Methods, devices, and terminals for assessing battery health using incomplete charging data

Abstract

This invention relates to the field of battery management technology, specifically to a method, apparatus, and terminal for assessing battery health using incomplete charging data. The method involves acquiring charging data, identifying constant current intervals, filtering effective voltage segments, and configuring effective segment identifiers. Features are extracted through resampling to construct a segment fusion health factor and data quality coefficient. The fusion weight is obtained by combining temperature and current changes to calculate operating condition reliability parameters and correcting sensitivity. A coarse estimate and aggregate reliability are calculated. Based on the coarse estimate, the discrimination distance is calculated and smoothing parameters are adjusted to output the soft membership degree of the degradation stage. The masked feature vector is input into the expert model for each degradation stage and weighted using the soft membership degree to output the health assessment result. This invention does not rely on a complete charging curve. The masking mechanism adaptively handles variable-length segments and effectively blocks sensor noise. The soft weighting completely eliminates output jumps at stage boundaries, ensuring continuous and smooth control of battery charging and discharging power.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and more specifically, to a method, apparatus, and terminal for assessing battery health using incomplete charging data. Background Technology

[0002] Lithium-ion batteries are widely used in new energy vehicles, energy storage systems, and portable electronic devices due to their high energy density, long cycle life, and fast power response. As the number of battery cycles increases, aging phenomena occur inside the battery, such as lithium inventory loss, active material loss, enhanced polarization, and increased internal resistance, leading to a gradual decrease in capacity. State of Health (SOH) is typically represented by the ratio of current usable capacity to initial or rated capacity. Accurately estimating SOH is crucial for safe battery operation, lifespan prediction, and maintenance decisions.

[0003] Many existing SOH estimation methods heavily rely on complete charge-discharge curves. However, in real-world vehicle operation or energy storage scheduling, charging behavior is typically highly random, often only obtaining partial curves or local voltage segments. Directly applying models trained on complete laboratory curves to local charging scenarios can easily lead to severe biases in the distribution of training and testing data, significantly reducing prediction accuracy and engineering applicability. Furthermore, existing methods typically input all candidate features directly into the model without fully considering the varying sensitivity of different voltage segments to SOH, and lack dynamic penalty mechanisms for alternating thermal conditions and underlying sensor noise. This makes the evaluation network highly susceptible to distortion due to feature redundancy and pseudo-polarization features.

[0004] Furthermore, the battery capacity degradation process exhibits significant stage heterogeneity, typically characterized by early-stage stable degradation, mid-stage nonlinear acceleration, and differentiated evolutionary features such as capacity fluctuations in the later stages. A single global model struggles to simultaneously adapt to the changing patterns of different degradation stages. Using a simple multi-model hard threshold switching approach inevitably leads to a step jump in the SOH (State of Health) assessment curve at the stage boundaries. This discontinuity in the underlying data output directly causes oscillations in the battery management system's charging and discharging power limiting commands, seriously threatening physical operational safety. Therefore, a method, device, and terminal for assessing battery health using incomplete charging data is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a method, apparatus, and terminal for assessing battery health using incomplete charging data, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention aims to provide a method for assessing battery health using incomplete charging data, comprising the following steps: S1. Obtain incomplete charging data of the battery to be evaluated, identify the constant current charging range and divide it into multiple candidate local voltage segments, select the effective local voltage segments and configure the effective segment identifier. The incomplete charging data includes at least battery voltage, charging current, cumulative capacity, and battery temperature. S2. Resample the effective local voltage segment to extract initial health features, construct a segment fusion health factor based on the feature evaluation index of the initial health features, and extract the segment data quality coefficient based on the resampled data state. S3. Calculate the static segment degradation sensitivity based on the feature evaluation index; and calculate the segment operating condition reliability parameter by combining the segment data quality coefficient and the change state of charging current and battery temperature corresponding to the effective local voltage segment. S4. Use the segment condition confidence parameter to correct the static segment degradation sensitivity to obtain the segment fusion weight; use the segment fusion weight, segment valid identifier and segment condition confidence parameter as mask parameters, and concatenate them with the segment fusion health factor to form a mask feature vector, and input it into the evaluation network to obtain a rough estimate of health status and aggregated condition confidence. S5. Calculate the discrimination distance to each degradation stage based on the rough estimate of the health status, and adjust the preset smoothing parameter using the credibility of the aggregated working condition, thereby calculating the soft membership degree of the degradation stage. S6. Input the mask feature vector into the expert fusion model of each stage respectively, and use the soft membership degree of the degradation stage to weight and sum the outputs of each model to obtain the health status assessment result.

[0007] As a further improvement to this technical solution, in step S1, determining the effective local voltage segment from the candidate local voltage segments involves the following steps: Based on the fluctuation of the charging current within a preset time window, candidate constant current charging intervals are determined from the incomplete charging data, and it is determined whether the voltage within the candidate constant current charging interval continues to rise. The candidate constant current charging interval is defined as the constant current charging interval if the voltage continues to rise and the duration and number of effective sampling points meet the preset requirements. Determine whether the constant current charging interval simultaneously covers the lower voltage boundary and the upper voltage boundary of each candidate local voltage segment. The candidate local voltage segment that simultaneously covers the lower voltage boundary and the upper voltage boundary of the segment and whose number of effective sampling points within the segment meets the preset requirements is determined as an effective local voltage segment. For candidate local voltage segments that do not simultaneously cover both the lower and upper voltage boundaries of the segment, they are not completed by extrapolating data outside the segment boundaries.

[0008] As a further improvement to this technical solution, in step S2, the segment fusion health factor corresponding to each effective local voltage segment is obtained, and the specific steps involved are as follows: The effective voltage segment is resampled by voltage domain interpolation according to a preset voltage difference step size, and the discrete time domain sampling sequence is converted into a resampled voltage sequence of fixed length, as well as a resampled current sequence, a resampled capacity sequence and a resampled time sequence aligned with the resampled voltage sequence. Within each effective voltage segment, based on the resampled time series and the resampled capacity series, multiple initial health features characterizing local charging capacity changes, charging time required per unit voltage change, and capacity-voltage curve shape are extracted. The feature evaluation index, which is pre-calibrated based on a historical battery aging dataset, is invoked. The feature evaluation index includes the Pearson correlation coefficient between each initial health feature and the actual health state of the battery, and the Spearman rank correlation coefficient of the initial health feature with the battery aging cycle. The feature fusion weights corresponding to each initial health feature are determined based on the Pearson correlation coefficient and the Spearman rank correlation coefficient. Multiple initial health features within the same effective local voltage segment are then weighted and fused according to the feature fusion weights to obtain the segment fusion health factor. The number of original valid sampling points for each effective local voltage segment before interpolation resampling is obtained, and the sampling point density coefficient is determined based on the number of original valid sampling points and the number of resampling target points, wherein the value of the sampling point density coefficient is not greater than 1. The interpolation interval penalty coefficient is determined based on the degree of deviation between the voltage interval between adjacent original voltage sampling points and the preset voltage difference step size; The fragment data quality coefficient is determined based on the sampling point density coefficient and the interpolation interval penalty coefficient. The larger the number of original valid sampling points and the smaller the maximum voltage interval between adjacent original voltage sampling points, the larger the fragment data quality coefficient.

