corresponding to the second period, and the third charging feature corresponding to the third period. If the preset screening conditions are met in the third time period, the first charging feature and the second charging feature are integrated, and the integrated data is used as the target feature data.

[0010] By precisely dividing the charging period into intervals and actively discarding the high-noise third period features, the fluctuation noise caused by the current interaction between the charging pile and the BMS is eliminated from the data source. This avoids invalid noise interfering with the model's extraction of the intrinsic electrochemical features of the battery cell, enabling the model to focus on the first and last effective data segments with higher information density, effectively improving the accuracy of feature extraction and the reliability of capacity deviation prediction.

[0011] Optionally, the target prediction model is trained using a differential learning rate and gradient monitoring mechanism.

[0012] By training the dual-tower prediction model using a differentiated learning rate strategy combined with a gradient monitoring mechanism, the training pace of the dual towers is actively balanced at the optimizer level. This effectively solves the training imbalance problem caused by gradient asymmetry in the dual towers in fast-sufficient segment scenarios, ensuring full utilization of features from both the preceding and following segments. The resulting target prediction model has higher prediction accuracy and stronger generalization ability.

[0013] Optionally, the target prediction model further includes a prediction head network connected to the first encoder and the second encoder respectively; training is performed using a differentiated learning rate using the following method: obtaining a training sample set; wherein the training sample set includes a first charging feature sample, a second charging feature sample, and a capacity deviation label; inputting the first charging feature sample into the first encoder to obtain a first context feature vector; inputting the second charging feature sample into the second encoder to obtain a second context feature vector; performing dimensional concatenation and fusion on the first context feature vector and the second context feature vector, and inputting the fused feature vector into the prediction head network to obtain a test deviation prediction value; calculating the training loss based on the test deviation prediction value and the capacity deviation label; updating the network parameters of the first encoder, the second encoder, and the prediction head network based on the training loss; wherein the first encoder uses a first learning rate for parameter updates; the second encoder and the prediction head network use a second learning rate for parameter updates; and the first learning rate is less than the second learning rate.

[0014] By setting differentiated learning rates for the dual-tower encoders, the parameter update rhythm of the two towers was actively balanced, effectively alleviating the gradient competition and degradation problems of the later tower caused by the difference in signal strength between the front and rear sections. This enabled the two towers to fully learn the electrochemical characteristics of their respective charging stages, significantly improving the model's feature utilization rate and capacity deviation prediction accuracy for fast charging data.

[0015] Optionally, training can be performed using a gradient monitoring mechanism by: within a preset training step size period, calculating the first gradient norm corresponding to the first encoder and the second gradient norm corresponding to the second encoder at the current training step; calculating the ratio of the second gradient norm to the first gradient norm to obtain the gradient norm ratio; if the gradient norm ratio meets a preset adjustment condition, adjusting the first learning rate of the first encoder or the second learning rate of the second encoder.

[0016] By periodically monitoring the gradient norm ratio of the two towers and dynamically adjusting the learning rate, the gradient imbalance during training can be detected in real time, and the update rhythm of the two towers can be corrected in a timely manner. This prevents the latter tower from gradually degrading during long-term training, further ensuring the balanced learning of the dual-tower encoder and improving the stability of model training and the reliability of the final prediction effect.

[0017] Optionally, the original feature data of the target battery cell can be obtained, including: obtaining the time-series operation data of the target battery cell; and performing feature preprocessing based on the time-series operation data to obtain the original feature data.

[0018] By performing multi-dimensional feature preprocessing on the time-series operational data, richer cell electrochemical information can be extracted from the original sampled signals, improving the stability and accuracy of subsequent model predictions.

[0019] Secondly, this application provides a battery capacity evaluation device, the device comprising: The data acquisition module is used to acquire the raw characteristic data of the target battery cell; The feature filtering module is used to perform feature filtering processing on fast charging events based on the original feature data to obtain target feature data; wherein, the target feature data includes a first charging feature and a second charging feature from the same fast charging event; The capacity assessment module is used to predict the battery capacity deviation of the target cell based on the target feature data using the target prediction model, so as to obtain the consistency assessment result of the target cell; wherein, the target prediction model is a dual-tower model including a first encoder and a second encoder; the first encoder is used to process the first charging feature; the second encoder is used to process the second charging feature; and the parameters of the first encoder and the second encoder are independent.

[0020] Thirdly, this application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0022] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above methods. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 A flowchart of a battery capacity evaluation method according to an embodiment of this application; Figure 2 This is a flowchart of a method for determining target feature data according to an embodiment of this application; Figure 3a This is a schematic diagram of the change curve of the loss value according to one embodiment of this application; Figure 3b This is a schematic diagram of the variation curve of the mean absolute error according to an embodiment of this application; Figure 3c This is a schematic diagram of the learning rate change curve provided in another embodiment of this application; Figure 4 This is a structural block diagram of a battery capacity evaluation device according to an embodiment of this application; Figure 5 This is an internal structural diagram of a computer device provided according to an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be noted that the current industry technology for monitoring the capacity and health status of power batteries mainly relies on relatively stable slow charging data. This involves acquiring physical signals such as voltage, current, and temperature through a data acquisition circuit, and then performing preprocessing and feature engineering analysis to calculate the capacity or deviation.

[0027] Specifically, it can include the following categories: 1. Mechanism calculation method based on complete slow charging curve. This type of method is currently the mainstream basic scheme for capacity calculation, with the ampere-hour integral method as a typical example. However, as mentioned earlier, the limitation of this type of method is that it is highly dependent on complete slow charging data, while actual operating vehicles cannot support high-frequency online monitoring.

[0028] 2. Fast Charging Prediction Methods Based on Traditional Machine Learning. With the accumulation of battery data in the cloud, the industry has begun to explore using broader and more frequent fast charging data to predict capacity deviation. This type of method combines physical models with traditional machine learning: manually extracted statistical features such as charging time and average voltage from fast charging time-series data are then input into a regression model to predict capacity deviation. The limitation of this method lies in the limited dimensionality of manually designed features, making it difficult to fully capture the complex electrochemical kinetic information in the time-series data; furthermore, the drastic current fluctuations and low signal-to-noise ratio during fast charging make manually designed features susceptible to noise interference, thus limiting the final prediction accuracy.

[0029] 3. Fast charging prediction method based on single-tower temporal neural network. To improve the ability to extract temporal features, the industry has introduced deep learning technology and adopted temporal neural networks such as RNN, LSTM, and GRU. The entire fast charging time series data is directly input into a single network structure for end-to-end modeling. The network automatically extracts deep features, bypassing the complex derivation of battery mechanisms to predict capacity-related indicators.

