Method for health management of energy storage devices for energy saving elevator systems and system therefor

Abstract

The application discloses a health management method and system for an energy storage device in an energy-saving elevator system, and the method comprises the following steps: applying a controlled pulse excitation to the energy storage device for a preset time length; collecting response time sequence data of each battery monomer at a fixed sampling frequency during the controlled pulse excitation and in a relaxation stage after the excitation ends; extracting multi-dimensional health features for representing the battery health state from the response time sequence data; constructing a group health representation space serving as a group health state reference benchmark by using a multivariate statistical analysis method based on the multi-dimensional health features of all battery monomers in the same battery pack; calculating the deviation degree of the health features of each battery monomer in the group health representation space, and identifying abnormal battery monomers according to the deviation degree; and calculating the relative health state parameter of each battery monomer based on the deviation degree, and combining the health state parameters in the historical detection period to predict the health trend of the battery monomer.

Description

Technical Field

[0001] This invention relates to the field of elevator system technology, and in particular to a health management method and system for energy storage devices in energy-saving elevator systems. Background Technology

[0002] With increasingly stringent energy conservation and environmental protection requirements, modern elevator systems have widely incorporated energy feedback and storage devices. These systems typically recover regenerative energy during elevator braking, light-load ascent, or heavy-load descent, storing it in energy storage devices (such as battery packs composed of multiple individual cells). During energy-consuming phases such as elevator acceleration or heavy-load ascent, the stored energy is fed back to the drive system, effectively reducing overall energy consumption. However, the actual operating conditions of the energy storage devices in these energy-saving elevator systems differ significantly from traditional constant-power, constant-current applications. Specifically, the charging and discharging behavior exhibits short-duration, medium-to-high rate, non-periodic pulse characteristics, rather than a smooth, continuous charging and discharging process. Furthermore, influenced by passenger behavior and usage time, the charging and discharging conditions exhibit a complex combination of high randomness and statistical periodicity over time.

[0003] Under the aforementioned complex operating conditions, energy storage devices face the following prominent problems: First, the accuracy of individual battery cell State of Health (SOH) assessment is severely limited. Most existing assessment methods are calibrated based on standard cycle conditions or laboratory testing conditions provided by manufacturers, making it difficult to accurately reflect the impact of factors such as high-frequency pulses, drastic polarization changes, and temperature fluctuations on battery degradation mechanisms in the actual operating environment of the elevator. This leads to a systematic deviation between the actual SOH assessment results and the true state, affecting the safety, controllability, and availability of the energy storage device. Second, the consistency of individual cells within the battery pack is a prominent issue. Due to manufacturing variations, uneven temperature distribution, and differences in aging rates caused by different installation locations, the voltage response, internal resistance characteristics, and usable capacity of individual cells within the same battery pack gradually diverge, forming a significant "weakest link" effect—the overall performance of the battery pack is limited by the worst-performing cell. This consistency degradation intensifies over time, but existing monitoring methods struggle to effectively quantify it in its early stages.

[0004] Furthermore, latent anomalies in individual cells are difficult to detect in a timely manner. In actual operation, individual battery cells may exhibit low usable capacity, decreased charge acceptance, or abnormal static voltage due to factors such as micro-leakage, internal short-circuit tendency, volt-ampere characteristic drift, or abnormal capacity decay. These anomalies typically do not trigger protection thresholds in the short term, but they can continuously drag down the overall energy storage performance and even lead to safety risks under certain conditions. Traditional monitoring methods based on single thresholds or simple statistics are insufficient to effectively identify such early anomalies under complex pulsed operating conditions. In addition, elevator energy storage devices operate in field environments, making it difficult to construct standardized health status comparison datasets or obtain sufficient labeled samples. Methods based on supervised learning or large-scale model training face high application barriers in such systems due to their reliance on large amounts of high-quality labeled data.