[0009] As a further improvement to this technical solution, in step S3, the specific steps involved in calculating the static fragment degradation sensitivity based on the feature evaluation index are as follows: Extract the first The Pearson correlation coefficient and Spearman rank correlation coefficient corresponding to each initial health feature within each effective voltage segment are used as the feature evaluation index; A single-feature sensitive parameter is constructed based on the product of the absolute values ​​of the Pearson correlation coefficient and the Spearman rank correlation coefficient. In the first Within each effective voltage segment, the single-feature sensitive parameters of the initial health features in all dimensions are averaged and aggregated to obtain the static segment degradation sensitivity.

[0010] As a further improvement to this technical solution, in step S3, the static fragment degradation sensitivity is corrected using the fragment condition confidence parameter to obtain the fragment fusion weight. The specific steps involved are as follows: Extract the first The variance of current fluctuations and the absolute temperature gradient within each effective local voltage segment are used to characterize the changing state; Based on the preset current ripple tolerance threshold and thermal response hysteresis threshold, a dimensionless penalty calculation is performed on the current fluctuation variance and the absolute temperature gradient. The result of the dimensionless penalty calculation is then combined with the data quality coefficient of the segment to calculate the reliability parameter of the segment's operating condition. The static segment degradation sensitivity is multiplicatively coupled with the segment condition confidence parameter, and global normalization is performed on all valid local voltage segments in the current evaluation loop to calculate the segment fusion weight.

[0011] As a further improvement to this technical solution, in step S4, the evaluation network adopts a multilayer perceptron (MLP) or a one-dimensional convolutional neural network (1D-CNN). The specific steps involved in obtaining a rough estimate of the health status and the reliability of the aggregated operating conditions are as follows: For the effective local voltage segments extracted in the current evaluation cycle, the segment fusion health factor, segment fusion weight, segment effective identifier and segment operating condition confidence parameter corresponding to each effective local voltage segment are spatially aligned and spliced ​​according to the preset segment sequence dimension to construct the mask feature vector; The mask feature vector is input into a pre-trained lightweight evaluation network. Through the forward propagation of the evaluation network, a coarse estimate of the health status is output, and the aggregated working condition confidence is obtained by nonlinear aggregation based on the working condition confidence of each segment.

[0012] As a further improvement to this technical solution, the specific steps involved in calculating the soft membership degree of the degradation stage in S5 are as follows: Based on a rough estimate of health status, a stage discrimination vector is constructed by combining aging characterization parameters extracted within a preset historical window. Calculate the discrimination distance between the stage discrimination vector and a plurality of preset degradation stage reference centers, wherein the plurality of degradation stages include at least an early stable degradation stage, a mid-term nonlinear accelerated degradation stage, and a late fluctuating degradation stage; Based on the reliability of the aggregated operating conditions, a dynamic smoothing temperature coefficient is generated, and the dynamic smoothing temperature coefficient is used to perform a normalized exponential mapping on each of the discrimination distances, outputting the soft membership degree of the current cycle corresponding to each of the degradation stages.

[0013] As a further improvement to this technical solution, the specific steps involved in S6 are as follows: The mask feature vectors are respectively input into the expert models of each degradation stage pre-deployed in the battery management system, and independent health status prediction values ​​corresponding to the early stable degradation stage, the intermediate nonlinear accelerated degradation stage, and the late fluctuating degradation stage are output in parallel. Using the soft membership degree of each degradation stage as a dynamic weighting factor, a linear weighted sum is performed on each of the independent health state prediction values ​​to output the final health state assessment value of the battery.

[0014] On the other hand, the present invention provides a battery health assessment device with incomplete charging data, comprising: The data acquisition and segment recognition module is used to acquire incomplete charging data and identify valid local voltage segments; The feature construction module is used to resample effective local voltage segments and construct segment health status features; The credibility correction module is used to determine the credibility of the segment's operating condition based on the deviation from the operating condition and to correct the segment's degradation sensitivity. The fusion computing module is used to generate fusion health characteristics and soft membership degrees for degradation stages; The evaluation output module is used to output battery health status evaluation results by fusing expert models of multiple degradation stages.

[0015] On the other hand, the present invention provides a battery health assessment terminal with incomplete charging data, including a processor, a memory, a battery data communication interface and a result output interface; The memory stores computer programs that can be executed by a processor; The battery data communication interface is used to receive time-series voltage data, current data, charging capacity data, and temperature data collected by the battery management system or battery testing equipment. When the processor executes the computer program, it implements the non-complete charging data battery health assessment method described in any of the above-mentioned claims. The result output interface is used to send the SOH assessment results to the battery management system, charging controller, energy management system, or maintenance management platform.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In this method and system for assessing battery health using incomplete charging data, the segment fusion weight, segment valid identifier, and segment operating condition confidence parameter are used as mask parameters, which are then concatenated with the segment fusion health factor in the spatial dimension to form a mask feature vector. This processing method adapts to the dynamic changes in the number of effective local voltage segments measured at the underlying level through a pre-set masking mechanism, completely isolating invalid and missing data from interfering with the gradient backpropagation of the assessment network.

[0017] 2. In this method and system for assessing battery health using incomplete charging data, addressing the stage heterogeneity of the battery degradation process, this invention calculates the discriminant distance from the reference center of each degradation stage based on a coarse estimate of the health status. Simultaneously, it dynamically adjusts preset smoothing parameters using aggregated operating condition confidence to calculate the soft membership degree of each degradation stage. This method enables the model to smoothly adapt to the degradation patterns of different lifecycle stages, completely avoiding misjudgments of stage boundaries caused by hard threshold division.

[0018] 3. In this incomplete charging data battery health assessment method and system, the mask feature vector is input into the expert model for each degradation stage, and the soft membership degree of the degradation stage is used as a dynamic weighting factor to perform weighted summation on the output of each model. This can significantly reduce abrupt changes in the prediction curve at the stage boundary and significantly enhance the continuity and smoothness of the SOH estimation results. The algorithm-level non-jump output is directly mapped downwards, guiding the underlying control loop to perform continuous and smooth adjustment of the battery charging and discharging power limits, thus preventing relay action oscillations or power surges caused by abrupt changes in the health state output. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a graph showing the stages of battery capacity decay and degradation according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: Please refer to Figure 1-2 As shown, this embodiment provides a battery health assessment method based on incomplete charging data, including the following steps: S1. Obtain incomplete charging data of the battery to be evaluated, identify constant current charging intervals from the incomplete charging data, divide the constant current charging intervals into multiple candidate local voltage segments according to preset voltage positions (i.e., the voltage intervals defined by the lower voltage boundary and upper voltage boundary of each candidate local voltage segment configured in advance), and determine the effective local voltage segments from them. The incomplete charging data includes at least voltage data, current data, charging capacity data, and temperature data arranged according to sampling time, as well as charging step numbers. Specifically, the incomplete charging data refers to the current charging process covering only one or more local voltage intervals within the complete charging voltage range, without requiring coverage of the complete charging process from the charging start voltage to the charging stop voltage.

