Battery health status prediction method

Abstract

This invention discloses a battery health state prediction method, belonging to the field of battery management. This method utilizes only the cross-cycle health feature sequence corresponding to the prediction time and a preset number of historical cycles prior to it for mode decomposition. This invention conforms to the information availability boundary of online prediction, avoiding future information leakage caused by pre-decomposition of the entire lifespan sequence. Furthermore, it employs a dual-branch temporal convolutional network with different time receptive fields to handle trends and fluctuations separately, and then fuses them through data-driven gating weights, thus taking into account both long-term degradation patterns and short-term local changes. Further, it constructs a cross-cycle comparable local IC sequence through a common available voltage range and a unified voltage grid, not requiring each cycle to reach a fully charged state, reducing the dependence of SOH prediction on the complete charging process.

Description

Technical Field

[0001] This invention belongs to the field of battery management technology, specifically relating to a method for predicting battery health status. Background Technology

[0002] Lithium-ion batteries experience capacity decay, internal resistance increase, and electrochemical reaction kinetics changes during cycling. State of Health (SOH) characterizes the degree to which a battery retains its current usable capacity relative to its nominal capacity, and is a crucial basis for developing charge / discharge control strategies, determining battery retirement timing, and performing safe maintenance.

[0003] Incremental Capacity (IC) analysis transforms the electrochemical reaction plateau in conventional charging curves into easily identifiable peak features by calculating the rate of change of charging capacity relative to voltage. Parameters such as IC peak value, peak position, peak area, and peak width exhibit regular changes with battery aging and can serve as sensitive health characteristics for SOH prediction. Existing IC analysis methods typically require obtaining complete constant-current charging data from a lower state of charge to a fully charged state. However, in the actual operation of electric vehicles and energy storage systems, user charging behavior is random; charging can start from any state of charge or terminate before reaching the full charge voltage. This leads to inconsistent start and end voltages across different cycles, often accompanied by data gaps and uneven sampling, making it difficult to obtain complete IC curves online.

[0004] To address the aforementioned issues, existing research has proposed a scheme based on local IC feature extraction. This scheme no longer relies on the complete charging curve, but instead dynamically extracts voltage and capacity data within a local voltage range around the IC peak position (hereinafter collectively referred to as "local IC segment"). A mapping relationship between this local IC segment and SOH (State of Health) is established using time-series or graph neural networks, thereby reducing the dependence of SOH prediction on complete constant current charging data. Building on this, subsequent research further arranges the health features (such as peak value and peak area) corresponding to the local IC segments extracted from different cycles in cyclic order, combining them into a cross-cycle health feature sequence. Time-series modeling is then used to capture the evolution of SOH with the number of cycles.

[0005] However, the aforementioned unified modeling approach across loops faces a fundamental contradiction when it is applied to online applications:

[0006] The cross-cycle health characteristic sequence contains a mixture of long-term degradation trends and short-term disturbances. These are two completely different types of information. The former is caused by the aging of internal battery materials and manifests as a slow, irreversible decline spanning tens to hundreds of cycles. The latter is caused by factors such as temperature changes, measurement noise, capacity regeneration, and fluctuations in operating conditions, and manifests as rapid changes with significant fluctuations within a few cycles. When these two are mixed in the same sequence and treated uniformly by the same model with the same receptive field, short-term disturbances may be misjudged as part of the long-term trend, and the long-term degradation pattern may be masked by local fluctuations, leading to biases in SOH prediction.

[0007] To address the aforementioned issue of timescale mixing, some studies have attempted to introduce Variational Mode Decomposition (VMD) to decompose cross-cycle health feature sequences, separating them into trend and fluctuation components. However, as a global decomposition algorithm, VMD typically requires a time-series signal of a certain length to determine the mode center and decomposition boundary when decomposing data from the current cycle. This means that the decomposition result of the current cycle may contain data information from future cycles. While this approach of incorporating future information into the current decomposition is feasible in offline retrospective analysis, it cannot be deployed in practical online prediction scenarios because future data has not yet been generated. Summary of the Invention

[0008] Purpose of the invention: In order to realize the deployment of VMD in actual online prediction scenarios, this invention proposes a battery health state prediction method. Under the constraint of relying only on current and historical data, this method can effectively separate long-term degradation trends and short-term disturbances in cross-cycle health feature sequences, so that the two types of information at different time scales can be modeled separately under their respective appropriate receptive fields, and finally adaptively fused to output accurate SOH prediction results.

[0009] Technical solution: This invention provides a method for predicting battery health status, comprising the following steps:

[0010] Acquire incomplete charging data of the battery under test in multiple consecutive historical cycles; and construct a cross-cycle health feature sequence based on the incomplete charging data.

[0011] At each prediction time, modal decomposition is performed using only the cross-cycle health feature sequence corresponding to that prediction time and the preset number of historical cycles before it to obtain the trend feature sequence characterizing the long-term degradation of the battery and the fluctuation feature sequence characterizing the local dynamic changes of the battery.

[0012] The trend feature sequence and the fluctuation feature sequence are extracted using a first time-series feature extraction branch with a first time receptive field and a second time-series feature extraction branch with a second time receptive field, respectively, to obtain trend coding features and fluctuation coding features; wherein the first time receptive field is larger than the second time receptive field.

[0013] Based on the gating weight jointly determined by the trend coding feature and the fluctuation coding feature, the trend coding feature and the fluctuation coding feature are fused, and the predicted health status value of the battery under test for at least one future cycle is output based on the fused features.

[0014] The core concept of this application is as follows: at each prediction time, a cross-cycle health feature window is constructed using only the current prediction time and a preset number of historical cycles. Rolling mode decomposition is performed without introducing real future cycle data, thereby reducing the right-end boundary effect of the finite historical window and ensuring that the trend-fluctuation decomposition process meets the information availability boundary of online prediction. Based on this, the reconstructed trend features and fluctuation features are respectively input into time series feature extraction branches with different time receptive fields, and adaptive fusion is performed through a gating mechanism, thereby taking into account both long-term degradation patterns and recent local changes, and realizing future cycle SOH prediction.

[0015] Compared with existing technologies, this invention has at least the following advantages: It performs rolling mode decomposition using only the current and previous historical cycles at each prediction time, conforming to the information availability boundary of online prediction and avoiding future information leakage caused by pre-decomposition of the entire lifetime sequence. Furthermore, it employs a dual-branch temporal convolutional network with different temporal receptive fields to handle trends and fluctuations separately, and then fuses them through data-driven gating weights, thus taking into account both long-term degradation patterns and short-term local changes.

[0016] Further utilizing correlation, nonlinear dependence, stability across batteries or operating conditions, and redundancy to jointly evaluate candidate windows enables the target window to be adaptively determined according to the range of available data, and reduces overfitting to a single battery, a single temperature, or adjacent repeated windows.

[0017] Furthermore, by constructing a local IC sequence that is comparable across cycles through a common available voltage range and a unified voltage grid, it is not required that each cycle reach a full charge state, thus reducing the dependence of SOH prediction on the complete charging process.

[0018] Furthermore, the trend component and fluctuation component are adaptively divided according to the center frequency interval, which reduces the subjectivity of manually assigning fixed mode numbers and enables the decomposition strategy to adapt to the spectral differences of different health characteristics.