[0005] In summary, there is an urgent need for a method that can accurately assess the health status of individual battery cells, identify anomalies, and predict health trends under complex pulse conditions in energy-saving elevator energy storage devices, without relying on standard control samples or manufacturer SOH calibration curves, in order to support the safe operation and predictive maintenance decisions of the system. Summary of the Invention

[0006] The main objective of this invention is to provide a health management method for an energy storage device in an energy-saving elevator system. The energy storage device includes multiple battery cells. The health management method includes: When the elevator system is in a non-operating period and meets the preset operating conditions, a controlled pulse excitation of preset duration is applied to the energy storage device. During the controlled pulse excitation process and the relaxation phase after the excitation ends, the response time series data of each battery cell are collected at a fixed sampling frequency. Extract multidimensional health features from response time series data to characterize battery health status; Based on the multidimensional health characteristics of all individual cells within the same battery pack, a group health characterization space is constructed using multivariate statistical analysis methods to serve as a reference benchmark for the group's health status. Calculate the deviation of the health characteristics of each battery cell in the population health characterization space, and identify abnormal battery cells based on the deviation. The relative health status parameters of each battery cell are calculated based on the deviation, and the health status parameters of the battery cells are combined with those of historical testing cycles to predict the health trend of the battery cells.

[0007] This invention also proposes a health management system for energy storage devices in energy-saving elevator systems, comprising: A controlled excitation unit is used to apply a controlled pulse excitation of a preset duration to the energy storage device; The data acquisition unit is used to acquire the response time series data of each battery cell at a fixed sampling frequency during the controlled pulse excitation process and the relaxation phase after the excitation ends. The data preprocessing unit is used to preprocess the response time series data to remove outliers and standardize and align the data. The feature extraction and modeling unit is used to extract multidimensional health features from the preprocessed response time series data to characterize the battery health status, and to construct a group health representation space as a reference benchmark for the group health status based on the multidimensional health features of all battery cells in the same battery pack. An anomaly identification unit is used to calculate the deviation of the health characteristics of each battery cell in the group health characterization space, and to identify abnormal battery cells based on the deviation. The health assessment unit is used to calculate the relative health status parameters of each battery cell based on the deviation, and to predict the health trend of the battery cells by combining the health status parameters of historical testing cycles. Each unit is used to implement a health management method for energy storage devices in an energy-saving elevator system. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram illustrating the steps of an embodiment of the health management method for an energy storage device in an energy-saving elevator system according to the present invention; Figure 2 This is a schematic diagram illustrating the steps of another embodiment of the health management method for an energy storage device in an energy-saving elevator system according to the present invention; Figure 3 This is a schematic diagram illustrating the steps of another embodiment of the health management method for an energy storage device in an energy-saving elevator system according to the present invention.

[0010] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] 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 a part of the embodiments of the present invention, and not all of the 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.

[0012] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0013] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0014] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0015] This invention proposes a health management method for energy storage devices in energy-saving elevator systems. In the first embodiment, as follows: Figure 1 As shown, health management methods include: When the elevator system is in a non-operating period and meets the preset operating conditions, a controlled pulse excitation of preset duration is applied to the energy storage device. During the controlled pulse excitation process and the relaxation phase after the excitation ends, the response time series data of each battery cell are collected at a fixed sampling frequency. Extract multidimensional health features from response time series data to characterize battery health status; Based on the multidimensional health characteristics of all individual cells within the same battery pack, a group health characterization space is constructed using multivariate statistical analysis methods to serve as a reference benchmark for the group's health status. Calculate the deviation of the health characteristics of each battery cell in the population health characterization space, and identify abnormal battery cells based on the deviation. The relative health status parameters of each battery cell are calculated based on the deviation, and the health status parameters of the battery cells are combined with those of historical testing cycles to predict the health trend of the battery cells.

[0016] It should be explained that the energy storage device includes multiple battery cells; these battery cells are connected in series or in series-parallel configuration to form a battery box or battery pack, which is used to recover regenerative energy when the elevator system brakes, moves upward under light load or moves downward under heavy load, and release energy when the elevator accelerates or moves upward under heavy load to reduce the overall energy consumption of the machine.

[0017] Because the operating conditions of elevators are highly random and pulse-like, and the on-site environment makes it difficult to provide standardized health status control samples, the method of this invention aims to achieve accurate assessment, anomaly identification, and trend prediction of the health status of individual battery cells by applying active excitation to the energy storage device and analyzing the group self-reference characteristics, without relying on manufacturer calibration curves and external control datasets.