[0022] When the incomplete charging data contains a valid charging step number, the continuous data interval of the step type constant current charging is determined as a candidate constant current charging interval; when the charging step number is missing or incomplete, the fluctuation state of the charging current is analyzed according to a preset time window, and the continuous data interval in which the charging current remains stable is determined as a candidate constant current charging interval. In this embodiment, the value range of the preset time window is 5 to 30 seconds, preferably 10 seconds. The preset time window is determined according to the data sampling frequency and the current adjustment cycle of the charging device.

[0023] In one embodiment of the present invention, the steady state of the charging current is characterized by the coefficient of variation of the charging current within a preset time window: ; In the formula, Indicates the first The coefficient of variation of current within a preset time window is a dimensionless quantity. This represents the average charging current within a preset time window, measured in amperes. This represents the standard deviation of the charging current within a preset time window, in amperes. This represents a positive number to prevent the denominator from being zero, and its unit is ampere, preferably... ; This indicates the preset time window.

[0024] Specifically, when the average charging current within a preset time window reaches the preset charging current requirement (i.e., the direction of the average charging current is consistent with the preset charging direction, and the absolute value of the average charging current is not lower than the preset minimum charging current), and the current variation coefficient is not greater than the preset current stability threshold, the data interval corresponding to the preset time window is determined as the constant current candidate interval; wherein, the value range of the preset minimum charging current is 0.02C to 0.10C charging current corresponding to the rated capacity of the battery, and in this embodiment, it is preferably 0.05C charging current, where C represents the charging rate based on the rated capacity of the battery. The preset minimum charging current is used to exclude the data interval corresponding to the resting stage, small leakage current, and zero-point drift of the current sensor; the value range of the preset current stability threshold is 0.01 to 0.05, and in this embodiment, it is preferably 0.03.

[0025] Constant current candidate data intervals with continuous time and similar average charging current are merged. The time interval threshold between adjacent constant current candidate data intervals ranges from 1 to 5 seconds, preferably 2 seconds in this embodiment. The relative deviation threshold of the average charging current between adjacent constant current candidate data intervals ranges from 3% to 10%, preferably 5% in this embodiment. When the time interval between two adjacent constant current candidate data intervals is no greater than 2 seconds and the relative deviation of the average charging current is no greater than 5%, the two are merged to obtain a candidate constant current charging interval.

[0026] Furthermore, based on whether the voltage within the candidate constant current charging interval remains continuously rising, the candidate constant current charging interval is confirmed. Considering that voltage sampling noise may cause a small amount of instantaneous drop, the proportion of adjacent sampling points that meet the preset voltage rise condition is calculated: ; In the formula, Indicates the candidate constant current charging range; This indicates the number of effective voltage sampling points within the candidate constant current charging interval; and These represent the battery voltages at adjacent sampling points, in volts. This indicates the permissible voltage drop tolerance, in volts. Based on the voltage sensor's resolution, measurement error, and sampling noise level, in this embodiment, the voltage fall-off tolerance is determined. The value range is 0.001 to 0.005V, preferably 0.002V; This represents an indicator function, which takes the value 1 when the condition inside the parentheses is true, and takes the value 0 otherwise. This indicates the effective proportion of voltage rise in the candidate constant current charging range.

[0027] When the effective voltage rise ratio of the candidate constant current charging interval reaches the preset ratio threshold (the preset ratio threshold ranges from 0.85 to 0.98, preferably 0.90), specifically, when the effective voltage rise ratio of the candidate constant current charging interval is not less than 0.90, the end voltage of the interval is higher than the start voltage of the interval, and the duration of the interval and the number of effective sampling points meet the preset requirements, the candidate constant current charging interval is determined as a constant current charging interval.

[0028] The minimum duration of the candidate constant current charging interval is in the range of 30 to 120 seconds, preferably 60 seconds; the minimum number of effective sampling points is determined by the product of the minimum duration and the data sampling frequency; in a preferred embodiment of the present invention, when the data sampling frequency is 1 Hz, the minimum number of effective sampling points is preferably 60.

[0029] The method also determines whether the cumulative charging capacity within the constant current charging interval generally increases. When the cumulative charging capacity is reset, continuously decreases, or abnormally jumps, and when the proportion of abnormal capacity sampling points to valid sampling points within the constant current charging interval exceeds the abnormal capacity ratio threshold, the corresponding constant current charging interval is removed. The abnormal capacity ratio threshold ranges from 5% to 15%, and in this embodiment, a value of 10% is preferred.

[0030] Furthermore, based on the electrochemical system of the battery to be evaluated, the rated voltage range, and the common coverage of the training data, a local segment analysis voltage range is determined, and this local segment analysis voltage range is divided into multiple candidate local voltage segments with fixed voltage positions. Then, the first... The candidate local voltage segments are: ; in, ; In the formula, Indicates the first One candidate local voltage segment; and These represent the lower voltage boundary and upper voltage boundary of the candidate local voltage segment, respectively, in volts; This represents the total number of candidate local voltage segments. It is worth noting that the training sample batteries and the batteries to be evaluated use the same candidate local voltage segment numbers and upper and lower voltage boundaries of the segments, so that candidate local voltage segments with the same number correspond to the same physical voltage position. The candidate local voltage segments can be divided according to a fixed voltage interval, or they can be divided into non-equidistant intervals according to the sensitivity of different voltage ranges to battery degradation.

[0031] In this embodiment, the segment width of the candidate local voltage segment ranges from 0.03 to 0.10V, preferably 0.05V. For NCM lithium-ion batteries, 3.650V to 4.150V is set as the local segment analysis voltage range, and the segment is divided into 10 candidate local voltage segments with a segment width of 0.050V. The voltage range and voltage interval are only one preferred embodiment.

[0032] For each candidate local voltage segment, it is determined whether the constant current charging interval actually covers both the lower voltage boundary and the upper voltage boundary of the segment. When the constant current charging interval covers both the lower voltage boundary and the upper voltage boundary of the segment, and the number of valid sampling points within the segment is not less than a preset segment sampling point threshold, the candidate local voltage segment is determined as a valid local voltage segment, and its valid segment identifier is set to a valid state. The preset segment sampling point threshold ranges from 10 to 50, and is preferably 20 in this embodiment. The preset segment sampling point threshold is determined based on the original data sampling frequency, the width of the candidate local voltage segment, and the amount of data required for subsequent voltage domain resampling.

[0033] When a candidate local voltage segment does not simultaneously cover both the lower voltage boundary and the upper voltage boundary of the segment, or when the number of valid sampling points within the segment is less than 20, its valid segment identifier is set to invalid, and data outside the current actual voltage range is not used to extrapolate and complete the candidate local voltage segment.

[0034] When adjacent measured voltage sampling points cover the segment boundary, interpolation is used to obtain the data at the segment boundary, and the interpolation does not exceed the actual voltage range of the current constant current charging interval.