[0030] However, this type of method still has shortcomings in fast charging scenarios: First, it suffers from severe noise interference. Not all fast charging data contains valid battery information. During the mid-charging phase, the charging current fluctuates dramatically due to the interaction strategy between the charging station and the Battery Management System (BMS). The voltage and current curves in this phase primarily reflect the control strategy characteristics of the charging station, rather than the electrochemical characteristics of the battery itself, resulting in an extremely low signal-to-noise ratio. Neural networks struggle to extract the true physical features reflecting the consistency of cell capacity from noisy data, easily leading to overfitting and memorizing data patterns rather than learning the battery's essential electrochemical characteristics.

[0031] Second, the generalization ability is insufficient. The fast charging control strategies of different car manufacturers and different battery packs are significantly different, resulting in no uniform pattern of current step states during fast charging and irregular fluctuations in voltage curves. This further increases the difficulty of extracting effective information, making it difficult for a single neural network model to adapt to fast charging conditions in multiple scenarios.

[0032] Third, the feature utilization rate is low. The fast charging process can be divided into multiple stages according to the State of Charge (SOC). The low SOC charging front stage and the high SOC charging back stage correspond to completely different electrochemical and physical processes. The front stage mainly involves the ohmic internal resistance response and polarization establishment process, while the back stage mainly involves the diffusion impedance change and the resting rebound process. The time-series neural network of the traditional single-tower structure cannot extract differentiated features from the two time-series data with completely different physical meanings, making it difficult to achieve the optimal representation of the two physical modes at the same time. This results in insufficient effective feature utilization of fast charging data and a bottleneck in prediction accuracy.

[0033] In summary, the relevant technologies are limited by the low frequency of slow charging data, making online monitoring impossible. Furthermore, in fast charging scenarios, they suffer from weak noise immunity, insufficient feature extraction targeting, and limited prediction accuracy, making it difficult to meet the high-precision evaluation requirements for cell capacity consistency.

[0034] Based on this, according to the embodiments of this application, an embodiment of a method for evaluating battery capacity is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment provides a method for evaluating battery capacity, such as... Figure 1 As shown, the method includes the following steps: S110. Obtain the original characteristic data of the target cell.

[0036] The target cell can refer to a single cell for which battery capacity evaluation is required. The raw feature data can refer to the structured feature sequence collected and processed in a single fast charging event, arranged in the order of time sampling. That is, it is a multi-dimensional time-series feature set constructed by preprocessing and feature engineering for a single fast charging event.

[0037] It is understandable that a fast charging event can refer to the complete process of charging the battery cell with a large current, that is, a single charging process covering four stages: pre-charging rest, constant current charging, reduced current charging, and post-charging rest.

[0038] For example, the raw feature data may include 16 time-series features, such as charging current features, average voltage features, cell voltage features, voltage difference features, single cell voltage change rate features, average voltage change rate features, average temperature features, state of charge (SOC) features, average voltage rolling window mean features, single cell voltage rolling window mean features, average voltage rolling difference features, single cell voltage rolling difference features, voltage cross rolling difference features, single cell capacity increment features, average capacity increment features, and capacity increment difference features.

[0039] Specifically, the original feature data of the target battery cell can be obtained by the following method: first, obtain the timing operation data of the target battery cell; then, perform feature preprocessing based on the timing operation data to obtain the original feature data.

[0040] The time-series operational data refers to the raw operational signal data sampled sequentially during fast charging. For example, it can be directly collected through the vehicle's BMS. After collection, the time-series operational data can be cleaned to remove null rows and abnormal transition points, ensuring data integrity and consistency.

[0041] After obtaining the time-series runtime data, feature engineering calculations can be performed on the time-series runtime data to obtain the original feature data.

[0042] Specifically, the charging current can be directly extracted from the original current sampling values ​​in the timing data to obtain the charging current characteristics.

[0043] For the average voltage, the average voltage sample value of the battery pack in the time-series running data can be directly extracted to obtain the average voltage characteristics.

[0044] For cell voltage, the target cell voltage sample value can be directly extracted from the timing data to obtain the cell voltage characteristics.

[0045] For voltage difference, the difference between cell voltage and average voltage in the timing data can be calculated to obtain voltage difference characteristics.

[0046] To determine the rate of change of individual unit voltage, the continuous sampled values ​​of individual unit voltage in the time-series operational data can be used to calculate the change of individual unit voltage over time within the moving time window, thereby obtaining the characteristics of the rate of change of individual unit voltage.

[0047] To determine the rate of change of average voltage, the continuous sampling values ​​of average voltage in the time-series running data can be used to calculate the change of average voltage over time within the moving time window, thereby obtaining the characteristic of the rate of change of average voltage.

[0048] For the average temperature, the average temperature feature can be obtained by directly extracting the average temperature sample value of the battery pack from the time-series running data.

[0049] For the state of charge (SOC), the SOC sample values ​​in the time-series running data can be directly extracted to obtain the state of charge characteristics.

[0050] For the average voltage rolling window mean, the average voltage sequence in the time series data can be used to calculate the average voltage of all sampling points within the moving time window, thereby obtaining the average voltage rolling window mean characteristic.

[0051] For the rolling window mean of individual unit voltage, the average value of individual unit voltage at all sampling points within the moving time window can be calculated using the individual unit voltage sequence in the time-series running data, thereby obtaining the characteristic of the rolling window mean of individual unit voltage.

[0052] To address the average voltage rolling difference, the difference between the average voltage of the first and last frames within the moving time window can be calculated using the average voltage sequence in the time-series running data, thereby obtaining the average voltage rolling difference characteristic.

[0053] To address the rolling difference in individual unit voltages, the individual unit voltage sequence in the time-series running data can be used to calculate the difference between the individual unit voltages in the first and last frames within the moving time window, thereby obtaining the characteristics of the rolling difference in individual unit voltages.

[0054] For voltage cross-rollover, the difference between the average voltage rollover and the individual voltage rollover can be calculated to obtain the voltage cross-rollover characteristics.

[0055] For the increase in single-cell capacity, the change in ampere-hour capacity corresponding to a unit voltage change, i.e., the single-cell dq / dv sequence, can be calculated using the single-cell voltage and charging current sequence in the time-series operation data, thereby obtaining the characteristics of the single-cell capacity increase.

[0056] For the average capacity increment, the average voltage and charging current sequences in the time-series operating data can be used to calculate the ampere-hour capacity change corresponding to a unit voltage change, i.e., the average dq / dv sequence, thereby obtaining the average capacity increment characteristics.