[0019] Furthermore, by using capacity-bounded mapping and monotonicity, smoothness, and boundary constraints, the prediction results are made to satisfy the physical range of battery capacity and SOH, and unreasonable abrupt changes are suppressed, thereby improving the stability and interpretability of the prediction results. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the battery health status prediction method of the present invention;

[0021] Figure 2 This is a schematic diagram of the battery incomplete charging curve and its partial IC curve of the present invention. The shaded area in the figure represents the common usable voltage range of 64 consecutive historical cycles.

[0022] Figure 3 This is a schematic diagram of the adaptive evaluation of the candidate voltage window and the selection of the target sensitive window in this invention;

[0023] Figure 4 This is a schematic diagram of the rolling VMD mode decomposition and trend-fluctuation reconstruction results of cross-cycle health characteristics according to the present invention.

[0024] Figure 5 This is a schematic diagram of the dual-branch TCN with different temporal receptive fields and the gated fusion structure of the present invention.

[0025] Figure 6 This is a schematic diagram illustrating the prediction results and prediction errors of future cycle SOH according to the present invention. Detailed Implementation

[0026] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0027] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0028] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0029] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0030] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0032] In the field of lithium-ion battery state of health (SOH) prediction technology, the most common application scenario is the online operation of battery management systems (BMS), such as daily charging of electric vehicles and scheduled charging and discharging of energy storage power stations. However, SOH prediction in these scenarios faces a practical problem: existing SOH prediction methods based on incremental capacity (IC) analysis typically require complete constant-current charging data from a low state of charge to a fully charged state. However, actual user charging behavior is random, making it difficult to obtain complete charging curves online. While schemes based on local IC feature extraction combined with cross-cycle time series modeling alleviate this problem to some extent, they model long-term degradation trends and short-term disturbances in the same sequence, leading to prediction bias. Although variational mode decomposition (VMD) can effectively separate these two types of components, its global decomposition characteristics depend on future data, making online deployment impossible.

[0033] Therefore, this application provides a battery health state prediction method in its embodiments. The core concept is that, under the constraint of relying only on current and historical data, by introducing a sequence decomposition strategy suitable for online scenarios, the long-term degradation trend and short-term perturbation in the cross-cycle health feature sequence are separated, so that the two types of information are modeled separately under their respective receptive fields and then adaptively fused, thereby achieving accurate prediction of SOH.

[0034] To facilitate understanding of the scheme described below, the following conventions are first established for key terms: a cycle refers to the completion of one effective charging event by the battery, providing the data samples required for local incremental capacity (IC) feature extraction; the last cycle refers to the most recent effective charging cycle completed before the current time; the cross-cycle health feature sequence refers to the time series formed by arranging the local IC health features extracted from each cycle in the order of cycle occurrence. Each training sample uses only the historical data of the last cycle and its preceding cycles at its last cycle to maintain the same information availability boundaries as in the online phase.

[0035] It should be understood that a valid charging cycle (or valid charging event, the two meanings are the same) requires that the voltage change during charging is greater than a preset threshold, such as a voltage increase exceeding 0.5V or a state of charge (SOC) increase exceeding 10%, to ensure that the charging segment carries sufficient electrochemical information for IC feature extraction. For example, if the user only charges for a few minutes (e.g., SOC from 80% to 83%, a change of 3%), this charging event is filtered out due to insufficient information and does not constitute a cycle; if the user charges from 30% to 80% SOC (a change of 50%), it constitutes a valid cycle, and the system records it as a cycle sample. It should be noted that the same continuous charging process is counted as one cycle, even if there are current fluctuations during charging, as long as charging is not interrupted, it is still considered a single cycle.

[0036] Based on the above terminology, this invention adopts a two-stage implementation architecture of offline training and online inference. The offline training stage utilizes historical charging data of the training battery and corresponding actual capacity or SOH tag to determine various parameters, specifically including:

[0037] Candidate window evaluation rules are used to adaptively filter the voltage range most sensitive to aging from local IC curves;

[0038] Feature standardization parameters are used to unify health characteristics of different dimensions to the same numerical range;

[0039] VMD parameters (including the number of modes and penalty factor, etc.) are used to control the granularity of sequence decomposition;

[0040] The parameters of the collaborative prediction model include the weights of the two-branch TCN and the gated fusion network.

[0041] All the above parameters are determined based on historical data and do not depend on any future information. During the online inference phase, the trained network weights are no longer updated. Instead, after obtaining new effective local charging segments of the battery under test, the parameters saved in the offline phase are called up, the most recent L effective historical loops are processed, and the predicted SOH value for future loops is output.

[0042] It is worth emphasizing that the training strategy of this invention ensures that each training sample uses only the historical data of the last loop and preceding loops at its final loop; in other words, for the first training sample... The training samples for each cycle are derived from input information only from the first to the last cycle. The loop does not include the first one. Any data after each cycle is processed to maintain the same information availability boundaries as during the online phase, fundamentally avoiding future information leakage problems.

[0043] Specifically, the following three stages of processing are performed sequentially: Stage 1, local IC health feature construction and sensitive window screening, which is used to extract health features sensitive to aging from the original charging data and construct a cross-cycle health feature sequence; Stage 2, rolling causal trend-fluctuation mode decomposition, which is used to separate the cross-cycle health feature sequence into long-term trend components and short-term fluctuation components without introducing future information constraints; Stage 3, dual-branch TCN and gated fusion prediction, which is used to extract time-series features from the two types of components and then adaptively fuse them to output the SOH prediction value.

[0044] The following will combine Figure 1 The specific steps of each stage are described in detail.

[0045] Phase 1 includes steps S1 to S5.

[0046] Step S1 involves acquiring incomplete charging data from the battery management system, charging equipment, or experimental testing apparatus. For the first... Each charging cycle records a set of data. In the formula, , and The first Time, voltage, and charging current at each sampling point; This refers to the number of sampling points in this cycle. Here, the number of sampling points refers to the total number of data entries, such as voltage, current, and temperature, actually recorded by the BMS at a fixed sampling frequency during this charging cycle.

[0047] Specifically, a BMS typically collects a set of data at fixed time intervals (such as every 1 second, every 0.1 seconds, or every 10 seconds). If the... The charging duration of each cycle is The sampling frequency is Then the total number of sampling points in this cycle For example, if the sampling interval is 1 second and a single charge lasts for 30 minutes (1800 seconds), then... One point.

[0048] Before executing step S2, the incomplete charging data obtained in step S1 is preprocessed. Specific processing steps include: for abnormal sampling points in the battery voltage or charging current that exhibit isolated abrupt changes relative to adjacent sampling data, but whose trends before and after the abrupt change remain continuous, these abnormal sampling points can be removed; for data segments with continuous abnormalities or missing lengths exceeding a preset threshold, and whose original trends cannot be reliably recovered, they will not participate in subsequent local IC sequence construction. After completing the abnormal data processing, the valid charging data from each historical cycle is used for determining the common available voltage range and unifying the voltage grid interpolation in step S2.

[0049] Step S2, determine the uniform voltage grid. For grids containing continuous voltages... The prediction window of each historical cycle is used to obtain the starting voltage of the effective segment of each cycle. and termination voltage The common usable voltage range can be expressed as: , In the formula, This indicates the number of valid charging cycles contained in the current historical cycle window; This indicates the circular index within the current history window. =1,2,…, ; and They represent the first The start and end voltages of each effective charging segment in a cycle; express The maximum value of the starting voltage of each cycle. express The minimum value among the cycle termination voltages. Therefore, [ , The above Each cycle actually covers the common available voltage range.