[0035] S2. Resample the effective local voltage segment to extract initial health features, construct a segment fusion health factor based on the feature evaluation index of the initial health features, and extract the segment data quality coefficient based on the resampled data state. In this embodiment, a fixed preset voltage differential step size is set. Regarding the first A number of effective local voltage segments are first constructed, with a fixed length of [missing information]. Resampled target voltage sequence : ; In the formula, Indicates the first The first effective local voltage segment One resampled target voltage node; This is the lower voltage boundary of the segment; Number the node index; This represents the total number of target nodes for resampling, and its calculation is based on... ,in For the voltage boundary of the segment, Indicates a round-down operation; voltage differential step size The value is set based on the lower limit of the actual sampling resolution of the BMS analog-to-digital converter (ADC) and the characteristic capture requirements of the polarization peak of the battery incremental capacity curve (IC curve) (too large a value will mask the local electrochemical phase transition characteristics, and too small a value will amplify high-frequency hardware noise). Its value range is 2mV to 5mV, and in this embodiment, 2mV is preferred.

[0036] Furthermore, the voltage sequence in the original discrete time domain is... As an independent input space, a conformal piecewise cubic interpolation algorithm is used as the mapping operator. For the original current sequence respectively Original capacity sequence and the original time series Perform voltage domain remapping of the dependent variable: ; In the formula, Represents the target voltage sequence Strictly aligned resampled current sequences; Represents the target voltage sequence Strictly aligned resampling capacity sequence; Represents the target voltage sequence Strictly aligned resampled time series; The mapping operator for the conformal piecewise cubic interpolation algorithm; Within each effective local voltage segment, based on the resampled time series and the resampled capacity series, multiple initial health features characterizing local charging capacity changes, charging time required per unit voltage change, and capacity-voltage curve shape are extracted. ( (as a feature number); specifically, the initial health features include, but are not limited to: the cumulative charging capacity within the effective local voltage segment (characterizing the change in local charging capacity), the ratio of the time difference to the voltage difference step size between adjacent resampling points within the segment (characterizing the charging time required for a unit voltage change), and the skewness or kurtosis coefficient calculated based on the capacity-voltage sequence (characterizing the geometric evolution of the capacity-voltage curve).

[0037] To eliminate the differences in physical dimensions of multidimensional features and filter out weakly correlated features, a feature evaluation index pre-calibrated based on a historical battery aging dataset is invoked; wherein, the feature evaluation index includes the Pearson correlation coefficient between each initial health feature and the battery's true state of health (SOH). The linear tracking capability used to characterize the features; and the Spearman rank correlation coefficient of the initial health features with battery aging cycles. It is used to characterize the global monotonicity of a feature as it degenerates.

[0038] It should be noted that the historical battery aging dataset is constructed as follows: Sample batteries belonging to the same electrochemical material system as the battery to be evaluated are selected, and a standard cyclic aging test covering the entire life cycle is performed under preset temperature gradient and charge / discharge rate conditions; at each aging cycle node, the complete charging process operation messages of the sample batteries are collected and recorded (including at least the voltage, current, and capacity sequences in the discrete time domain), and the true state of health (SOH) label of the sample batteries under that aging cycle is obtained simultaneously through standard capacity calibration tests (such as the static ampere-hour integration method); the extracted charging process operation messages of each cycle are mapped and bound with the corresponding true state of health labels to form the historical battery aging dataset.

[0039] By combining the Pearson correlation coefficient and the Spearman rank correlation coefficient, the feature fusion weights corresponding to each initial health feature are determined using a normalized exponential function. : ; In the formula, and All represent preset weight adjustment coefficients, and satisfy the following conditions: In this embodiment, The value range is 0.3 to 0.45. The value range is 0.55 to 0.7, preferably... and ; This represents the total number of dimensions of the initial health features extracted; and The range of values ​​is strictly distributed within Within the interval, when the value approaches 1 or -1, it indicates that there is a very strong positive or negative correlation between the initial health feature and the actual health state of the battery (for example, as the battery ages, a certain feature value shows a strict physical increasing or decreasing trend). When the value approaches 0, it indicates that the feature has no substantial linear or monotonic correlation with the battery degradation trajectory.

[0040] Among them, according to the feature fusion weight The weighted fusion of multiple initial health features within the same effective local voltage segment (it is worth noting that the multiple initial health features have been dimensionlessly normalized) yields the segment fusion health factor: ; In the formula, Indicates the first The segment fusion health factor corresponding to each effective local voltage segment; Indicates the index number of the effective local voltage segment; The dimension index number of the initial health feature indicates the current dimensional index. The first of the multiple features extracted within a voltage segment A specific feature; This indicates the weighting logic used to determine the first weight. Feature fusion weights corresponding to each initial health feature; In the The first effective local voltage segment extracted from the first Initial health characteristics; It is worth noting that this reduces the dimensionality of multidimensional discrete features to convergence into a single fragment fused health factor, which greatly reduces the fitting difficulty and overfitting risk of the subsequent network.

[0041] Meanwhile, to quantify the distortion risk introduced by data extrapolation during interpolation resampling, a segment data quality coefficient is constructed for each effective local voltage segment. Specifically, the number of original valid sampling points for each valid local voltage segment before interpolation resampling is obtained. And based on the original number of valid sampling points and the number of resampling target points Determine the sampling point density coefficient : ; Furthermore, the maximum voltage interval between adjacent original voltage sampling points within this effective local voltage segment is extracted. And based on its step size with respect to the preset voltage difference. The degree of deviation between them is used to calculate the interpolation interval penalty coefficient. : ; In the formula, The preset distortion penalty attenuation factor ranges from 0.1 to 0.5, preferably 0.2. Specifically, the value of the attenuation factor is based on the error divergence characteristics of the conformal cubic interpolation algorithm during the data window period: the closer the original maximum voltage interval is to the preset differential step size, the smaller the deviation and the higher the penalty coefficient. The closer it is to 1, the more likely it is to deviate from 1. Conversely, if local sensor data loss causes a sudden increase in voltage range, the risk of non-physical deformation caused by polynomial fitting will increase exponentially. In this case, the penalty coefficient will be applied through... The adjustment of rapid decay; maximum voltage interval The effective range of values ​​sets a physical tolerance limit: that is (In the preferred state of this embodiment, The upper limit is 20mV), and in this embodiment, its value range is within... to between; The value range is 2mV to 5mV, and in this embodiment, 2mV is preferred.

[0042] It is worth noting that the closer the original maximum voltage interval is to the preset voltage differential step size, the smaller the deviation and the smaller the penalty coefficient. The closer it is to 1, the greater the penalty coefficient becomes; conversely, if a large amount of data loss occurs locally, causing a sudden increase in voltage, the penalty coefficient will decrease. It decays exponentially.

[0043] Specifically, the data segment quality coefficient is determined based on the sampling point density coefficient and the interpolation interval penalty coefficient: ; In the formula, This represents the quality coefficient of the data segment.

[0044] S3. Calculate the static segment degradation sensitivity based on the feature evaluation index; and calculate the segment operating condition reliability parameter by combining the segment data quality coefficient and the change state of charging current and battery temperature corresponding to the effective local voltage segment; use the segment operating condition reliability parameter to correct the static segment degradation sensitivity to obtain the segment fusion weight.

[0045] Because the main aging side reactions that occur in batteries at different states of charge (SOC) ranges have physical differences, the sensitivity of different local voltage segments to health status is inherently heterogeneous.

[0046] Specifically, to quantify this static physical difference, the first... Pearson correlation coefficients for each initial health characteristic within each effective local voltage segment correlation coefficient with Spearman's rank As the feature evaluation index.