[0057] To address the capacity increment difference, the difference between the individual capacity increment and the average capacity increment can be calculated to obtain the capacity increment difference characteristic.

[0058] Furthermore, after obtaining all 16 time-series features, the Z-score standardization method can be used to standardize each feature.

[0059] For example, standardization can be performed using the following formula: In the formula, Z represents the standardized result, and X represents a certain feature; The mean value can be obtained by fitting the historical training set during model training. The standard deviation is represented by the value, which can be obtained by fitting the historical training set during model training.

[0060] After performing the Z-score standardization process on all 16 time-series features, the values ​​of each feature dimension are uniformly mapped to the standard normal distribution interval with a mean of 0 and a standard deviation of 1, so as to eliminate the differences in dimensionality and numerical scale between different physical quantities.

[0061] Subsequently, the standardized 16-dimensional features can be time-aligned and combined according to the original sampling time sequence to form a multi-dimensional time-series feature matrix that corresponds one-to-one with the original fast charging event time step. This matrix is ​​the original feature data after preprocessing, which can be directly used for subsequent charging event segmentation and model prediction input.

[0062] By performing multi-dimensional feature preprocessing on the time-series operational data, richer cell electrochemical information can be extracted from the original sampled signals, improving the stability and accuracy of subsequent model predictions.

[0063] S120. Perform feature filtering processing on the fast charging event based on the original feature data to obtain the target feature data.

[0064] As mentioned earlier, the middle segment of a fast charging event is affected by the current interaction strategy between the charging pile and the battery management system, resulting in drastic current fluctuations. The voltage and current curves in this segment reflect the characteristics of the charging pile's control strategy more than the electrochemical characteristics of the battery itself, and the signal-to-noise ratio is extremely low. In contrast, the physical information contained in the first and last segments of charging more accurately reflects the cell's capacity status. Therefore, it is necessary to segment and filter charging events to retain the effective characteristic segments.

[0065] Among them, target feature data can refer to the effective time series feature set that can be directly input into the target prediction model, that is, the data obtained by filtering from the original feature data, retaining only the data of specific time periods with high information density and low noise during fast charging.

[0066] Specifically, the target feature data may include a first charging feature and a second charging feature derived from the same fast charging event.

[0067] Among them, the first charging characteristic and the second charging characteristic can be the time sequence characteristics corresponding to two different physical periods in the fast charging event. They are independent of each other and can reflect the electrochemical characteristics of the battery cell at different charging stages.

[0068] For example, for a fast charging event from 15% SOC to 92%, the timing characteristics from the initial resting state to 30% SOC can be used as the first charging characteristic. This first charging characteristic corresponds to the low SOC high-current charging stage and includes physical signals highly correlated with cell capacity aging, such as open-circuit voltage, current step response, and ohmic internal resistance polarization establishment. The timing characteristics from 80% SOC to the end of charging and then resting can be used as the second charging characteristic. This second charging characteristic corresponds to the high SOC low-current charging stage and includes physical signals reflecting diffusion impedance and polarization degree, such as current-limiting voltage drop jump, polarization dissipation, and resting voltage rebound.

[0069] Specifically, the process begins by filtering fast charging events based on the original feature data to ensure completeness. Feature data corresponding to charging events that meet the start and end SOC requirements are retained as intermediate feature data. Then, charging periods are split and filtered based on this intermediate feature data. Three charging periods are divided according to a preset SOC threshold. Feature data from the middle period is then removed, retaining only the feature data from the first and last two valid periods. Finally, the target feature data is obtained by integrating these elements.

[0070] S130. Using the target prediction model, predict the battery capacity deviation of the target cell based on the target feature data to obtain the consistency evaluation result of the target cell.

[0071] The target prediction model can refer to a deep learning prediction model used to output the cell capacity deviation based on the fast charging timing characteristics.

[0072] Specifically, the target prediction model can be a dual-tower model containing a first encoder and a second encoder. The dual towers refer to the first encoder and the second encoder, also known as the front tower and the back tower. It can extract segmented time-series features through the dual-tower coding structure, and after fusion and mapping, obtain the capacity deviation prediction result, which serves as the consistency evaluation result of the target battery cell. Among them, the first encoder can refer to a time-series coding network in the dual-tower model, which is used to process the first charging features and can extract deep time-series features during the low SOC charging period.

[0073] The second encoder can refer to another temporal coding network in the dual-tower model, used to process the second charging features and capable of extracting deep temporal features during high SOC charging periods.

[0074] It should be noted that the parameters of the first encoder and the second encoder are independent, meaning that the network weight parameters of the two encoders are not shared and are updated and optimized independently. This design allows the two encoders to adapt to the electrochemical modes of different charging stages, learn the characteristic patterns of the corresponding time periods, and avoid the decrease in feature extraction accuracy caused by a single tower model simultaneously fitting two significantly different physical modes.

[0075] Furthermore, the target prediction model also includes a prediction head network connected to the first encoder and the second encoder, respectively.

[0076] The prediction head network can refer to a fully connected network that performs nonlinear mapping on the fused features output by the dual-tower encoder. It can be used to convert high-dimensional abstract features into continuous capacity bias values ​​and output the final prediction result.

[0077] Specifically, the first charging feature is first input into the first encoder, and after encoding, a first context prediction vector is output. Then, the second charging feature is input into the second encoder, and after encoding, a second context prediction vector is output. Next, the first and second context prediction vectors are concatenated and fused along the feature dimension to obtain a fused prediction vector.

[0078] Subsequently, the fused prediction vector is input into the prediction head network. After nonlinear mapping is performed by the fully connected layer of the prediction head network, the final output is the predicted value of the battery capacity deviation of the target cell, which serves as the consistency evaluation result of the target cell.

[0079] The consistency assessment result can be a quantitative evaluation value used to measure the deviation between the target cell capacity and the reference capacity, that is, the ampere-hour deviation value of the actual capacity of the target cell relative to the reference capacity. It can be directly used to judge the degree of cell capacity degradation and the consistency level within the battery pack. The reference capacity can be directly obtained by calibrating the target cell using the BMS system.