[0050] when When the width is not less than the preset minimum available width, the voltage interval shall be used within this range. Establish a unified grid : The time and current of each cycle are interpolated to this grid. The common available range here can be determined for a fixed training dataset or dynamically updated based on a rolling historical cycle window during online prediction.

[0051] Where ΔV represents the voltage interval of the uniform voltage grid, which can be predetermined during the training phase based on the voltage sampling resolution, measurement noise level, and analytical accuracy of the target IC peak, and remains consistent with the training phase during the online inference phase. The endpoints of the equally spaced grids do not completely coincide; the value can be no more than [a certain value]. The largest grid point is taken as the endpoint.

[0052] For example, if the voltage coverage ranges of three valid charging events are 3.68V–4.10V, 3.72V–4.08V, and 3.70V–4.05V respectively, then the common usable voltage range is [max(3.68,3.72,3.70),min(4.10,4.08,4.05)] = [3.72V,4.05V]. When ΔV = 0.002V, a unified grid of 3.720V, 3.722V, 3.724V, ..., 4.050V can be formed, with a total of 166 grid points. The time and current data of each historical charging event are interpolated according to this unified voltage coordinate, so that the local IC characteristics of different events can be compared at the same voltage position.

[0053] like If the common voltage range is less than the preset minimum available width, it is considered that there is no sufficiently wide common voltage range among the current L valid charging events: during the offline training phase, the training samples corresponding to this historical window are discarded; during the online inference phase, the update of the SOH prediction value based on the current window is temporarily suspended, or the prediction result that most recently meets the common available width requirement is output and marked as having decreased confidence, until a new valid charging event makes the common available voltage width meet the minimum available width again.

[0054] For the first in the unified voltage grid For each point, the local charging capacity is calculated using the integral recursive method based on ampere-hours:

[0055] ;

[0056] in, Indicates the unified voltage grid point number. and They represent the current number. cyclic interpolation to Time and charging current at the point, This represents the i-th voltage point in the aforementioned unified voltage grid; This indicates the accumulation from the current local charging segment's starting point to... Local charging capacity at the location, This represents the local charging capacity corresponding to the previous grid point. It should be noted that the unit of current is ampere, the unit of time is second, and the unit of capacity is ampere-hour. Due to the relative change in capacity with voltage in subsequent uses, the capacity at the starting point of the local segment... It can be set to zero; when the current sign is defined as the charging current taking a negative value, for Take the absolute value or change the sign before the integral. Here, 3600 is the unit conversion factor between seconds and hours. When the current unit is A and the time unit is s, the capacity unit after dividing the integral result by 3600 is Ah. If the time has been uniformly converted to h, the corresponding integral does not need to be divided by 3600 again.

[0057] Calculate the local IC value using central difference:

[0058] ;

[0059] In the formula, Indicates the current number The cycle is in the 1st... Local incremental capacity values ​​at a uniform voltage grid point; and These represent the local charging capacity corresponding to the adjacent next grid point and the previous grid point, respectively; and These represent the voltages corresponding to the adjacent next grid point and the previous grid point, respectively.

[0060] A local incremental capacity sequence is generated by sorting the local incremental capacity values.

[0061] To reduce the amplification effect of differentiation on measurement noise, Savitzky-Golay smoothing is performed on the IC sequence after differentiation; the smoothing window should be smaller than the typical width of the IC peak to be identified in order to preserve peak value and peak position information.

[0062] Step S3: Generate candidate voltage windows within the common usable voltage range and extract candidate health features that reflect the battery aging state. The j-th candidate voltage window is denoted as: Where j represents the candidate window number, and These represent the lower and upper boundary voltages of the j-th candidate window, respectively. All windows can have the same width. and step size Slide to generate, here is the window width , This indicates the sliding step size between the starting positions of adjacent candidate windows; different widths can also be set according to the extreme points and inflection points of the IC curve. Optionally, Take a voltage ranging from 0.05V to 0.20V. Select a value from 0.005V to 0.05V, depending on the battery system, sampling accuracy, and common available voltage range.

[0063] For each candidate voltage window, all historical cycles are iterated one by one, and health characteristics are calculated from the charging data and local IC data corresponding to each historical cycle within that window, thereby obtaining the sequence of changes in each health characteristic with the number of cycles under that candidate window. The candidate health characteristics are quantitative representations used in step S4 to evaluate the aging sensitivity, cross-condition stability, and redundancy of the candidate window, including at least one of interval charging time, IC curve area, IC peak value, peak position, peak width, average slope, and kurtosis.

[0064] The interval charging time is defined as:

[0065] ;

[0066] In the formula, This represents the interval charging time required for the current cycle to pass through the j-th candidate voltage window; and These represent the time-voltage interpolation relationship of the current cycle in terms of voltage. and The corresponding time value, therefore It reflects the change in the speed at which the battery passes through this local voltage range.

[0067] The area of ​​the IC curve can be calculated using trapezoidal integrals:

[0068] ;

[0069] In the formula, This represents the area of ​​the local IC curve within the j-th candidate voltage window of the current cycle; It represents the function of local incremental capacity as a function of voltage V after being calculated and smoothed by a unified voltage grid; and Let represent the lower boundary voltage and upper boundary voltage of the j-th candidate window, respectively, and let represent the lower limit of integration and the upper limit of integration, respectively. It is used to characterize the intensity of cumulative capacity change within this local voltage range.

[0070] In addition to the above-mentioned interval charging time and IC curve area In addition, it can also calculate peak-shaped health characteristics within the window: IC peak value. The maximum IC value within the j-th window, peak position. For the corresponding voltage; peak width The voltage difference at the intersection of the left and right points at half-peak height can be taken; average slope The difference between the IC values ​​at the top and bottom boundaries of the window and the window width can be taken. Ratio; kurtosis The fourth central moment of the normalized IC sample values ​​within the window can be determined. For a window containing multiple local peaks, the main peak and secondary peaks can be extracted separately. , and Alternatively, calculate the statistics for each peak parameter. (The above...) , , , , , and All of these are candidate health features, which can be used individually or combined into a candidate health feature vector. The purpose is to quantify the response of the voltage window to battery aging from different perspectives such as charging speed, local capacity change, and peak shape evolution.

[0071] Step S4: Establish a window evaluation model. First, using only training battery data, calculate the evaluation metrics between candidate health features and the true state of health (SOH). Correlation metrics. The absolute value of either the Pearson correlation coefficient or the Spearman rank correlation coefficient can be used; nonlinear dependence index Normalized mutual information is acceptable; stability index Used to evaluate whether the correlation direction and normalized change trend of the same feature are consistent across different training batteries, temperature ranges, or rate conditions; redundancy index It is used to characterize the degree of feature correlation between a candidate window and the selected window or other candidate windows.

[0072] In one specific implementation, firstly , , and Normalize to 0 to 1 respectively, then calculate:

[0073] ;

[0074] In the formula, This represents the overall evaluation value of the j-th candidate window; , , and Let α, β, and γ represent the correlation index, nonlinear dependence index, cross-cell or cross-operating condition stability index, and feature redundancy index, respectively. α, β, and γ are the weights of the three positive indices (correlation, nonlinear dependence, and stability) to control their relative contribution to the overall evaluation value, and α + β + γ = 1. λ is the redundancy penalty coefficient, used to suppress candidate windows that contain highly redundant information with the selected window. As an example, α = 0.35, β = 0.25, γ = 0.40, and λ = 0.20 can be chosen, where a larger γ indicates that cross-cell or cross-operating condition stability is prioritized in this example. These values ​​are for illustrative purposes only and can be determined through cross-validation using the training and validation sets.