[0018] In this embodiment, a controlled pulse excitation of a preset duration is first applied to the energy storage device when the elevator system is in a non-operating period and meets preset operating conditions. The "non-operating period" refers to the time interval during which the elevator system does not perform passenger or cargo transportation tasks, such as at night, holidays, or system idle periods. Detection during this period avoids interference with normal elevator operation caused by the excitation process and also reduces the contamination of battery response data by external load fluctuations. The "preset operating conditions" include at least the state of charge (SOC) of the energy storage device and the battery temperature. Specifically, to ensure data comparability between different detection cycles, the SOC of the energy storage device is typically required to reach or exceed a preset threshold (e.g., 90%) before applying the pulse excitation. This is because the battery voltage response at low SOC is significantly affected by the state of charge, interfering with the accuracy of health feature extraction. Simultaneously, the battery temperature should be within a preset allowable range (e.g., 20°C to 35°C), as extreme temperatures significantly alter the battery's internal resistance and polarization characteristics, causing the response data to fail to accurately reflect the battery's intrinsic health state. If the SOC does not meet the requirements, the energy storage device will be recharged first; if the temperature exceeds the allowable range, the detection will be delayed until the temperature returns to the appropriate range.

[0019] "Controlled pulse excitation" refers to a constant-power charging pulse applied for a short period of time with precisely controlled parameters. This embodiment uses a constant-power mode instead of a constant-current mode because in actual elevator systems, the energy storage converter typically operates according to power commands, and a constant-power pulse better reflects the thermal effects and polarization characteristics under actual operating conditions. The preset pulse duration is, for example, 10 seconds, and the excitation power is typically set to the power value corresponding to a rate of 0.2C to 0.5C (C being the battery's rated capacity). To avoid random errors in a single pulse, it can be applied multiple times (e.g., 3 times) within a single detection cycle, with a sufficient rest interval (e.g., 600 seconds) between adjacent pulses to allow for sufficient relaxation of battery polarization. The working principle of this controlled pulse excitation is to stimulate the ohmic polarization, electrochemical polarization, and concentration polarization processes inside the battery through short-term constant-power charging without significantly changing the overall SOC of the energy storage device. Since battery cells in different health states have different internal resistance, polarization capacitance, and diffusion impedance, the voltage response waveforms of each cell will show distinguishable changes under the same pulse excitation, thus providing information-rich excitation response data for subsequent health feature extraction.

[0020] During the controlled pulse excitation process and the relaxation phase after excitation, this method collects the terminal voltage time series data of each battery cell at a fixed sampling frequency (e.g., once per second). The "relaxation phase" refers to the period after the pulse excitation stops, during which the internal polarization of the battery gradually dissipates and the terminal voltage slowly decreases. This phase typically lasts about 30 seconds, and its voltage decay curve contains key health information such as the total accumulated polarization and the polarization dissipation time constant. Therefore, the collected "response time series data" should cover at least two consecutive time intervals: the pulse charging phase and the relaxation phase. For ease of subsequent analysis, the sampled data needs to be stored in association with information such as the battery cell number, pulse sequence number, and timestamp. If the local control system has limited computing power, the raw data can be encapsulated and uploaded to a cloud service platform or remote analysis server via wired or wireless communication modules for subsequent complex calculations in the cloud. In case of communication failure, the data is temporarily stored locally and retransmitted after the network is restored to ensure data integrity.