[0080] For example, taking a specific battery cell in a vehicle as an example, the BMS system calibrates the cell's baseline capacity as 100Ah. According to the aforementioned classification rules, the first charging feature of the current fast charging event corresponds to time-series data with a State of Charge (SOC) of 10% to 30%, and the second charging feature corresponds to time-series data with an SOC of 80% to 90%. The first charging feature is input into the first encoder, where the temporal dependencies are extracted using a bidirectional gated recurrent unit. Then, an additive attention mechanism is used to weight and aggregate the hidden states at each time step, resulting in a 128-dimensional first context prediction vector front=[f1,f2,…,fn]. The second charging feature is then input into the second encoder, where it is processed by an independent network with the same structure to obtain a 128-dimensional second context prediction vector back=[b1,b2,…,bn]. By directly concatenating two feature vectors, a 256-dimensional fused feature vector can be obtained (e.g., comb=[ f1,f2,…,fn,b1,b2,…,bn]). This vector is then input into the prediction head network. Assuming the final output capacity deviation prediction value is -2.8Ah, this result is the consistency assessment result of the target cell, indicating that the current actual capacity of the target cell is 2.8Ah lower than the reference capacity.

[0081] In the above implementation, fast charging events are screened to actively eliminate noisy data, reducing the interference of charging pile interaction noise on prediction at the data level. Simultaneously, a parameter-independent dual-tower coding structure is employed to extract targeted features from two different physical meanings of feature data, fully utilizing the electrochemical information of fast charging events. This method does not rely on complete slow charging data; it can achieve high-precision prediction of cell capacity deviation using only fast charging data. It can support large-scale battery consistency monitoring of vehicles in operation, effectively improving evaluation efficiency and coverage.

[0082] In some implementation methods, please refer to the appendix. Figure 2 Based on the original feature data, feature filtering processing of fast charging events is performed to obtain target feature data.

[0083] S210. Perform fast charging event filtering processing based on the original feature data to obtain intermediate feature data.

[0084] Specifically, the initial SOC and final SOC values ​​can be extracted from the original feature data corresponding to each fast charging event. The initial SOC value is compared with a preset starting threshold, and the final SOC value is compared with a preset ending threshold. The feature data corresponding to fast charging events that simultaneously satisfy the condition that the initial SOC is not greater than the preset starting threshold and the final SOC is not less than the preset ending threshold are retained and used as intermediate feature data.

[0085] For example, suppose the preset starting threshold is 20% and the preset ending threshold is 90%. Then, multiple collected fast charging events can be iterated through. If the starting SOC value of a fast charging event is 15% and the ending SOC value is 92%, satisfying both threshold requirements, then all the original feature data corresponding to that event is retained and included in the intermediate feature data. If the starting SOC value of a fast charging event is 35% and the ending SOC value is 70%, satisfying neither the starting nor ending requirements, then the original feature data corresponding to that event is removed and not included in subsequent processing. If the starting SOC value of a fast charging event is 12% and the ending SOC value is 75%, satisfying the starting requirement but not the ending requirement, then its corresponding original feature data is also removed.

[0086] S220. Based on the intermediate feature data, perform segmentation and filtering of charging periods to obtain target feature data.

[0087] The charging period can refer to the time segmentation obtained by dividing a single fast charging event according to the state of charge interval. Different time periods correspond to different electrochemical response characteristics and signal noise levels of the battery cell. For example, it can be divided into the pre-charging period, the middle charging period, and the post-charging period.

[0088] Specifically, the charging period can include a first period, a second period, and a third period.

[0089] The first period can refer to the low SOC charging period from the start of charging to the first threshold, i.e. the pre-charging period. The data in this period includes physical signals that are highly related to cell capacity aging, such as open circuit voltage, current step response, and ohmic internal resistance polarization establishment.

[0090] The second time period can refer to the high SOC charging period from the second threshold to the end of the charging stage, i.e. the later stage of charging. The data in this period includes physical signals that reflect the diffusion impedance and polarization degree, such as current limiting voltage drop jump, polarization dissipation, and static voltage rebound.

[0091] The third time period can refer to the mid-SOC charging period between the first and second boundary thresholds, i.e. the charging mid-term. The data in this period contains high-noise physical signals that are affected by the interaction strategy between the charging pile and the BMS.

[0092] For example, the first threshold can be set to 30%, and the second threshold can be set to 80%. In a single fast charging event, the time period from the initial static state before charging to 30% is the first period, the time period from the initial static state to the final static state after charging is the third period, and the time period from the initial static state to the final static state after charging is the second period.

[0093] Specifically, the target feature data can be obtained by the following method: First, the intermediate feature data is split based on the charging period to obtain the first charging feature corresponding to the first period, the second charging feature corresponding to the second period, and the third charging feature corresponding to the third period; if the third period meets the preset screening conditions, the first charging feature and the second charging feature are integrated, and the integrated data is used as the target feature data.

[0094] The third charging characteristic can refer to the set of time-series characteristics corresponding to the middle stage of charging. Since the third period is greatly affected by the current interaction strategy between the charging pile and the BMS, the current fluctuates frequently. The corresponding third charging characteristic mainly reflects the characteristics of the charging strategy rather than the electrochemical characteristics of the battery itself, and the signal-to-noise ratio is low.

[0095] Specifically, feature splitting can be performed as follows: First, determine the first and second boundary thresholds based on user needs or analysis precision. Then, iterate through each time sampling point of the intermediate feature data, read the SOC value corresponding to each sampling point, and classify it into the corresponding time period sampling set based on the SOC value of each sampling point. Finally, extract all feature dimension time-series data corresponding to the three sampling sets, and after processing, obtain the first charging feature of the first time period, the second charging feature of the second time period, and the third charging feature of the third time period.

[0096] By way of example, a first demarcation threshold is set to 30%, and a second demarcation threshold is set to 80%. For intermediate characteristic data of a fast charging event with a starting SOC of 15% and an ending SOC of 92%, SOC values are matched point by point in chronological order: all sampling points with SOC ≤ 30% are classified into a first time period set, and corresponding 16-dimensional time sequence features are extracted to form a first charging characteristic. All sampling points with 30% < SOC < 80% are classified into a third time period set, and corresponding 16-dimensional time sequence features are extracted to form a third charging characteristic. All sampling points with SOC ≥ 80% are classified into a second time period set, and corresponding 16-dimensional time sequence features are extracted to form a second charging characteristic. After splitting is completed, it is necessary to ensure that the three segments of features each retain the original chronological order.

[0097] Subsequently, when the third time period satisfies a preset screening condition, the first charging characteristic and the second charging characteristic are integrated, and the integrated data is used as target characteristic data.

[0098] Wherein, the preset screening condition is a determination rule for determining whether the third time period is a high-noise interval that needs to be eliminated.