[0075] according to Candidate windows are selected from high to low. When multiple target sensitive voltage windows are selected, the redundancy index of the remaining windows relative to the selected windows is recalculated after each window is added. The preset number of windows represents the upper limit of the number of target windows that can be selected; the evaluation value threshold represents the minimum acceptable comprehensive evaluation level of the candidate windows; the validation set performance gain represents the degree of improvement in the prediction error of the collaborative prediction model on the validation set after adding a new candidate window. Selection can be stopped when the upper limit of the target number of windows is reached, the comprehensive evaluation value of the remaining candidate windows is lower than the threshold, or the validation set performance gain after adding a new window is less than the preset gain threshold. This preserves the local intervals that are closely related to SOH and stable across operating conditions, while suppressing the duplication of information between adjacent windows.

[0076] Step S5: Arrange the health characteristics of the target sensitive voltage window in each cycle according to the cycle number. Let the total number of selected characteristics be... The number of consecutive historical cycles is Then, a cross-cycle health feature matrix is ​​formed at the current prediction time. Each feature dimension is standardized using the mean and standard deviation of the training set; the standardized parameters saved during the training phase are used for online prediction, and are not re-estimated using future iterations.

[0077] The above steps complete the construction of the cross-cycle health feature sequence, including the cross-cycle health feature matrix. In the formula, X represents the cross-cycle health feature matrix; L represents the number of effective charging cycles contained in the historical cycle window; D represents the number of health feature dimensions extracted in each cycle; n represents the cycle position arranged in the order of sequence within the window, n=1,2,…,L; d represents the health feature dimension index, d=1,2,…,D; Let X represent the d-th standardized health feature of the nth cycle. The nth row of matrix X corresponds to the nth historical cycle, and the d-th column corresponds to the d-th health feature.

[0078] However, this sequence contains a mixture of long-term degradation trends and short-term disturbances, and directly feeding both into the same model for time series prediction would lead to bias. To address this, in Phase Two of this embodiment, a rolling causal trend-fluctuation mode decomposition is performed on the above sequence under the constraint of relying only on current and historical data, in order to separate information from the two different time scales.

[0079] Phase two includes step S6.

[0080] Step S6 is used to separate long-term degradation changes and short-term local changes in cross-cycle health features without introducing information from future cycles. Specifically, when the current last cycle number is k, only the health features from the (k-L+1)th to the kth cycle are extracted to form the cross-cycle health feature matrix. Where L is the length of the historical loop window. For the d-th dimension of the health feature matrix, the variational mode decomposition method is used to decompose it into... The original sequence contains narrowband modal components with different center frequencies. Variational mode decomposition, under the constraint that the sum of the modal components can reconstruct the original healthy feature sequence, minimizes the sum of the modal bandwidths to distribute components with different rates of change in the original sequence to different frequency ranges, thus providing a basis for subsequent reconstruction of long-term trend features and short-term fluctuation features.

[0081] The constrained variational problem is:

[0082] ;

[0083] ;

[0084] In the formula, Indexing for health feature dimensions, ; For VMD modal index, ; For the current length is The time position within the historical loop window. ; For the first Health characteristics in location Standardized values ​​at the location; Indicates the first The first health characteristic One modal component; This indicates the center frequency corresponding to the mode; Indicates the total number of VMD modes; This refers to an operator that converts real-valued modes into analytic signals; Indicates relative to the cycle position Discrete difference operator; It is the imaginary unit.

[0085] The augmented Lagrange multiplier and alternating direction multiplier methods are used to iteratively update each modal component, center frequency, and Lagrange multiplier until the relative change between two adjacent iterations is less than the convergence threshold or the maximum number of iterations is reached. Optionally, the number of modes... The values ​​are 3 to 6, the quadratic penalty parameter is 1000 to 5000, and the convergence threshold is [value missing]. to The parameters are determined by the reconstruction error of the training set, the center frequency separation, and the prediction performance.

[0086] The same health feature dimension d corresponds to Center frequency Arrange them in ascending order and calculate the interval Δ between adjacent center frequencies. = - Where m = 1, 2, ..., k-1; then determine the division between trend mode and fluctuation mode according to the following formula: In the formula, Δ The center frequency interval between the m-th and (m+1)-th adjacent modes of the d-th health feature; This represents the modal segmentation position of the d-th health feature, i.e., the low-frequency side modal index corresponding to the maximum adjacent center frequency interval. Using the maximum center frequency interval as the segmentation basis is mainly to reduce the subjectivity of manually assigning the first few modes to the trend, and to enable different health features to adaptively determine the trend-fluctuation boundary based on their own spectral separation.

[0087] From the 1st to the The low-frequency modes are superimposed to form a trend sequence: , will the To the The high-frequency modes are superimposed to form a wave sequence: In the formula, This represents the trend component of the d-th dimension of health characteristics at position n. The d-th dimension of health features represents the fluctuation component at position n; the low-frequency mode is used to characterize the slowly changing degradation trend across multiple cycles, while the high-frequency mode is used to characterize local dynamic information such as recent degradation rate changes, capacity regeneration, temperature disturbances, and measurement fluctuations.

[0088] Repeat the above process for all D-dimensional health features, and combine them according to feature dimensions to obtain the trend feature matrix. and fluctuation characteristic matrix .in, The dth column is , The dth column is Both have the same cycle length L and feature dimension D as the original cross-cycle health feature matrix X, theoretically satisfying + ≈X, where the difference arises solely from numerical iteration and termination errors. If the maximum center frequency interval of a certain health feature is not significant, it can be... The number of trend modes is limited to between the minimum and maximum trend modes preset during the training phase to avoid all modes being classified into the same category.

[0089] This embodiment reuses the data up to the current cycle to perform decomposition at each prediction time. The construction of training samples also follows the same rule: for training samples with the actual capacity or health status of the (k+h)th cycle as the supervision label, where k represents the last cycle number in the current historical cycle window, and h represents the prediction step index, h=1,2,…,H, the input data of the training samples only includes historical cycle data up to the kth cycle and before, excluding the actual data from the (k+1)th cycle and after. This ensures that there is no future information leakage during the training, validation, and online prediction phases.

[0090] After obtaining the trend feature matrix and fluctuation feature matrix through the above decomposition, their temporal features need to be extracted separately using appropriate receptive fields, and then the two types of information are adaptively fused. To this end, in the third stage of this embodiment, a dual-branch TCN and gating fusion mechanism are used to execute steps 7 to 8.

[0091] Step S7, convert the trend feature matrix and fluctuation characteristic matrix The input consists of a trend TCN branch and a fluctuation TCN branch. The trend branch uses a larger temporal receptive field to represent long-term degradation patterns spanning many historical cycles, while the fluctuation branch uses a smaller temporal receptive field to extract local dynamic changes in recent cycles, thus obtaining trend encoding features and fluctuation encoding features respectively. Each branch is composed of multiple temporal convolutional residual units connected in series, and each residual unit includes two layers of dilated causal convolution, nonlinear activation, and residual connection.