[0021] After obtaining the raw voltage time series data, this method further extracts multidimensional health features to characterize the battery's health status from these data. Before feature extraction, the raw data typically needs to be preprocessed to eliminate interference such as measurement noise, communication anomalies, and sampling asynchrony. Preprocessing steps include: integrity verification (checking the number of sampling points, timestamp continuity, and pulse number consistency), identification and correction of abnormal sampling points (e.g., using the median of a sliding window to detect amplitude abrupt changes and replacing them with linear interpolation or the median), time alignment (using the start time of each pulse as the time zero point), stage segmentation (dividing the sequence into charging stage subsequences and relaxation stage subsequences), and baseline correction (shifting the sequence based on the pulse start voltage to reflect relative voltage changes rather than absolute voltage values). After the above preprocessing, multidimensional health features are extracted for different stages. Specifically, the extracted features include, but are not limited to, the following categories: Transient response characteristics during the charging phase: (1) Initial voltage transition; reflects the magnitude of the battery's equivalent ohmic internal resistance, and is calculated using the following formula: ,in The most recent sampling point before charging began. This is the first sampling point after charging begins. This feature is used to characterize the equivalent ohmic internal resistance of the battery.

[0022] (2) Initial voltage rise rate. Obtained through linear fitting. During the initial time window of the charging phase. Within this process, a linear fit is performed on the voltage-time series: , The slope is solved using the least squares method. Defined as

[0023] This characteristic reflects the battery's ability to rapidly polarize under high-rate excitation.

[0024] (3) Maximum slope and slope dispersion. Calculate the first-order difference of the voltage during the charging phase. And extract the maximum slope and slope standard deviation It is used to reflect the stability of the polarization establishment process.

[0025] (4) Voltage curvature characteristics characterize the degree of concentration polarization. Calculate the second-order voltage difference: The mean and maximum value can be extracted as features.

[0026] Characteristics of polarization release during relaxation phase: (5) The relaxation voltage decay amplitude reflects the total cumulative polarization: .

[0027] (6) Relaxation time constant fitting. A multi-exponential model is used for the voltage time series during the relaxation phase: The time constant is estimated using the nonlinear least squares method. Amplitude coefficient The time constant is used to distinguish between fast polarization and slow diffusion processes.

[0028] (7) Relaxation fitting residual energy. Calculate the model fitting residual. And calculate the residual energy: It is used to identify abnormal relaxation behavior.

[0029] Integral and energy-based characteristics: (8) Integral voltage change during charging phase: It is used for the reaction polarization accumulation process.

[0030] (9) Release integrals during the relaxation phase: .

[0031] Repetitive pulse stability characteristics: Within one detection cycle, K pulse detection results are obtained for the same battery cell (K is the number of pulse repetitions, typically 3).

[0032] For any feature f, calculate: Mean: ; Standard deviation: ; Coefficient of variation: .

[0033] Mean, standard deviation, and coefficient of variation are used to assess the consistency of battery response.

[0034] Relative deviation characteristics of the group: Calculate the group mean for similar characteristics of N individual cells within the same battery pack. Then calculate the relative deviation of the individual: It is used to characterize consistency degradation and latent anomalies.

[0035] The aforementioned features are combined in a preset order to form a health feature vector for each battery cell. This feature vector characterizes the battery's health status from multiple physical dimensions (resistance, polarization, diffusion, and uniformity), overcoming the limitation that a single indicator cannot fully reflect health degradation under complex operating conditions.

[0036] After extracting the multidimensional health features of all individual battery cells, this method constructs a group health representation space based on the multidimensional health features of all individual battery cells within the same battery pack using multivariate statistical analysis, which serves as a reference benchmark for the group's health status. In this embodiment, Principal Component Analysis (PCA) is preferably used as the "multivariate statistical analysis method." The specific construction process is as follows: The health feature vectors of all individual battery cells are arranged in rows to form a feature matrix, where the number of rows represents the number of battery cells and the number of columns represents the feature dimensions. Each column of this matrix is ​​standardized (subtracting the mean and dividing by the standard deviation) to eliminate the influence of dimensional differences between different features. Then, the covariance matrix of the standardized feature matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​are sorted from largest to smallest, and the top K principal components (eigenvectors) with a cumulative contribution rate reaching a preset threshold (e.g., 85% or 90%) are selected. The subspace spanned by these K principal components is called the "main health subspace," which reflects the common health pattern of most battery cells within the current testing period, representing the normal and mainstream health state of the group. The subspace spanned by the remaining feature vectors is called the "singular deviation subspace," which mainly contains information on individual differences, abnormal behaviors, and consistency degradation. The "group health representation space" is a high-dimensional mathematical space jointly constructed by the main health subspace and the singular deviation subspace. In this space, the health feature vector of each battery cell can be projected onto the main health subspace (to obtain its coordinates in the group's common pattern) and the singular deviation subspace (to obtain its component that deviates from the group's common pattern). Since this method does not rely on any external standard samples or manufacturer calibration curves, but instead uses the current data of all cells within the same battery pack as a reference, it is called the "group self-reference" method. This method is particularly suitable for field environments with no control dataset and complex operating conditions, and can dynamically adapt to the gradual aging of the entire battery pack.