[0099] Specifically, the target characteristic data can be determined by the following method: first, it is determined whether the third time period belongs to a preset discard time period, so as to determine whether the preset screening condition is satisfied. If the condition is satisfied, the third time period can be determined as an invalid noise segment, and all third charging characteristics are eliminated and not included in the model input. Subsequently, the first charging characteristic and the second charging characteristic are retained, which respectively maintain their respective time sampling order, and together form the target characteristic data as two independent time sequence characteristics.

[0100] It can be understood that, as described above, the third time period is severely affected by the current interaction strategy between the charging pile and the BMS, and the current fluctuates frequently, so the preset screening condition may be whether the third time period falls into a preset discard time period, and the time interval of the preset discard time period may be consistent with the middle charging section, so as to ensure that the unnecessary charging characteristics are indeed discarded by means of re-inspection.

[0101] By way of example, suppose the preset screening condition is that the third time period belongs to the preset discard time period of 30% to 80%. For the third charging characteristic obtained by the above splitting, the corresponding SOC interval is 30% to 80%, which is consistent with the interval range of the preset discard time period, so it is determined that the preset screening condition is satisfied, all characteristic data of this segment can be eliminated, and do not participate in subsequent model calculation.

[0102] Subsequently, the first charging characteristic (covering SOC 15% to 30%) and the second charging characteristic (covering SOC 80% to 92%) corresponding to the fast charging event are retained, and spliced according to the original time sequence order of the two segments of characteristics to form the final target characteristic data.

[0103] By precisely dividing the charging period into intervals and actively discarding the high-noise third period features, the fluctuation noise caused by the current interaction between the charging pile and the BMS is eliminated from the data source. This avoids invalid noise interfering with the model's extraction of the intrinsic electrochemical features of the battery cell, enabling the model to focus on the first and last effective data segments with higher information density, effectively improving the accuracy of feature extraction and the reliability of capacity deviation prediction.

[0104] In the above implementation, the two-step feature filtering process of event integrity and time segmentation first ensures that the input event contains complete and valid information segments from the source, and then further removes invalid noise data in the middle. This not only ensures the integrity of the valid information, but also achieves efficient noise reduction and purification of fast charging data, providing a high-quality input foundation for the accurate prediction of the subsequent dual-tower model.

[0105] In some implementations, the target prediction model is trained using a differential learning rate and gradient monitoring mechanism.

[0106] Specifically, an initial prediction model with a dual-tower structure can be constructed first, and then end-to-end training with multiple rounds of iterations can be performed on the initial prediction model. Each round of training outputs a corresponding intermediate prediction model. If the current intermediate prediction model meets the preset iteration termination condition, then the intermediate prediction model is used as the final target prediction model. If the current intermediate prediction model does not meet the preset iteration termination condition, then the intermediate prediction model is used as a new initial prediction model, and the training steps of differential learning rate combined with gradient monitoring mechanism are repeated to obtain a new intermediate prediction model. The iteration stops when the new intermediate prediction model meets the preset iteration termination condition, and then it is used as the target prediction model.

[0107] It should be noted that the entire training process adopts an end-to-end single-stage training approach, without phased pre-training or fine-tuning. All network parameters are updated and optimized synchronously within the same training process.

[0108] Optionally, training can be performed using a differentiated learning rate as follows: First, a training sample set is obtained; then, the first charging feature sample is input into the first encoder to obtain the first context feature vector; next, the second charging feature sample is input into the second encoder to obtain the second context feature vector; then, the first context feature vector and the second context feature vector are dimensionally concatenated and fused, and the fused feature vector is input into the prediction head network to obtain the test bias prediction value; then, the training loss is calculated based on the test bias prediction value and the capacity bias label; finally, the network parameters of the first encoder, the second encoder, and the prediction head network are updated based on the training loss.

[0109] The training sample set can be a labeled dataset used to train the dual-tower prediction model. It can consist of real data from multiple historical fast charging events, along with the real labeled values ​​of these real data.

[0110] Specifically, the training sample set includes a first charging feature sample, a second charging feature sample, and a capacity deviation label.

[0111] The first charging feature sample can refer to the time-series feature data corresponding to the low SOC charging stage in the training samples, the second charging feature sample can refer to the time-series feature data corresponding to the high SOC charging stage in the training samples, and the capacity deviation label can refer to the true labeled value of the corresponding cell capacity deviation, which can be used as the benchmark true value for model fitting target.

[0112] For example, assuming historical fast charging data of a certain battery cell is selected as the sample source, each sample corresponds to a complete fast charging event. First, the time-series data of the preceding and following segments are obtained according to a preset threshold. Then, after feature engineering and standardization, the first charging feature sample and the second charging feature sample are generated respectively. At the same time, the actual capacity deviation value of the battery cell is obtained by calibration through the slow charging ampere-hour integration method or the BMS system, which serves as the corresponding capacity deviation label. The three together form a complete training sample.

[0113] After obtaining the training sample set, the first charging feature sample can be input into the first encoder to obtain the first context feature vector, and the second charging feature sample can be input into the second encoder to obtain the second context feature vector.

[0114] The first context feature vector can refer to the feature vector output by the first encoder after performing time-series encoding on the first charging feature sample. Correspondingly, the second context feature vector can refer to the feature vector output by the second encoder after performing time-series encoding on the second charging feature sample.

[0115] It should be noted that the first encoder and the second encoder use the exact same network structure, but their parameters are independent of each other. Both consist of two parts: a bidirectional gated recurrent unit (BiGRU) and an additive attention mechanism.

[0116] Before inputting the data to the encoder, variable-length sequence packing can be performed. Since the temporal lengths of different samples vary, dynamic padding can be applied to the sequences within a batch, aligning all samples to the longest sequence length in the current batch. Then, the padded sequences are packed using `pack_padded_sequence`, ensuring that subsequent BiGRU layers only perform loop calculations on valid time steps, avoiding the introduction of invalid information from padded frames.

[0117] Subsequently, the packaged sequence is input into a multi-layer BiGRU network, enabling the network to extract temporal dependencies from both the forward and reverse directions simultaneously, and output the hidden state sequence corresponding to each time step.

[0118] It's important to note that if the BiGRU has more than one layer, a Dropout layer can be added between adjacent network layers to randomly deactivate some neurons and suppress overfitting. After computation, the output is restored to the padded tensor format using pad_packed_sequence.