[0092] For the l-th residual unit of the b-th TCN branch, where b distinguishes between trend and fluctuation branches, and l represents the residual unit layer number in this branch, the calculations of the first and second layer dilated causal convolutions are as follows:

[0093] ;

[0094] ;

[0095] In the formula, n represents the time position within the historical loop window; q represents the sampling position inside the convolution kernel; This indicates the length of the convolution kernel used in the b-th branch; This represents the expansion coefficient of the l-th residual unit in the b-th branch; This represents the input characteristics of the residual unit; and These represent the outputs of the first and second layer dilated causal convolutions of the residual unit, respectively. and These represent the weight matrices of the two convolutional layers at convolution position q, respectively. and These represent the bias vectors corresponding to the two convolutional layers, respectively. This represents a nonlinear activation function. The subscripts 1 and 2 in the formula are used to distinguish between the first and second convolutional layers within the same residual unit.

[0096] To ensure that the temporal convolution process satisfies causality, for input positions in the convolution operation that exceed the starting position of the current historical loop window, the corresponding input is set to zero, or the boundary value of the starting position of the health feature sequence is used for filling, so that the convolution output at position n is determined only by the historical health features at position n and before it, without using the information after position n.

[0097] For the l-th temporal convolutional residual unit in the b-th temporal feature extraction branch, the output of the second layer dilated causal convolution is residually concatenated with the input of this residual unit. The output can be expressed as:

[0098] ;

[0099] In the formula, This represents the residual mapping matrix of the l-th residual unit in the b-th branch; when the number of input and output channels is the same, Using unit mapping, when the number of channels is different, 1×1 convolution or equivalent linear mapping is used to match the channel dimension so that the feature dimension at both ends of the residual connection remains consistent. This represents the output of the l-th residual unit at position n after residual connection and nonlinear activation.

[0100] like Figure 5 As shown, the two branches are configured with different combinations of expansion coefficients, making the temporal receptive field of the trend branch larger than that of the fluctuation branch. For the case where each residual unit contains two convolutional layers with a kernel length of 3, the branch receptive field... Calculation. When the historical window length L=64, the trend branch is set with 4 residual units and the inflation coefficients are 1, 2, 4, and 8 respectively, and its receptive field... The fluctuation branch is set with two residual units and expansion coefficients of 1 and 2 respectively, and its receptive field is... . Figure 5 The upper branch corresponds to the trend TCN, and the lower branch corresponds to the fluctuation TCN. Thus, the trend branch can cover most of the cycles in the historical window to extract slow degradation patterns, while the fluctuation branch focuses on more recent cycles to extract local changes, avoiding the use of the same receptive field to process two types of time-scale information at the same time.

[0101] Temporal aggregation is performed on the final sequence outputs of the two branches respectively. The temporal aggregation operator can take the features from the last valid time position, global average pooling, or trainable weighted pooling; then the two aggregation results are linearly mapped to the same C-dimensional feature space to obtain the trend encoding features. and fluctuation coding features , where C represents the dimension of the unified coding features before gating fusion.

[0102] Step S8: Concatenate the two encoded features and calculate the vector gating weights.

[0103] ;

[0104] In the formula, This indicates that the trend is encoded as a feature. With fluctuation coding features Perform vector concatenation; the resulting dimension is 2C. The gated mapping weight matrix, The gating bias vector; sigmoid(·) represents the sigmoid function; g Let g be the gating weight vector, where each element of g is between 0 and 1. The fused features are:

[0105] ;

[0106] In the formula, ∈ ⊙ represents the fused encoding feature; ⊙ represents element-wise multiplication; 1-g represents the complementary gating weights corresponding to the fluctuation branch. When the value of g on a certain feature dimension is close to 1, that dimension focuses more on trend encoding features; when the value is close to 0, that dimension focuses more on fluctuation encoding features.

[0107] Therefore, when the long-term degradation patterns of the current sample are more reliable, the trained gating module can increase the trend encoding weights on the corresponding feature dimensions; when local temperature changes, capacity regeneration, or recent degradation rate changes are more important for prediction, the fluctuation encoding weights can be increased. The gating weights are determined jointly by the two branches, rather than by a pre-set fixed fusion ratio.

[0108] Fusion features Input health status prediction module. The first prediction mapping matrix converts the C-dimensional fused features into... The second prediction mapping matrix transforms the predicted hidden state into H capacity prediction intermediates:

[0109] ;

[0110] ;

[0111] Desirable , Predicting steps or ; This is a per-element nonlinear activation function. When performing a single-step prediction, it outputs the SOH of the next cycle; when performing a multi-step prediction, it outputs the SOH of multiple future cycles at once. To ensure the capacity results are within a reasonable range, the [function name] is [function name]. Perform bounded mapping on each intermediate variable:

[0112] ;

[0113] In the formula, It can be determined based on a conservative lower limit of capacity below the retirement threshold. This can be determined based on the new battery's nominal capacity and allowable initial capacity fluctuations. An optional setting is... , The final predicted SOH value is:

[0114] ;

[0115] In the formula, This represents the predicted health status value for the (k+h)th cycle in the future. This indicates the reference capacity used in the definition of SOH. If SOH is expressed as a percentage, then... Multiply by 100%. It can be the nominal capacity given by the manufacturer, or the initial capacity calibration value completed according to preset standards when the battery is put into use; the training label and online output should use the same definition.

[0116] In a preferred embodiment, to improve the accuracy of subsequent IC curve construction, before executing step S2, the raw charging data obtained in step S1 can be subjected to quality checks and preprocessing. Specifically, this includes: performing quality checks on the data for each charging cycle; identifying valid charging segments based on the charging current direction and current fluctuation range; checking the voltage changes of adjacent sampling points, the deviation of current from local neighborhood statistics, the continuity of timestamps, and the absence of sampling points. Sampling points that exhibit significant abrupt changes relative to adjacent sampling data, and where the abrupt change only lasts for one or a few consecutive sampling points while maintaining a continuous trend in data change before and after the abrupt change, are identified as isolated outliers and can be replaced by the median or interpolated value of the neighborhood before and after the point. For missing intervals whose duration does not exceed a preset missing length threshold, linear interpolation or conformal piecewise cubic interpolation can be used for repair. For segments that are continuously abnormal or whose missing length exceeds the threshold and cannot reliably recover the original trend, they are not included in the subsequent candidate window generation.

[0117] It should be understood that the raw charging data refers to the set of charging cycle data obtained in step S1. That is, based on sampling time Battery voltage and charging current The time-series data is composed of these elements.

[0118] Steps S1 to S8 above describe the forward computation process of this embodiment under given model parameters. To achieve offline deployment of the above process, it is necessary to pre-determine the window evaluation index, feature standardization parameters, VMD parameters, and network weights of the dual-branch TCN, gated fusion, and prediction modules using training data. Therefore, this embodiment further provides the following collaborative prediction model training scheme.

[0119] Different training cells are divided into training, validation, and test sets based on individual cells. Different cycles of the same cell are not mixed across sets to reduce data leakage caused by the similarity of adjacent cycles. Window evaluation metrics, feature normalization parameters, VMD parameters, and network hyperparameters are all determined solely based on the training and validation sets.

[0120] For each allowed final loop in the training set Construct a length of [length] according to steps S1 to S6. The trend feature matrix and volatility feature matrix, and with the first To the The actual capacity or state of health (SOH) of each cycle is used as a monitoring label. For lifetime sequence tails that are insufficient... Samples with individual labels are either discarded or the loss is calculated using a label mask.