[0037] After constructing the population health representation space, this method calculates the deviation of the health characteristics of each battery cell within this space and identifies abnormal battery cells based on the deviation. "Deviation" is a quantitative indicator used to measure the degree of difference between the health behavior of a battery cell and the mainstream health pattern of the population. In this embodiment, deviation is primarily characterized by the SquaredPredictionError (SPE) index. Specifically, the standardized feature vector of a battery cell is projected onto the squared deviation subspace to obtain a projection vector. Then, the squared Euclidean norm of this projection vector is calculated, which is the squared predictionerror index. The larger this index, the more components of the health characteristics of the cell cannot be explained by the main health subspace of the population, meaning the higher the degree to which the cell deviates from the normal pattern of the population. In addition to the squared predictionerror index, the distance from the projected coordinates of the cell in the main health subspace to the population health center (e.g., the origin or the projection point of the population mean) can also be calculated, called the dominant modal distance. This distance reflects the difference between the overall health level of the cell and the average level of the population. In actual anomaly detection, singular energy indicators can be used alone, or a weighted combination of singular energy indicators and dominant mode distance can be obtained to obtain a comprehensive anomaly score. The anomaly threshold is determined adaptively, for example, by taking the 95th percentile of all individual singular energy indicators as the threshold. When the singular energy indicator of a single cell exceeds this threshold, or the comprehensive anomaly score is greater than 1, the cell is determined to be an abnormal battery cell. To reduce false alarms, a continuous confirmation mechanism can be introduced: only when the same cell is determined to be abnormal in multiple consecutive detection cycles (e.g., three consecutive cycles) is it confirmed as a stable anomaly and a maintenance reminder is triggered. Furthermore, by analyzing the contribution of the projection components of abnormal cells in the singular deviation subspace, the health characteristic category of the dominant anomaly (such as internal resistance anomaly, polarization anomaly, relaxation anomaly, or consistency degradation) can be identified, thus providing more targeted fault cause references for on-site maintenance.

[0038] Finally, this method calculates the relative health state parameters of each battery cell based on the deviation and combines them with the health state parameters from historical testing cycles to predict the health trend of the battery cells. It is important to emphasize that the health state parameters in this method are not the absolute SOH (State of Health) relative to the manufacturer's rated capacity in the traditional sense, but rather a "relative health state parameter," typically defined as [0,1], where 1 indicates that the cell has the best health level in the current group, and 0 indicates severe degradation or failure. This relative definition avoids the evaluation difficulties caused by the lack of manufacturer calibration curves or the inability to measure absolute capacity, while accurately reflecting the relative performance differences of batteries under field operating conditions. In specific calculations, the principal mode distance can be normalized and mapped to the principal health component (e.g., ...). The exotic energy index is mapped to an exotic health component (e.g., through an exponential decay function) using an exponential decay function. The model introduces a stability correction factor based on the number of consecutive abnormal cycles. Then, the main health component and the singular health component are weighted and summed to obtain the relative health state parameters of each individual cell. For the battery pack as a whole, a "barrel effect" weighted model can be used, taking the minimum value of all individual cell health state parameters and considering a variance penalty term, or a low quantile (such as the 5th percentile) can be used as the battery pack health state parameter to reflect the limitation of the weakest link on the overall performance.