[0119] Next, additive attention computation is performed on the BiGRU output. First, the hidden state at each time step is mapped through the first linear transformation layer, processed by the Tanh activation function and Dropout, and then mapped to a scalar attention score through the second linear transformation layer. The masked_fill operation sets the attention score corresponding to the filled position to negative infinity, and after Softmax normalization, ensures that the weight at the filled position is 0. Finally, the attention weights are weighted and summed with the hidden state at the corresponding time step to obtain a fixed-dimensional context feature vector, which serves as the final output of the corresponding encoder.

[0120] Subsequently, the first and second context feature vectors can be fused by dimensional concatenation. The two vectors can be directly concatenated end-to-end along the feature dimension, fully preserving all the information from each feature segment. The fused feature vector is then input into the prediction head network, enabling the initial prediction model to perform capacity bias prediction and obtain the test bias prediction value.

[0121] The test bias prediction value can refer to the capacity bias estimate output by the model during the training phase.

[0122] Specifically, the prediction head network can be a multilayer fully connected neural network (MLP). After the feature vector is fused into the prediction head network, it will pass through the fully connected layer and the nonlinear transformation of the activation function in sequence, and finally map out a single continuous value as the test bias prediction value.

[0123] The training loss can then be calculated based on the predicted test bias and the capacity bias label.

[0124] It should be noted that the training process can use the Huber loss function to calculate the training loss. This loss function combines the advantages of mean squared error and mean absolute error, and is more robust to outliers. Its specific expression is as follows: In the formula, Represents capacity deviation label; This represents the predicted value of the test deviation; This represents the hyperparameters used during model training and can be set to 1.

[0125] From the above formula, it can be seen that when the absolute value of the prediction error does not exceed... When the absolute value of the prediction error exceeds a certain threshold, the loss can be calculated to ensure convergence accuracy. In this case, linear error is used to calculate the loss, reducing the interference of abnormal samples on training.

[0126] After obtaining the training loss, the network parameters of the first encoder, the second encoder, and the prediction head network can be updated based on the training loss.

[0127] In this system, the first encoder uses a first learning rate to update its parameters, while the second encoder and the prediction head network use a second learning rate to update their parameters, and the first learning rate is less than the second learning rate.

[0128] Understandably, this setup is designed to address the gradient competition issue that may arise during dual-tower training. Specifically, the first charging feature sample has a higher signal strength and a larger gradient magnitude. If both towers use the same learning rate, the first encoder's update speed will be significantly faster than the second encoder, resulting in a "Matthew effect" where the first tower dominates training while the second tower's representation degrades. Therefore, it's necessary to lower the learning rate of the first encoder to actively balance the update speeds of the two towers at the optimizer level, ensuring that the second encoder can effectively learn the feature patterns of the corresponding time period.

[0129] Specifically, a baseline learning rate can first be determined during model training, and 0.3 to 0.5 times the baseline learning rate can be set as the first learning rate, which is then allocated to the first encoder for parameter updates. The baseline learning rate can then be directly used as the second learning rate, allocated to the second encoder and the prediction head network for parameter updates.

[0130] For example, the learning rate scheduling can adopt a cosine annealing warm restart scheduler. The initial baseline learning rate can be set to 0.001. During training, the learning rate is periodically adjusted according to the cosine law, and the warm restart mechanism helps the model escape local optima and improve the final generalization performance.

[0131] When updating parameters, backpropagation can be performed based on the training loss to calculate the parameter gradients of the three network modules respectively, and then gradient descent can be performed synchronously according to their respective learning rates to complete one round of parameter update.

[0132] By setting differentiated learning rates for the dual-tower encoders, the parameter update rhythm of the two towers was actively balanced, effectively alleviating the gradient competition and degradation problems of the later tower caused by the difference in signal strength between the front and rear sections. This enabled the two towers to fully learn the electrochemical characteristics of their respective charging stages, significantly improving the model's feature utilization rate and capacity deviation prediction accuracy for fast charging data.

[0133] Furthermore, training can also be performed using a gradient monitoring mechanism in the following way: First, within a preset training step size period, calculate the first gradient norm corresponding to the first encoder and the second gradient norm corresponding to the second encoder at the current training step; then calculate the ratio of the second gradient norm to the first gradient norm to obtain the gradient norm ratio; if the gradient norm ratio meets the preset adjustment conditions, then adjust the first learning rate of the first encoder or the second learning rate of the second encoder.

[0134] The training step size period can refer to the number of rounds in a single model training iteration, which includes the number of training steps. For example, every 100 training steps can be considered as one training round. The current training step can refer to the single iteration step in the current training process where parameter updates are being performed.

[0135] The first gradient norm can refer to the L2 norm of the gradients of all trainable parameters of the first encoder, which can quantify the overall gradient strength of the first encoder. The second gradient norm can refer to the L2 norm of the gradients of all trainable parameters of the second encoder, which can quantify the overall gradient strength of the second encoder.

[0136] Specifically, at each training step period, the gradient tensors of all trainable parameters of the first encoder under the current training step can be extracted first, and their L2 norm can be calculated to obtain the first gradient norm. Similarly, the gradient tensors of all trainable parameters of the second encoder are extracted, and their L2 norm is calculated to obtain the second gradient norm.

[0137] The ratio of the second gradient norm to the first gradient norm can then be calculated to obtain the gradient norm ratio. This gradient norm ratio is defined as the relative ratio of the gradient strength of the second encoder to the gradient strength of the first encoder, and can quantitatively measure the gradient balance state between the two towers.

[0138] Specifically, the gradient norm ratio can be calculated using the following formula: In the formula, Represents the gradient norm ratio; Represents the second gradient norm; This represents the first gradient norm.

[0139] Furthermore, if the gradient norm ratio meets the preset adjustment conditions, the first learning rate of the first encoder or the second learning rate of the second encoder is adjusted.

[0140] The preset adjustment condition can be set to the gradient norm ratio being consistently lower than a preset threshold. For example, if the gradient norm ratio is less than 0.2 for three consecutive training step cycles, it means that the gradient strength of the second encoder may be weaker than that of the first encoder, indicating an imbalance in the dual-tower training.

[0141] Specifically, the gradient norm ratio can be statistically analyzed in real time for multiple consecutive monitoring periods. If the value remains below a preset threshold, it is determined that gradient imbalance has occurred in the dual-tower training. At this point, the first learning rate of the first encoder can be further reduced, or the second learning rate of the second encoder can be appropriately increased to widen the learning rate gap between the two, thereby enhancing the update incentive for the second encoder until the gradient norm ratio returns to the preset balance range, maintaining the balanced progress of dual-tower training.