[0047] This embodiment calculates the static fragment degradation sensitivity based on extracted feature evaluation indicators. : ; In the formula, Indicates the first Within the first effective local voltage segment Pearson correlation coefficients corresponding to each initial health characteristic; Indicates the first Within the first effective local voltage segment Spearman rank correlation coefficients corresponding to each initial health characteristic; Indicates the first Static segment degradation sensitivity of an effective local voltage segment.

[0048] The alternating operating conditions of real vehicles or energy storage power stations can generate severe charging current ripple and dynamic temperature gradients, which can violate the isothermal constant current assumption, cause severe polarization shifts not caused by battery aging, and thus lead to the failure of static degradation sensitivity.

[0049] Based on this, this embodiment calculates a dimensionless penalty for the current fluctuation variance and the absolute temperature gradient based on preset current ripple tolerance thresholds and thermal response hysteresis thresholds, and jointly constrains them with the segment data quality coefficient to calculate the segment operating condition reliability parameter. : ; In the formula, Indicates the quality coefficient of the data segment; This represents the variance of current fluctuations as parsed from the CAN message; This represents the absolute temperature gradient parsed from the CAN message; Indicates the first The segment operating condition reliability parameter for each effective local voltage segment.

[0050] It is worth noting that the preset current ripple tolerance threshold (Preferred) ) and thermal response hysteresis threshold (Preferred) The range of the Hall current sensor and the maximum thermal conductivity of the battery pack liquid cooling plate are strictly anchored respectively. This represents the preset current condition penalty factor. This represents the preset temperature penalty factor, where, and The range of values ​​is In this embodiment, the preferred value range is... Specifically, when the target battery is an energy storage power station battery with relatively stable operating conditions, its system has a low tolerance for charge and discharge ripple. Preferred (In one embodiment, the optimal value is 1.2) to perform strong attenuation penalty on disturbance segments that deviate from constant current; when the target battery is a new energy vehicle power battery under frequent alternating load conditions, its charging current is inevitably accompanied by high-frequency load shedding of the accessories. In order to prevent too many local segments from being "falsely killed" by the algorithm, resulting in data gaps, Preferred (In this second embodiment, the optimal value is 0.6); Furthermore, when the target battery adopts a liquid-cooled thermal management architecture, and the cell is a high-nickel ternary lithium-ion battery that is extremely sensitive to high-temperature polarization and thermal runaway, Preferred (In one embodiment, the optimal value is 1.2) to ensure that when the underlying NTC sensor detects an abnormal temperature gradient, the weight contribution of the heat-generating segment is exponentially and rapidly cut off, preventing pseudo-polarization characteristics from contaminating the evaluation network; and when the target battery adopts a natural air-cooling architecture or is a lithium iron phosphate battery with high thermal stability, it has reasonable thermodynamic hysteresis and temperature rise tolerance. Preferred (In this second embodiment, the optimal value is 0.8).

[0051] Furthermore, after obtaining the aforementioned parameters, this embodiment uses the fragment condition confidence parameter to perform a substantial correction on the static fragment degradation sensitivity, and calculates the fragment fusion weight. : ; In the formula, This represents the total number of valid local voltage segments successfully captured and retained in the current evaluation loop; Indicates the first The segment fusion weights corresponding to each effective local voltage segment; This indicates the first iteration in the current loop. The operating condition reliability parameters corresponding to each effective local voltage segment; This indicates the first iteration in the current loop. Static segment degradation sensitivity corresponding to each effective local voltage segment; Indicates the summation index number ( The range of values ​​is to ).

[0052] S4. The fragment fusion weight, fragment valid identifier and fragment working condition credibility parameter are used as mask parameters and concatenated with the fragment fusion health factor to form a mask feature vector, which is then input into the evaluation network to obtain a rough estimate of the health status and the aggregate working condition credibility.

[0053] Due to the actual number of effective local voltage segments acquired Given dynamic uncertainties (e.g., the current loop only captures one local segment, or multiple consecutive segments), to enable the evaluation network to adaptively handle such variable-length inputs, this embodiment focuses on the successfully retained segments in the current evaluation loop. One effective local voltage segment, extract segment fusion health factors Fragment fusion weights Valid segment identifier and segment working condition reliability parameters ; Among them, fragment fusion health factors This is used to characterize the core degradation state of the segment after feature dimensionality reduction and correlation weighting; Fragment fusion weights This is used to characterize the true and reliable contribution of the segment in the current loop after experiencing the current fluctuation variance and temperature gradient exponential penalty in step S3. Fragment valid identifier The binary status bit configured in step S1 (with a value of 0 or 1) is used to fill the missing bits with a structured mask when the number of segments does not reach the preset maximum value, so as to completely isolate the interference of invalid data on the network gradient. Fragment working condition reliability parameters It is a continuous physical scalar that characterizes the degree of packet loss and electrochemical polarization distortion of the underlying sensor, calculated by step S3.

[0054] The above four types of parameters are rigorously spatiotemporally aligned and concatenated along the segment sequence dimension to obtain a dimension of [dimensionality missing]. mask feature vector : ; In the formula, This represents the masked feature vector input to the evaluation network, and its underlying data structure contains... A two-dimensional tensor with 4 rows and 4 columns; The sequence index representing the effective local voltage segment; This represents the total number of valid local voltage segments that were successfully captured and retained in the current evaluation loop.

[0055] It is worth noting that, to ensure the lightweight evaluation network possesses high-precision generalization capabilities before actual vehicle installation or online operation, the evaluation network needs to be pre-trained with accelerated life test (ALT) data throughout its entire lifecycle in the cloud or offline server cluster. Specifically: data is collected on the target battery of the same model at different ambient temperatures ( ), different charge / discharge rates and different aging stages (SOH from decay to The system uses massive amounts of full and incomplete charging history data; through the offline feature extraction pipeline in steps S1 to S3, it generates corresponding historical mask feature vectors in batches. The input features of the training samples are used, and the true SOH values ​​obtained from the corresponding offline static test are used as the true labels for supervised learning. The mean squared error (MSE) loss function and the Adam optimizer are used for multiple rounds of iterative backpropagation training until the convergence error of the network on the validation set is less than a preset threshold (preferred). The offline solidification of network weight parameters is completed; after the model training is completed, the solidified network weight matrix is ​​burned into the non-volatile memory of the vehicle-side BMS or energy storage edge controller.

[0056] Furthermore, in order to perfectly adapt to automotive or energy storage edge computing nodes (such as those with a main frequency lower than...), Given the extremely limited computing power and RAM constraints of conventional automotive-grade MCUs or low-power AI accelerator chips, the evaluation network implemented strict lightweight constraints in its network architecture design. The lightweight evaluation network preferably uses a total number of parameters controlled within 150,000 and fewer than [a certain number of] floating-point operations (FLOPs). Miniature multilayer perceptron (MLP) or single-layer one-dimensional convolutional neural network (1D-CNN); the specific network layers include: input layer (receiving dimension is... (mask feature vector), compressed to 1 channel The hidden feature extraction layer and the coarse health status estimate are output in parallel through fully connected layers. With regard to the reliability of the aforementioned aggregation conditions ; Ensure that the time taken for a single forward inference is in the millisecond range ( ); Within a certain range, the engineering risks of BMS underlying real-time control thread blocking or memory overflow caused by excessively heavy deep learning models are completely avoided.