[0121] Prediction error loss can be calculated using mean squared error, mean absolute error, or smoothed L1 loss. Taking mean squared error as an example:

[0122] ;

[0123] In the formula, B represents the number of samples in a training batch; s represents the sample index within the batch, s=1,2,…,B; h represents the prediction step index, h=1,2,…,H; This represents the true SOH corresponding to the s-th sample in the h-th prediction step. This represents the model's predicted SOH value for that location; This represents the loss due to prediction error. To avoid confusion in sign usage, subsequent monotonicity, smoothness, and boundary constraints will all use the predicted value. For calculation objects.

[0124] To allow for measurement errors or slight capacity regeneration, a non-negative tolerance threshold is set. ,in This represents the maximum allowable slight increase in SOH between adjacent prediction steps; monotonicity constraint loss. Used to punish exceeding The predicted increase in SOH can be expressed as:

[0125] ;

[0126] Smoothness constraint loss The second-order difference calculation based on the predicted SOH sequence is used to suppress unreasonable and drastic curvature changes between adjacent prediction steps:

[0127] ;

[0128] when When this term is used, it can be calculated on a sequence of multiple consecutive single-step prediction samples, or the term can be set to zero.

[0129] Let the allowed lower and upper bounds of SOH be respectively and ,in , These represent the lower and upper bounds of the acceptable range of predicted SOH, respectively; boundary constraint loss. Predictions that exceed this physical range are penalized.

[0130] ;

[0131] The total loss is:

[0132] ;

[0133] Total loss Loss due to prediction error Monotonicity constraint loss Smoothness constraint loss and boundary constraint loss Weighted composition; , and These represent the weighting coefficients of the last three constraint losses relative to the prediction error loss. Optionally, , , We set the values ​​to 0.10, 0.01, and 0.10 respectively, and adjusted them using the validation set. For single-step prediction models, prediction error loss and boundary constraint loss can be mainly used.

[0134] The Adam optimizer can be used for training, and the initial learning rate can be set to... The batch size can be 16 to 64, the number of training epochs can be 100 to 300, and the learning rate decays and early stopping is performed based on the validation set loss. The network weights are randomly initialized, and the model parameters with the smallest prediction error on the validation set are saved.

[0135] In a specific parameter example, select , , , , and The trend branch uses 4 residual units, and the fluctuation branch uses 2 residual units; all convolutional layers have 64 output channels, and the random inactivation rate is 0.1%. This parameter combination is only used to illustrate tensor relationships and implementation processes and does not constitute a limitation of the present invention.

[0136] After the above training is completed, save the window evaluation parameters, normalization parameters, VMD parameters, and collaborative prediction model weights for use in the online inference stage.

[0137] After completing offline training and saving all parameters, this embodiment performs online inference according to the following process.

[0138] When running online, the battery management system performs the following processing each time a new valid local charging segment is obtained: adding the new segment to the historical loop cache; retaining the most recent segment. A series of loops that meet data quality requirements are established; the available voltage range covered by these loops is calculated and a local IC sequence is constructed; a window evaluation model is invoked within the available range to determine the target sensitive window; cross-loop health features are extracted and standardized.

[0139] Subsequently, rolling VMD is performed on the current historical window to reconstruct the trend feature matrix and fluctuation feature matrix; future trends are obtained through a dual-branch TCN, a gated fusion module, and a health status prediction module. The predicted SOH value for each cycle. The prediction results can be stored in the battery management system and used for health alerts, charging strategy adjustments, or maintenance plan development.

[0140] It is particularly important to emphasize that "rolling" here refers to the fact that at each prediction time (last cycle number k), the decomposition input only takes the historical window [k-L+1,k], excluding any data from future cycles (such as k+1 and beyond). The construction of training samples also strictly follows the same rule: for a sample with label k+h, the decomposition input only extends to cycle k. This is the key constraint for achieving online VMD deployment in this invention, and it is fundamentally different from existing offline global VMD decomposition methods.

[0141] When the available voltage range of a new cycle changes relative to the previous prediction time, a new candidate window can be generated, and a target sensitive window that is still within the common available range can be selected. If the current common available range is less than the preset minimum width, the output of the new SOH prediction value is temporarily suspended, or the prediction result of the most recent time that met the data quality requirements is continued to be output and marked as having decreased confidence.

[0142] Taking a battery with a nominal capacity of 2Ah as an example, assuming the current cycle number is 300, the most recent 64 effective charging cycles collectively cover a voltage range of 3.72V to 4.05V, with a uniform grid interval of 0.002V. Candidate windows with a width of 0.10V and a sliding step size of 0.02V are generated within this range. After evaluation, two target windows are selected. From each window, the interval charging time, IC area, and main peak position are extracted, resulting in a total of 6-dimensional health characteristics.

[0143] The formation The health feature matrix is ​​decomposed into four modes one-dimensionally and reconstructed according to the maximum center frequency interval. The trend matrix and fluctuation matrix are obtained. After dual-branch TCN and gated fusion, a 64-dimensional fused feature is obtained. Then, the intermediate capacity of the next 5 cycles is output by predicting the hidden state using 32 dimensions and mapped to SOH. The above values ​​are used to illustrate the data size and calculation order. The actual window range and model parameters are determined by the battery system and training data.

[0144] In summary, the core improvement of this invention lies in the following: at each prediction time, only the historical window up to the current cycle is used as the input to VMD, fundamentally cutting off the path for future information to participate in the current decomposition, thus enabling VMD to be deployed online; the training sample construction adopts the same principle of the last cycle and previous data as online inference, avoiding the performance inflation caused by the leakage of future information during offline training; the trend component and fluctuation component obtained from the decomposition are respectively sent to the TCN branches with different time receptive fields, so that long-term degradation patterns and short-term disturbances are modeled at their respective appropriate scales; a gating mechanism is introduced to dynamically adjust the fusion weights of the two types of components according to the current data characteristics, rather than fusion at a fixed ratio, thereby maintaining prediction robustness under different aging stages and operating conditions.

[0145] Therefore, while retaining VMD's ability to effectively separate trends and fluctuations, this invention breaks through its dependence on future data, achieving a leap from offline analysis to online prediction.

[0146] The following combination Figure 2 , Figure 3 , Figure 4 and Figure 6 The beneficial effects of the present invention will be further explained.

[0147] Figure 2This illustrates how incomplete charging data, after processing in steps S1 and S2, forms comparable local IC data. Specifically, using the example of the current last cycle number 300 and a historical window length L=64 as a background, several representative cycles from cycle 237 to 300 are selected to plot local charging voltage-time curves; subsequently, effective segment filtering, unified voltage grid mapping, ampere-hour integration, and center difference are performed on each cycle, and peak-preserving smoothing is applied to obtain the local IC curve. Figure 2 The shaded area in the figure below corresponds to the 3.72V to 4.05V voltage range covered by the aforementioned 64 historical cycles. This figure illustrates that even though the start and end points and durations of charging differ between cycles, a local IC sequence that can be aligned and compared across cycles can still be obtained through a common available voltage range and a unified voltage grid.

[0148] Figure 3 This is used to illustrate the evaluation process of candidate voltage windows and the determination process of target sensitivity windows. Figure 2 Within the common available voltage range shown, candidate windows are generated with a window width of 0.10V and a sliding step size of 0.02V. For each candidate window, candidate health features such as interval charging time, IC area, and peak position are extracted. Then, the comprehensive evaluation value of each candidate window is calculated according to step S4. And combine redundancy penalties to select the target sensitive window. Figure 3 The figure illustrates two target windows: 3.72V to 3.82V and 3.86V to 3.96V. This figure is used to illustrate that the target window is not predetermined by a fixed voltage range or a single IC peak position, but is determined comprehensively based on the correlation between candidate features and SOH, nonlinear dependence, cross-condition stability, and information redundancy.