[0142] By periodically monitoring the gradient norm ratio of the two towers and dynamically adjusting the learning rate, the gradient imbalance during training can be detected in real time, and the update rhythm of the two towers can be corrected in a timely manner. This prevents the latter tower from gradually degrading during long-term training, further ensuring the balanced learning of the dual-tower encoder and improving the stability of model training and the reliability of the final prediction effect.

[0143] For example, the overall architecture and training hyperparameters of the target prediction model can be set as follows: the input feature dimension is 16, corresponding to 16-dimensional temporal features. The BiGRU hidden layer dimension is 128, and the network has 3 layers. The dropout ratio is set to 0.3. A bidirectional GRU structure is also used. The hidden dimension of the fully connected layer in the prediction head is 64, and the intermediate dimension of the attention mechanism is 64. The activation function is GELU, and the initial baseline learning rate is 1e-3.

[0144] Further, please refer to Figure 3a The graph shows the change curve of the loss value during a certain training process. The horizontal axis represents the training rounds, and the vertical axis represents the Huber loss value. The curve includes the training set loss (blue line) and the validation set loss (red line), which can intuitively reflect the convergence process and fitting status of the model.

[0145] Please refer to Figure 3b The graph shows the change in mean absolute error (MAE) during a certain training process. The horizontal axis represents the number of training rounds, and the vertical axis represents the MAE value. It also includes two curves for the training set (blue line) and the validation set (red line), which are used to measure the change in the accuracy of the model's capacity bias prediction.

[0146] Please refer to Figure 3cThe diagram shows the learning rate change curves of the two towers during a certain training process. The horizontal axis represents the training rounds, and the vertical axis represents the learning rate value. The blue line represents the change of the second learning rate, and the red line represents the change of the first learning rate. This reflects the change pattern of the differentiated learning rate under cosine annealing scheduling, and allows for a direct observation of the difference in the learning rates of the two towers and the trend of synchronous scheduling.

[0147] In the above implementation, the dual-tower prediction model is trained by combining a differentiated learning rate strategy with a gradient monitoring mechanism. This actively balances the training rhythm of the dual towers at the optimizer level, effectively solving the training imbalance problem caused by gradient asymmetry in the dual towers in the fast-sufficient segment scenario. It ensures the full utilization of features in the preceding and following segments, and the final target prediction model has higher prediction accuracy and stronger generalization ability.

[0148] In some implementations, after using a target prediction model to predict the battery capacity deviation of the target cell based on target feature data to obtain the consistency evaluation result of the target cell, the method further includes: The contributions of the first encoder and the second encoder in the target prediction model are verified by ablation experiments to obtain the first contribution rate corresponding to the first encoder and the second contribution rate corresponding to the second encoder; the first learning rate is adjusted by the first contribution rate and the second learning rate is adjusted by the second contribution rate.

[0149] Ablation experiment verification can refer to a verification method that quantifies the contribution of a specific module in a directional zeroing model to the overall predictive performance of the model. For example, by comparing the changes in prediction error before and after module failure, the contribution ratio of the corresponding module to the final prediction result can be calculated, which can evaluate the actual role and value of the module.

[0150] The first contribution rate can refer to a numerical indicator that quantifies the contribution of the first encoder to the capacity deviation prediction performance, while the second contribution rate can refer to a numerical indicator that quantifies the contribution of the second encoder to the capacity deviation prediction performance.

[0151] Specifically, ablation experiment validation and learning rate adjustment can be performed through the following steps: First, the full target prediction model can be run on a unified validation dataset, and the mean absolute error (MAE) between the capacity bias prediction values ​​and capacity bias labels of all validation samples can be calculated. This error is recorded as the full model error. .

[0152] Next, all the first context feature vectors output by the first encoder can be set to zero, and only the second context feature vectors output by the second encoder can be retained and input into the prediction head network. The mean absolute error of the prediction results is then recalculated on the same validation dataset and denoted as the second encoder error. The second encoder error is then substituted into the following formula to obtain the second contribution rate: In the formula, Represents the second contribution rate. Represents the error of the complete model; This represents the error of the second encoder.

[0153] Similarly, the second context feature vector output by the second encoder can be set to zero, and only the first context feature vector output by the first encoder can be retained and input into the prediction head network. The mean absolute error of the prediction results can then be recalculated on the same validation dataset and denoted as the first encoder error. The first encoder error is then substituted into the following formula to obtain the first contribution rate: In the formula, Represents the highest contribution rate. Represents the error of the complete model; This represents the error of the first encoder.

[0154] Then, based on the calculated first contribution rate and second contribution rate, the first learning rate and second learning rate can be adjusted accordingly.

[0155] Specifically, if the first contribution rate is significantly higher than the second contribution rate, it indicates that the first encoder is too dominant in prediction, meaning the feature value of the second encoder is not being fully utilized. In this case, the first learning rate can be further reduced or the second learning rate can be appropriately increased to guide the model to strengthen its learning and utilization of the second encoder's features. If the second contribution rate is significantly higher than the first contribution rate, the first learning rate can be increased or the second learning rate decreased accordingly to maintain the contributions of the two towers within a balanced and reasonable range, avoiding the marginalization of features from a single tower.

[0156] In the above implementation, the actual performance contribution of the dual-tower encoder is quantified through ablation experiments. This provides reverse guidance for the dynamic adjustment of the learning rate from the perspective of prediction performance, avoiding the waste of feature information caused by single-tower dominance and improving the targeting and rationality of learning rate adjustment. Simultaneously, the quantified contribution rate directly reflects the value ratio of the charging data before and after the two stages, enhancing the interpretability of the model structure and providing a clear quantitative basis for subsequent model optimization and parameter tuning.

[0157] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple steps or stages, which are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0158] This specification also provides a battery capacity evaluation device 400, such as... Figure 4 As shown, it includes: a data acquisition module 410, a feature filtering module 420, and a capacity assessment module 430, wherein: The data acquisition module 410 is used to acquire the original characteristic data of the target battery cell.

[0159] The feature filtering module 420 is used to perform feature filtering processing on fast charging events based on the original feature data to obtain target feature data; wherein, the target feature data includes a first charging feature and a second charging feature from the same fast charging event.

[0160] The capacity assessment module 430 is used to predict the battery capacity deviation of the target cell based on the target feature data using the target prediction model, so as to obtain the consistency assessment result of the target cell; wherein, the target prediction model is a dual-tower model including a first encoder and a second encoder; the first encoder is used to process the first charging feature; the second encoder is used to process the second charging feature; and the parameters of the first encoder and the second encoder are independent.