[0057] Through forward inference of this evaluation network, the system outputs a coarse estimate of the health status in parallel. and the credibility of aggregated operating conditions ; ; In the formula, This represents a rough estimate of the health status of the network output. This represents a forward nonlinear mapping operator composed of the micro multilayer perceptron or a one-dimensional convolutional neural network; This represents the set of parameters, including the network weight matrix and bias vector, that have been solidified through multiple iterations during the pre-offline training phase of the evaluation network.

[0058] Rough estimate of health status The baseline SOH estimate for large-scale coarse positioning throughout the entire life cycle is directly used as a target discrimination parameter to calculate the discrimination distance to each degradation stage in subsequent steps; Aggregate operating condition reliability To obtain the working condition confidence parameter for all valid segments in the current loop We obtain the following by performing weighted summation and aggregation:

[0059] Among them, the reliability of the aggregation condition As a macroscopic representation of environmental operating conditions, it is used as a smoothing parameter for dynamically adjusting the soft membership calculation during the degradation stage.

[0060] S5. Calculate the discrimination distance to each degradation stage based on the coarse estimate of the health status, and adjust the preset smoothing parameter using the credibility of the aggregated working condition, thereby calculating the soft membership degree of the degradation stage.

[0061] Because the capacity degradation of lithium-ion batteries throughout their entire lifespan is not a single, stationary process, but rather exhibits significant stage heterogeneity, this embodiment avoids the predictive jumps (abrupt boundary changes) caused by traditional methods of defining degradation stages using a single fixed threshold. Therefore, this embodiment uses the health state as a coarse estimate. Using the core anchor point and combining aging characterization parameters extracted within a preset historical sliding window, a stage discrimination vector is constructed. : ; In the formula, This represents the current capacity decay slope calculated based on historical window data. This represents the local characteristic fluctuation coefficient calculated based on historical window data, used to characterize the degree of oscillation in capacity degradation. Specifically, in this embodiment, a fixed length is preferably used. A sliding window (within a cycle) is used to calculate the coefficient of variation of the SOH sequence or the characteristic sequence of the local constant current segment within that window. In the actual degradation process of lithium-ion batteries, the fluctuation coefficient... The range of values ​​is ; The shape index of the incremental capacity (IC) curve is specifically the attenuation of the main peak or the shift of the peak position.

[0062] Preliminary offline statistical or single-cell life tests (such as...) Figure 2 As shown), the early stable degradation stage is established ( ), mid-term nonlinear accelerated degradation stage ( ) and the later stage of fluctuation and degradation ( () stage reference center; It is worth noting that, Figure 2In the graph, the horizontal axis "Cycle" represents the number of battery cycles, and the vertical axis "SOH" represents the battery's state of health. Multiple curves in the graph illustrate the changes in SOH as the number of cycles increases along different capacity decay trajectories. Specifically, the legend "#1" indicates... Figure 2 The first curve in the chart shows the capacity decay trajectory of SOH as a function of cycle number; "#2" indicates the attached curve. Figure 2 The second curve in the chart shows the capacity decay trajectory of SOH as a function of cycle number; "#3" indicates the attached curve. Figure 2 The third curve in the figure shows the capacity decay trajectory of SOH as a function of cycle number; "#4" indicates the attached curve. Figure 2 The fourth curve in the figure shows the capacity decay trajectory of SOH as a function of cycle number; "#5" indicates the appendix. Figure 2 The fifth curve in the chart shows the capacity decay trajectory of SOH as a function of cycle number; "#6" indicates the appendix. Figure 2 The sixth curve in the chart shows the capacity decay trajectory of SOH as a function of cycle number; "#7" indicates the appendix. Figure 2 The seventh curve in the chart shows the capacity decay trajectory of SOH as a function of cycle number; "#8" indicates the appendix. Figure 2 The eighth curve shows the capacity decay trajectory of SOH as a function of the number of cycles.

[0063] Let the first Each stage of degradation ( The reference center coordinates of ) are Then the current stage discriminant vector To the Weighted discriminant distance of reference centers at each degradation stage Defined as: ; In the formula, Indicates the first offline calibration The baseline coordinates of the health status at each stage of degradation; Indicates the first offline calibration The reference center coordinates of the decay slope of each degradation stage; Indicates the first offline calibration The reference center coordinates of the fluctuation coefficient for each degradation stage; Indicates the first offline calibration The reference center coordinates of the IC curve topology morphology in each degradation stage; These are dimensional weight coefficients calibrated offline and embedded in the BMS controller, used to adjust the engineering contribution rate of each physical quantity in stage discrimination, taking into account the actual degradation physical characteristics of energy storage and vehicle power batteries. Similarly, due to Then Substituting each into the weighted discriminant distance formula, we obtain the following results. , as well as Specifically, Indicates the current battery state as it is approaching the early stable degradation stage. (Reference Center) Weighted Manhattan distance; Indicates the current battery state as being far from the mid-term nonlinear accelerated degradation stage. (Reference Center) Weighted Manhattan distance; Indicates the current battery state and the distance from the later stage of fluctuation and degradation. Reference Center Weighted Manhattan distance; The scale parameters representing health status characteristics use the same scale as SOH. The characteristic scale parameter of the fluctuation coefficient is a dimensionless quantity. The characteristic scale parameter of the IC curve shape index is represented by... Same dimensions; The characteristic scale parameter representing the capacity decay slope is... The dimensions are the same; it is worth mentioning that all the above scale parameters are positive numbers, which can be determined by the standard deviation, interquartile range or the difference between the maximum and minimum values ​​of the corresponding features in the historical aging dataset, to ensure that the difference terms are all unified to the dimensionless space before weighting.

[0064] It is worth noting that, in this embodiment, the preferred... The value range is set to , The value range is set to , The value range is set to , The value range is set to .

[0065] Furthermore, to completely block the error propagation from the underlying hardware sensing noise to the top-level model inference, the reliability of the aggregated operating conditions will be... As a penalty variable, the dynamic smoothing temperature coefficient is calculated. : ; In the formula, This indicates the system's set baseline temperature hyperparameter, and its value range is... In this embodiment, The preferred value is 0.15; This represents the hardware smoothing compensation gain (a constant) for BMS underlying sampling noise calibration. Its value range is In this embodiment, The preferred value is 0.20.

[0066] It is worth noting that when NCM system batteries enter the mid-term nonlinear accelerated degradation stage, the surge in their polarization resistance causes the voltage response of the local constant current charging segment to be highly volatile; preferably set Cooperate The benchmark anchoring is able to Break Under high-noise operating conditions, Instantly rise to In this way, at the level of mathematical normalization, the Euclidean distance exponent output in the early and middle stages is instantly leveled out, and at the level of physics, the false triggering jump caused by the local charging and discharging oscillations caused by the high activity of NCM batteries is perfectly suppressed.