[0149] Figure 4 This is used to illustrate the rolling VMD and trend-fluctuation reconstruction process. Taking the current cycle 300 as the last cycle, the VMD of step S6 is performed on the one-dimensional standardized health feature sequence composed of the most recent 64 cycles. In the example, it is decomposed into 4 modes and sorted from low to high according to the center frequency. After determining the mode division position based on the maximum adjacent center frequency interval, the low-frequency mode 1 and mode 2 are superimposed to form the trend component, and the high-frequency mode 3 and mode 4 are superimposed to form the fluctuation component. Figure 4 From top to bottom, the first chart compares the original cross-cycle health characteristics with the reconstruction trend; the second chart shows each VMD mode; and the third chart shows the final trend component and fluctuation component. These three charts illustrate that this method can separate slowly changing long-term degradation information from rapidly changing local fluctuation information using only data up to the current cycle, thus providing... Figure 5 The two-branch TCN with different receptive fields at different times provides the corresponding inputs.

[0150] The above Figures 2 to 4This is mainly used to illustrate the data processing relationships and implementation process of the present invention. The values ​​in the figure are exemplary data and should not be taken as conclusions about actual performance.

[0151] Figure 6 This invention is used to illustrate the prediction results of future cycle SOH and its prediction error. Figure 6 In the figure above, the horizontal axis represents the target cycle number to be predicted, the vertical axis represents SOH, the reference curve represents the actual or calibrated SOH of the corresponding cycle, and the prediction curve represents the SOH prediction value output by this invention. When a confidence interval or retirement threshold is set in the figure, the shaded area represents the range of prediction uncertainty, and the horizontal dashed line represents the preset SOH retirement threshold. Figure 6 The figure below shows the prediction error for the corresponding prediction cycle. The prediction error can be expressed as the difference between the predicted SOH and the reference SOH, absolute error, or relative error. This figure is used to illustrate that the model can output the future cycle SOH and corresponding error evaluation results while adhering to the boundaries of current and historical information. The exemplary data in the figure is used to illustrate the prediction output format and evaluation method, and is not intended to limit the actual prediction accuracy of a specific battery.

[0152] The method of this invention can be implemented by a processor executing program instructions in memory, or it can be deployed in a vehicle controller, battery management system controller, energy storage station edge computing device, or cloud server. The training and online phases can be completed on different computing devices; the online device only needs to save window evaluation parameters, normalization parameters, VMD parameters, and parameters of the trained collaborative prediction model.

[0153] The above embodiments illustrate the principles and implementable steps of the present invention. Any equivalent substitutions or conventional transformations made by those skilled in the art, without departing from the concept of the present invention, regarding the candidate window generation method, health feature combination, modal decomposition solution method, number of TCN layers, gating structure, or number of prediction steps, should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting battery health status, characterized in that, Includes the following steps: Acquire incomplete charging data of the battery under test in multiple consecutive historical cycles; and construct a cross-cycle health feature sequence based on the incomplete charging data. At each prediction time, modal decomposition is performed using only the cross-cycle health feature sequence corresponding to that prediction time and the preset number of historical cycles before it to obtain the trend feature sequence characterizing the long-term degradation of the battery and the fluctuation feature sequence characterizing the local dynamic changes of the battery. The trend feature sequence and the fluctuation feature sequence are extracted using a first time-series feature extraction branch with a first time receptive field and a second time-series feature extraction branch with a second time receptive field, respectively, to obtain trend coding features and fluctuation coding features; wherein the first time receptive field is larger than the second time receptive field. Based on the gating weight jointly determined by the trend coding feature and the fluctuation coding feature, the trend coding feature and the fluctuation coding feature are fused, and the predicted health status value of the battery under test for at least one future cycle is output based on the fused features.

2. The battery health status prediction method according to claim 1, characterized in that, The step of constructing a cross-cycle health feature sequence based on the incomplete charging data includes: The incomplete charging data is preprocessed to obtain a local incremental capacity sequence; Multiple candidate voltage windows are generated based on the available voltage range covered by multiple consecutive historical cycles, and candidate health features are extracted from the local incremental capacity sequence and charging data corresponding to each candidate voltage window. The pre-established window evaluation model is invoked to determine at least one target sensitive voltage window based on the degree of correlation between the candidate health features and the battery health status, the stability of the candidate health features across different batteries or operating conditions, and the degree of feature redundancy between different candidate voltage windows. The health features corresponding to the target sensitive voltage window are arranged according to the battery cycle order to construct a cross-cycle health feature time series.

3. The battery health status prediction method according to claim 2, characterized in that, The local incremental capacity sequence is generated by sorting the calculated local incremental capacity values; the calculation steps for the local incremental capacity values ​​are as follows: Interpolate data from different historical cycles to a uniform voltage grid with the same voltage interval; Calculate the first by ampere-hour integration Local charging capacity corresponding to each voltage grid point: ； In the formula, To accumulate from the current local charging segment starting point to Local charging capacity at the location, To standardize the voltage grid point numbering; This represents the local charging capacity corresponding to the previous grid point; and Each is the current number cyclic interpolation to Time and charging current at the point; And calculate the local incremental capacity value based on the first derivative of the capacity with respect to voltage or its discrete approximation: ； In the formula, For the current number The cycle is in the 1st... Local incremental capacity values ​​at a uniform voltage grid point; and These represent the local charging capacity corresponding to the adjacent next grid point and the previous grid point, respectively; and These represent the voltages corresponding to the adjacent next grid point and the previous grid point, respectively.

4. The battery health status prediction method according to claim 3, characterized in that, The window evaluation model is determined according to the following formula: Evaluation values ​​for each candidate voltage window: ； In the formula, For the first The overall evaluation value of each candidate window; , , and These represent the correlation index, nonlinear dependence index, cross-battery or cross-operating condition stability index, and feature redundancy index, respectively; α, β, and γ are the weights of the three positive indices of correlation, nonlinear dependence, and stability, respectively, used to control the relative contribution of the three indices to the comprehensive evaluation value. λ is the redundancy penalty coefficient, used to suppress candidate windows that contain highly repetitive information with the selected window.

5. The battery health status prediction method according to any one of claims 1 to 4, characterized in that, The incomplete charging data includes sampling time, battery voltage, and charging current.

6. The battery health status prediction method according to any one of claims 2 to 4, characterized in that, The candidate health features include one or more of the following: The interval charging time required for the battery to pass through the candidate voltage window; The area under the local incremental capacity curve within the candidate voltage window; Local incremental capacity peak; Battery voltage corresponding to the peak value of local incremental capacity; Peak width of the local incremental capacity peak; The slope of the local incremental capacity curve; Kurtosis of the local incremental capacity curve.