[0161] In some implementations, the feature filtering module 420 is further configured to perform feature filtering processing on fast charging events based on the original feature data to obtain target feature data, including: performing fast charging event filtering processing on the original feature data to obtain intermediate feature data; and performing charging period segmentation filtering processing on the intermediate feature data to obtain target feature data.

[0162] In some implementations, the charging period includes a first period, a second period, and a third period; the feature filtering module 420 is further configured to perform feature splitting processing on intermediate feature data based on the charging period to obtain a first charging feature corresponding to the first period, a second charging feature corresponding to the second period, and a third charging feature corresponding to the third period; if the third period meets the preset filtering conditions, the first charging feature and the second charging feature are integrated, and the integrated data is used as the target feature data.

[0163] In some embodiments, a battery capacity assessment device 400 further includes a model training module for training a target prediction model using a differential learning rate and gradient monitoring mechanism.

[0164] In some implementations, the target prediction model further includes a prediction head network connected to the first encoder and the second encoder, respectively. The model training module is further configured to train the model using a differentiated learning rate through the following methods: obtaining a training sample set; wherein the training sample set includes a first charging feature sample, a second charging feature sample, and a capacity deviation label; inputting the first charging feature sample into the first encoder to obtain a first context feature vector; inputting the second charging feature sample into the second encoder to obtain a second context feature vector; performing dimensional concatenation and fusion on the first and second context feature vectors, and inputting the fused feature vector into the prediction head network to obtain a test deviation prediction value; calculating a training loss based on the test deviation prediction value and the capacity deviation label; updating the network parameters of the first encoder, the second encoder, and the prediction head network based on the training loss; wherein the first encoder uses a first learning rate for parameter updates; the second encoder and the prediction head network use a second learning rate for parameter updates; and the first learning rate is less than the second learning rate.

[0165] In some implementations, the model training module is also used to train using a gradient monitoring mechanism by means of the following method: within a preset training step period, calculating the first gradient norm corresponding to the first encoder and the second gradient norm corresponding to the second encoder at the current training step; calculating the ratio of the second gradient norm to the first gradient norm to obtain the gradient norm ratio; if the gradient norm ratio meets a preset adjustment condition, adjusting the first learning rate of the first encoder or the second learning rate of the second encoder.

[0166] For specific limitations regarding a battery capacity evaluation device, please refer to the limitations of a battery capacity evaluation method described above, which will not be repeated here. Each module in the aforementioned battery capacity evaluation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0167] In this embodiment, a battery capacity evaluation device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0168] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 5 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations. Figure 5 Take a processor 10 as an example.

[0169] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0170] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0171] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0172] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0173] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0174] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0175] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0176] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

[0177] The apparatus, module, or unit described in the above embodiments can be implemented by a computer chip or entity, or by a product having a certain function. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0178] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0179] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0180] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0181] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0182] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0183] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0184] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0185] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for evaluating battery capacity, characterized in that, The method includes: Obtain the raw feature data of the target battery cell; Based on the original feature data, feature filtering processing of fast charging events is performed to obtain target feature data; wherein, the target feature data includes a first charging feature and a second charging feature from the same fast charging event; Using a target prediction model, the battery capacity deviation of the target cell is predicted based on the target feature data to obtain the consistency evaluation result of the target cell; wherein, the target prediction model is a dual-tower model including a first encoder and a second encoder; the first encoder is used to process the first charging feature; the second encoder is used to process the second charging feature; and the parameters of the first encoder and the second encoder are independent.

2. The method according to claim 1, characterized in that, The feature filtering process for fast charging events based on the original feature data to obtain target feature data includes: Based on the original feature data, fast charging event filtering processing is performed to obtain intermediate feature data; The intermediate feature data is used to perform a segmentation and filtering process for charging periods to obtain the target feature data.

3. The method according to claim 2, characterized in that, The charging period includes a first period, a second period, and a third period; the process of splitting and filtering the charging period based on the intermediate feature data to obtain the target feature data includes: Based on the charging period, the intermediate feature data is split into features to obtain a first charging feature corresponding to the first period, a second charging feature corresponding to the second period, and a third charging feature corresponding to the third period. If the preset screening conditions are met in the third time period, the first charging feature and the second charging feature are integrated, and the integrated data is used as the target feature data.

4. The method according to claim 1, characterized in that, The target prediction model is trained using a differential learning rate and gradient monitoring mechanism.

5. The method according to claim 4, characterized in that, The target prediction model further includes a prediction head network connected to the first encoder and the second encoder respectively; it is trained using a differential learning rate through the following method: Obtain a training sample set; wherein the training sample set includes a first charging feature sample, a second charging feature sample, and a capacity deviation label; The first charging feature sample is input into the first encoder to obtain the first context feature vector; The second charging feature sample is input into the second encoder to obtain the second context feature vector; The first context feature vector and the second context feature vector are dimensionally concatenated and fused, and the fused feature vector is input into the prediction head network to obtain the test bias prediction value. The training loss is calculated based on the predicted test deviation value and the capacity deviation label. The network parameters of the first encoder, the second encoder, and the prediction head network are updated based on the training loss; wherein the first encoder uses a first learning rate for parameter update; the second encoder and the prediction head network use a second learning rate for parameter update; and the first learning rate is less than the second learning rate.

6. The method according to claim 4, characterized in that, Training can be performed using a gradient monitoring mechanism through the following method: Within a preset training step period, calculate the first gradient norm of the first encoder and the second gradient norm of the second encoder for the current training step. Calculate the ratio of the second gradient norm to the first gradient norm to obtain the gradient norm ratio; If the gradient norm ratio meets the preset adjustment conditions, then the first learning rate of the first encoder or the second learning rate of the second encoder is adjusted.

7. The method according to claim 1, characterized in that, The acquisition of the original feature data of the target battery cell includes: Obtain the timing operation data of the target battery cell; Feature preprocessing is performed based on the time-series running data to obtain the original feature data.

8. A battery capacity evaluation device, characterized in that, The device includes: The data acquisition module is used to acquire the raw characteristic data of the target battery cell; The feature filtering module is used to perform feature filtering processing on fast charging events based on the original feature data to obtain target feature data; wherein, the target feature data includes a first charging feature and a second charging feature from the same fast charging event; A capacity assessment module is used to predict the battery capacity deviation of the target cell based on the target feature data using a target prediction model, so as to obtain the consistency assessment result of the target cell; wherein, the target prediction model is a dual-tower model including a first encoder and a second encoder; the first encoder is used to process the first charging feature; the second encoder is used to process the second charging feature; and the parameters of the first encoder and the second encoder are independent.

9. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

Images