[0067] The above dynamic smoothing temperature coefficient Inject a normalized exponential function to calculate the soft membership degree of the current loop belonging to each degradation stage: ; In the formula, This indicates that the current cycle to be evaluated belongs to the first... The weighting factors for each degradation stage, due to ,therefore, , , Let represent the soft membership degrees corresponding to the early, middle, and late degradation stages, respectively, and satisfy . and .

[0068] It is worth noting that when When the value is large, the weight distribution between different degradation stages is forced to become flatter; when Smaller (i.e., excellent operating conditions) Approaching When the sample membership is 0, the membership degree sharply points to the nearest physical degradation stage.

[0069] S6. Input the mask feature vector into the expert fusion model of each stage respectively, and use the soft membership degree of the degradation stage to weight and sum the outputs of each model to obtain the health status assessment result.

[0070] Because the capacity decay mechanism of lithium-ion batteries involves fundamental physical transitions between the early slow SEI film growth stage, the mid-stage accelerated loss of active material (LAM) stage, and the late-stage local lithium plating oscillation stage, a single global model is prone to underfitting in local areas. On the other hand, using a multi-model hard threshold switching method can cause a step jump in the SOH output curve at the stage boundary, which in turn can trigger oscillations in the BMS thermal management system or charge / discharge current limiting commands.

[0071] Therefore, in this embodiment, three independent lightweight degradation stage expert models are pre-programmed and deployed in the non-volatile memory of the vehicle-side BMS or energy storage controller, which are defined as early expert models. Mid-term expert model With post-expert models To meet the real-time computing power constraints of vehicle-mounted and edge-side energy storage nodes, the three expert models mentioned above are required to adopt the same basic network architecture as the evaluation network described in step S4, and are subject to completely consistent lightweight constraints.

[0072] Specifically, the early, middle, and late-stage expert models preferably use a total parameter count controlled within a certain range. Within 10,000, the number of floating-point operations (FLOPs) is less than The evaluation method can be implemented using a miniature multilayer perceptron (MLP) or a single-layer one-dimensional convolutional neural network (1D-CNN). Furthermore, when the evaluation method is deployed on a cloud-based battery lifecycle big data analysis platform with high computing power, the expert model is freed from the aforementioned lightweight constraints and can independently or in combination with an ensemble learning and high-dimensional nonlinear regression architecture. In this case, the expert model can employ CatBoost, ExtraTrees, RandomForest, XGBoost, ELM, DELM, TCN, or other regression models. Specifically, for the mid-term nonlinear accelerated degradation stage, an XGBoost or CatBoost architecture based on gradient boosting trees is preferred, utilizing its built-in symmetric tree growth mechanism and histogram acceleration algorithm to process the mask feature vector. High-dimensional nonlinear segmentation is performed to accurately capture local feature heteroscedasticity caused by abrupt changes in internal resistance. For the later fluctuation and degradation stage, a TCN (Temporal Convolutional Network) architecture is preferred, which extracts long-term dependencies in the feature vector through dilated causal convolution kernels to smooth and suppress high-frequency temporal perturbations caused by local lithium plating. After parallel extrapolation of the above heterogeneous models, they also output their respective independent health status predictions. .

[0073] When performing online inference, the dimension size is... mask feature vector As shared input features, these features are fed into the three expert models in parallel to obtain independent health status predictions for each physical degradation subspace: ; ; ; In the formula, , , These represent the corresponding proprietary parameter sets assuming the current battery is absolutely in the early, middle, or late degradation stage. The refined SOH evaluation value is obtained from parallel inference of the homologous network. It is worth noting that the weight parameter sets of each expert model are absolutely independent and isolated; the weight parameter sets... In the pre-offline training phase, the Accelerated Lifetime Trial (ALT) dataset is strictly divided based on the physical degradation stage, and then trained independently for each stage. This ensures that each homologous model can achieve extremely high optimization accuracy in its own local physical degradation subspace.

[0074] Based on the reliability of the underlying hardware operating conditions The soft membership vector with dynamic smoothing constraints is used to perform a linear weighted summation at the tensor dot product level on the above independent predicted values, outputting a smooth and abrupt final health status assessment value. : ; because This fusion mechanism forms a rigorous convex hull mapping in mathematical topology. When the battery is in the boundary blind zone of the degradation stage or suffers from local test noise interference, it forcibly smooths out the extreme value spikes that a single expert model may output through the smooth transition of soft membership.

[0075] Furthermore, the final health status assessment value The data is written to the master control register of the BMS via the internal data bus, serving as the macroscopic constraint boundary of the underlying hardware actuator.

[0076] Specifically, the BMS underlying control thread, based on the received... The attenuation curve is used to calculate and generate dynamic adjustment instructions for the current battery charge and discharge power limits by looking up tables or calling dynamic current limiting functions; based on the dynamic smoothing temperature coefficient. And the soft-weighted fusion mechanism, the It possesses extremely strong continuity of the first derivative.

[0077] As the charging and discharging current limiting command gradually decreases with the decay of SOH, the current limiting threshold changes in a smooth, gradual curve. This completely avoids engineering disasters that seriously endanger driving and energy storage safety, such as frequent closing and opening of relays or sudden cliff-like power loss caused by the SOH jump of traditional hard classification.

[0078] Example 2: This example provides a battery health assessment device for incomplete charging data, characterized in that it includes a data acquisition and fragment identification module, used to acquire incomplete charging data and identify valid local voltage fragments; It also includes a feature construction module, which is used to resample effective local voltage segments and construct segment health status features; It also includes a credibility correction module, which is used to determine the credibility of the segment's operating condition based on the deviation from the operating condition and to correct the segment's degradation sensitivity. It also includes a fusion computing module for generating fusion health characterization and soft membership of degradation stages; It also includes an evaluation output module, which is used to output battery health status evaluation results by fusing expert models of multiple degradation stages.

[0079] Example 3: This example also provides a battery health assessment terminal with incomplete charging data, including a processor, a memory, a battery data communication interface, and a result output interface; The memory stores computer programs that can be executed by a processor; The battery data communication interface is used to receive time-series voltage data, current data, charging capacity data, and temperature data collected by the battery management system or battery testing equipment. When the processor executes the computer program, it implements the non-complete charging data battery health assessment method described in any of the above-mentioned claims. The result output interface is used to send the SOH assessment results to the battery management system, charging controller, energy management system, or maintenance management platform.

[0080] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for assessing battery health using incomplete charging data, characterized in that, Includes the following steps: S1. Obtain incomplete charging data of the battery to be evaluated, identify the constant current charging range and divide it into multiple candidate local voltage segments, select the effective local voltage segments and configure the effective segment identifier. The incomplete charging data includes at least battery voltage, charging current, cumulative capacity, and battery temperature. S2. Resample the effective local voltage segment to extract initial health features, construct a segment fusion health factor based on the feature evaluation index of the initial health features, and extract the segment data quality coefficient based on the resampled data state. S3. Calculate the static segment degradation sensitivity based on the feature evaluation index; and calculate the segment operating condition reliability parameter by combining the segment data quality coefficient and the change state of charging current and battery temperature corresponding to the effective local voltage segment. The static segment degradation sensitivity is corrected using the segment condition confidence parameter to obtain the segment fusion weight; S4. The fragment fusion weight, fragment valid identifier and fragment working condition credibility parameter are used as mask parameters and concatenated with the fragment fusion health factor to form a mask feature vector, which is then input into the evaluation network to obtain a rough estimate of health status and aggregate working condition credibility. S5. Calculate the discrimination distance to each degradation stage based on the rough estimate of the health status, and adjust the preset smoothing parameter using the credibility of the aggregated working condition, thereby calculating the soft membership degree of the degradation stage. S6. Input the mask feature vector into the expert fusion model of each stage respectively, and use the soft membership degree of the degradation stage to weight and sum the outputs of each model to obtain the health status assessment result.