7. The battery health status prediction method according to claim 4, characterized in that, The modality decomposition steps include: Continuous The corresponding historical charging cycles Health features are arranged in cyclical order to construct a cross-cycle health feature matrix: ； In the formula, X represents the cross-cycle health feature matrix; L represents the number of effective charging cycles contained in the historical cycle window; D represents the number of health feature dimensions extracted in each cycle; n represents the cycle position arranged in the order of sequence within the window, n=1,2,…,L; d represents the health feature dimension index, d=1,2,…,D; Let X represent the d-th standardized health feature of the nth cycle. The nth row of matrix X corresponds to the nth historical cycle, and the d-th column corresponds to the d-th health feature. For each dimension of the health feature sequence in the cross-cyclic health feature matrix: The health feature sequence is decomposed into the following values ​​using the variational mode decomposition method: The system has several modal components with different center frequencies, and each modal component is determined by solving the following constrained variational problem. and its corresponding center frequency : ； ； In the above formula, Indexing for health feature dimensions, ; For VMD modal index, ; For the current length is The time position within the historical loop window. ; For the first Health characteristics in location Standardized values ​​at the location; Indicates the first The first health characteristic One modal component; This indicates the center frequency corresponding to the mode; Indicates the total number of VMD modes; This refers to an operator that converts real-valued modes into analytic signals; Indicates relative to the cycle position Discrete difference operator; The imaginary unit; Arranged in order of center frequency from low to high The modal components are rearranged to satisfy: ; Calculate the center frequency interval between adjacent modal components and determine the position corresponding to the maximum center frequency interval as the th . Modal partitioning position of the health feature sequence: In the formula, This represents the modality partitioning position of the d-th dimension of health features. The number of trend modes is limited to between the minimum and maximum trend modes preset during the training phase; Δ The center frequency interval between the m-th and (m+1)-th adjacent modes of the d-th health feature; The modal components with center frequencies located before and at the modal division position are superimposed and reconstructed to obtain the first modal component. Trend feature sequence corresponding to health characteristics: The modal components whose center frequencies are located after the modal division position are superimposed and reconstructed to obtain the first modal component. Fluctuation feature sequence corresponding to the health characteristics: ;in, This represents the trend component of the d-th dimension of health characteristics at position n. This represents the fluctuation component of the d-th dimension health characteristic at position n; All The trend feature sequences corresponding to the health characteristics are combined according to the feature dimensions to obtain the trend feature matrix: ; and all The fluctuation feature sequences corresponding to the health characteristics are combined according to the feature dimensions to obtain the fluctuation feature matrix: .

8. The battery health status prediction method according to claim 7, characterized in that, The step of extracting features from the trend feature sequence and the fluctuation feature sequence includes: The trend feature matrix Input the first time-series feature extraction branch, and extract the fluctuation feature matrix. Input the second temporal feature extraction branch, where, The length of the historical loop window. For health characteristics dimension; The first and second time-series feature extraction branches are uniformly represented as the first... Temporal feature extraction branch, and the first The input representation of the temporal feature extraction branch is as follows: ； in, This represents the initial input feature matrix for the b-th temporal feature extraction branch; The first The temporal feature extraction branch includes For the sequentially connected temporal convolutional residual units, for the , Each temporal convolutional residual unit is transformed using dilated causal convolution to obtain: ； ； In the formula, n represents the time position within the historical loop window; q represents the sampling position inside the convolution kernel; This indicates the length of the convolution kernel used in the b-th branch; This represents the expansion coefficient of the l-th residual unit in the b-th branch; This represents the input characteristics of the residual unit; and These represent the outputs of the first and second layer dilated causal convolutions of the residual unit, respectively. and These represent the weight matrices of the two convolutional layers at convolution position q, respectively. and These represent the bias vectors corresponding to the two convolutional layers, respectively. Represents a nonlinear activation function; Set the corresponding input to zero or fill it with the boundary values ​​of the health feature sequence, so that the first The convolution output corresponding to the nth historical loop is only determined by the nth historical loop. The health characteristics of each historical cycle and its preceding periods are determined; The output of the second dilated causal convolution is combined with the first... The inputs of each temporal convolutional residual unit are concatenated to obtain: In the formula, This represents the residual mapping matrix of the l-th residual unit in the b-th branch; when the number of input and output channels is the same, Using unit mapping, when the number of channels is different, 1×1 convolution or equivalent linear mapping is used to match the channel dimension so that the feature dimension at both ends of the residual connection remains consistent. This represents the output of the l-th residual unit at position n after residual connection and nonlinear activation; The first step is determined based on the kernel length and dilation coefficient of each temporal convolution residual unit. Temporal receptive field of the temporal feature extraction branch: ； In the formula, This represents the length of the temporal receptive field of the b-th temporal feature extraction branch; This indicates the length of the convolution kernel used in the b-th branch; This indicates the number of temporal convolutional residual units contained in this branch; l is the residual unit layer number; the coefficient 2 indicates that each residual unit contains two layers with the same dilation coefficient. causal convolution; This represents the expansion coefficient of the l-th residual unit in the b-th branch; The final outputs of the first and second time-series feature extraction branches are time-series aggregated to obtain trend time-series features and fluctuation time-series features: ； ； In the formula, (·)and (·) represent the temporal aggregation and linear mapping operators of the first and second temporal feature extraction branches, respectively; and These represent the encoded features obtained from the trend branch and the fluctuation branch, respectively. and These represent the number of temporal convolutional residual units contained in the first temporal feature extraction branch and the second temporal feature extraction branch, respectively. The first time-series feature extraction branch is used to extract long-term degradation changes, and the second time-series feature extraction branch is used to extract short-term fluctuations and local degradation changes.

9. The battery health status prediction method according to claim 8, characterized in that, The gating weights are calculated as follows: ； In the formula, This indicates that the trend is encoded as a feature. With fluctuation coding features Perform vector concatenation; the resulting dimension is 2C. For the gated mapping weight matrix, The gating bias vector; sigmoid(·) represents the sigmoid function; g For the gating weight vector, each element of g is between 0 and 1; The fusion features are obtained in the following manner: ； In the formula, ∈ Indicates fused coding features; ⊙ indicates element-wise multiplication; 1-g represents the complementary gating weights corresponding to the fluctuation branches; when the value of g on a certain feature dimension is close to 1, that dimension focuses more on trend encoding features; when the value is close to 0, that dimension focuses more on fluctuation encoding features.

10. The battery health status prediction method according to claim 9, characterized in that, The step of outputting a predicted health status value of the battery under test for at least one future cycle based on the fused features includes: Fusing temporal features The health status prediction module receives input and obtains the predicted latent state through nonlinear feature mapping. ； in, To integrate temporal features, This is the first prediction mapping matrix. This is the first prediction bias vector. It is a non-linear activation function. To predict the hidden state; when the dimension of the fused temporal features is C, the dimension of the predicted hidden state is... hour, , , ; The predicted hidden state is mapped to the output to obtain the future. Intermediate values ​​for capacity prediction corresponding to each cycle: ； ； in, This is the second prediction mapping matrix. This is the second prediction bias vector. This is the intermediate vector for capacity prediction. To preset the number of prediction steps, For the first The intermediate quantity of capacity prediction for each prediction step, h=1,2,…,H; Perform a bounded mapping on the intermediate quantities of the capacity prediction to obtain the future... Predicted sequence of available capacity for each cycle: ； ； in, represents the last loop number in the current historical loop window; h represents the prediction step index relative to the last loop k, h=1,2,…,H, so k+h represents the h-th loop to be predicted after the current last loop; Indicates the first Predicted available capacity for each cycle, and These represent the preset lower limit and upper limit of capacity, respectively; Calculate the predicted health status for each future cycle based on the predicted available capacity and the reference capacity: ； in, This represents the predicted health status value for the (k+h)th cycle in the future. The nominal capacity of the battery under test.

Images