2. The battery health assessment method based on incomplete charging data according to claim 1, characterized in that, In step S1, determining the effective local voltage segment from the candidate local voltage segments involves the following steps: Based on the fluctuation of the charging current within a preset time window, candidate constant current charging intervals are determined from the incomplete charging data, and it is determined whether the voltage within the candidate constant current charging interval continues to rise. The candidate constant current charging interval is defined as the constant current charging interval if the voltage continues to rise and the duration and number of effective sampling points meet the preset requirements. Determine whether the constant current charging interval simultaneously covers the lower voltage boundary and upper voltage boundary of each candidate local voltage segment. The candidate local voltage segment that simultaneously covers the lower voltage boundary and upper voltage boundary and whose number of effective sampling points within the segment meets the preset requirements is determined as an effective local voltage segment.

3. The battery health assessment method based on incomplete charging data according to claim 2, characterized in that, In step S2, the segment fusion health factor corresponding to each effective local voltage segment is obtained. The specific steps involved are as follows: The effective local voltage segment is resampled by voltage domain interpolation according to a preset voltage difference step size, and the discrete time domain sampling sequence is converted into a resampled voltage sequence of fixed length, as well as a resampled current sequence, a resampled capacity sequence and a resampled time sequence aligned with the resampled voltage sequence. Based on the resampled time series and the resampled capacity series, multiple initial health features are extracted to characterize local charging capacity changes, charging time required for unit voltage changes, and the shape of the capacity-voltage curve. The feature evaluation index, which is pre-calibrated based on a historical battery aging dataset, is invoked. The feature evaluation index includes the Pearson correlation coefficient between each initial health feature and the actual health state of the battery, and the Spearman rank correlation coefficient of the initial health feature with the battery aging cycle. The feature fusion weights corresponding to each initial health feature are determined based on the Pearson correlation coefficient and the Spearman rank correlation coefficient. Then, multiple initial health features within the same effective local voltage segment are weighted and fused according to the feature fusion weights to obtain the segment fusion health factor.

4. The battery health assessment method based on incomplete charging data according to claim 3, characterized in that, In step S3, the static fragment degradation sensitivity is calculated based on the feature evaluation index, and the specific steps involved are as follows: Extract the first The Pearson correlation coefficient and Spearman rank correlation coefficient corresponding to each initial health feature within each effective voltage segment are used as the feature evaluation index; A single-feature sensitive parameter is constructed based on the product of the absolute values ​​of the Pearson correlation coefficient and the Spearman rank correlation coefficient. In the first Within each effective voltage segment, the single-feature sensitive parameters of the initial health features in all dimensions are averaged and aggregated to obtain the static segment degradation sensitivity.

5. The battery health assessment method based on incomplete charging data according to claim 4, characterized in that, In step S3, the static fragment degradation sensitivity is corrected using the fragment condition confidence parameter to obtain the fragment fusion weight. The specific steps involved are as follows: Extract the first Current fluctuation variance and absolute temperature gradient within an effective local voltage segment; Based on the preset current ripple tolerance threshold and thermal response hysteresis threshold, a dimensionless penalty calculation is performed on the current fluctuation variance and the absolute temperature gradient. The result of the dimensionless penalty calculation is then combined with the data quality coefficient of the segment to calculate the reliability parameter of the segment's operating condition. The static segment degradation sensitivity is multiplicatively coupled with the segment condition confidence parameter, and global normalization is performed on all valid local voltage segments in the current evaluation loop to calculate the segment fusion weight.

6. The battery health assessment method based on incomplete charging data according to claim 5, characterized in that, In step S4, the evaluation network employs a multilayer perceptron or a one-dimensional convolutional neural network. The specific steps involved in obtaining a rough estimate of the health status and the reliability of the aggregated operating conditions are as follows: For the effective local voltage segments extracted in the current evaluation cycle, the segment fusion health factor, segment fusion weight, segment effective identifier and segment operating condition confidence parameter corresponding to each effective local voltage segment are spatially aligned and spliced ​​according to the preset segment sequence dimension to construct the mask feature vector; The mask feature vector is input into a pre-trained lightweight evaluation network. Through the forward propagation of the evaluation network, a coarse estimate of the health status is output, and the aggregated working condition confidence is obtained by nonlinear aggregation based on the working condition confidence of each segment.

7. The battery health assessment method based on incomplete charging data according to claim 6, characterized in that, In step S5, the specific steps involved in calculating the soft membership degree during the degradation stage are as follows: Based on a rough estimate of health status, a stage discrimination vector is constructed by combining aging characterization parameters extracted within a preset historical window. Calculate the discrimination distance between the stage discrimination vector and a plurality of preset degradation stage reference centers, wherein the plurality of degradation stages include at least an early stable degradation stage, a mid-term nonlinear accelerated degradation stage, and a late fluctuating degradation stage; Based on the reliability of the aggregated operating conditions, a dynamic smoothing temperature coefficient is generated, and the dynamic smoothing temperature coefficient is used to perform a normalized exponential mapping on each of the discrimination distances, outputting the soft membership degree of the current cycle corresponding to each of the degradation stages.

8. The battery health assessment method based on incomplete charging data according to claim 7, characterized in that, The specific steps involved in S6 are as follows: The mask feature vectors are input into the pre-deployed expert models for each of the degradation stages, and independent health status prediction values ​​corresponding to the early stable degradation stage, the intermediate nonlinear accelerated degradation stage, and the late fluctuating degradation stage are output in parallel. Using the soft membership degree of each degradation stage as a dynamic weighting factor, a linear weighted sum is performed on each of the independent health state prediction values ​​to output the final health state assessment value of the battery.

9. A battery health assessment device with incomplete charging data, characterized in that, include: The data acquisition and segment recognition module is used to acquire incomplete charging data and identify valid local voltage segments; The feature construction module is used to resample effective local voltage segments and construct segment health status features; The credibility correction module is used to determine the credibility of the segment's operating condition based on the deviation from the operating condition and to correct the segment's degradation sensitivity. The fusion computing module is used to generate fusion health characteristics and soft membership degrees for degradation stages; The evaluation output module is used to output battery health status evaluation results by fusing expert models of multiple degradation stages.

10. A battery health assessment terminal with incomplete charging data, characterized in that, This includes the processor, memory, battery, data communication interface, and result output interface; The memory stores computer programs that can be executed by a processor; The battery data communication interface is used to receive time-series voltage data, current data, charging capacity data, and temperature data collected by the battery management system or battery testing equipment. When the processor executes the computer program, it implements the battery health assessment method for incomplete charging data as described in any one of claims 1-8. The result output interface is used to send the SOH assessment results to the battery management system, charging controller, energy management system, or maintenance management platform.

Images