[0039] After obtaining the health status parameters for each monitoring cycle, a time series is constructed. By performing linear regression or nonlinear fitting (e.g., an exponential decay model) on this time series, the degradation rate of the health status parameters can be estimated. Based on the degradation rate and a preset failure threshold (e.g., 0.6 to 0.8), the time required to reach the failure threshold in the future, i.e., the remaining usable life, can be predicted. Simultaneously, the prediction interval can be calculated to quantify uncertainty. The above prediction results can be used to guide maintenance decisions: for example, generating a replacement recommendation when the predicted remaining life is below a certain threshold; triggering an early warning when the degradation rate accelerates.

[0040] Through the above steps, the method of this invention, without relying on standard control samples or manufacturer SOH calibration curves, utilizes only short-term controlled pulse excitation during the off-peak hours of the elevator system and multivariate statistical analysis based on group self-reference to accurately assess the health status of individual battery cells in an energy storage device, identify anomalies, and predict health trends. This method fully considers the high-random-pulse charging and discharging characteristics of energy-saving elevator systems, amplifies the response differences between batteries in different health states through active excitation, and eliminates dependence on absolute calibration values ​​through group self-reference. It has advantages such as strong engineering applicability, good interpretability, and online update capability.

[0041] In one embodiment, the population health representation space includes a main health subspace and a singular deviation subspace; like Figure 2 As shown, the population health representation space, constructed using multivariate statistical analysis methods as a benchmark for population health status, includes: The multidimensional health characteristics of all battery cells are used to construct a feature matrix and then standardized. Calculate the covariance matrix based on the standardized feature matrix; Eigenvalue decomposition is performed on the covariance matrix, and principal component analysis is used to extract principal components that reflect the common health patterns of the group. The main health subspace and the singular deviation subspace are determined. The main health subspace is determined by the feature vectors corresponding to the top K principal components whose cumulative contribution rate reaches a preset threshold, and is used to characterize the common health pattern of the group. The singular deviation subspace is determined by the remaining feature vectors and is used to characterize the deviation of individual battery cells from the common pattern of the group. K is a positive integer. Calculating the deviation of the health characteristics of each individual battery cell from the population health characterization space includes: Calculate the projection vector of each battery cell in the singular deviation subspace, and calculate the singular energy index based on the projection vector. The singular energy index is used to quantitatively characterize the deviation.

[0042] It should be noted that traditional battery health assessment methods typically rely on external standard samples or manufacturer-defined health status benchmarks, such as calculating State of Health (SOH) by comparing the current battery capacity with the factory-rated capacity. However, in energy-saving elevator systems, on-site conditions are complex, charging and discharging behaviors are highly random, and it is impossible to obtain a standard health dataset of batteries of the same model under the same operating conditions. Therefore, this invention proposes a group self-reference approach: treating all battery cells within the same battery pack (or battery box) as a group, and dynamically constructing a mathematical space reflecting the mainstream health pattern of the group by analyzing the statistical structure among the health characteristics of each cell within this group, using this space as a reference benchmark. Under this benchmark, normal battery cells should exhibit similar statistical behavior, while abnormal cells will show significant deviations. This method does not rely on any external control data, is entirely driven by the measured data of the current testing cycle, and can adaptively track the gradual aging of the entire battery pack.

[0043] Specifically, the process of constructing the population health representation space is as follows.

[0044] I. Constructing the Feature Matrix and Standardization Process In the same battery box or parallel battery pack, assume there are N battery cells. For each battery cell i, after the completed controlled pulse excitation, data acquisition, and feature extraction steps, a P-dimensional health feature vector is obtained: , Where P represents the number of dimensions of the health features, and the specific types of features may include, but are not limited to: initial voltage transition, initial voltage rise rate, maximum slope, and voltage curvature characteristics during the charging phase; voltage decay amplitude, relaxation time constant, and fitted residual energy during the relaxation phase; as well as integral features, repetitive pulse stability characteristics, and population relative deviation characteristics. These features characterize the battery's health state from different physical aspects (ohmic internal resistance, polarization capability, diffusion characteristics, and consistency).

[0045] Arrange the feature vectors of all N individuals in rows to form an N x P group feature matrix X: Because different features have different physical dimensions and numerical ranges (e.g., internal resistance may be on the order of milliohms, while time constant may be on the order of seconds), directly using the raw values ​​for subsequent statistical calculations will lead to features with larger dimensions dominating the analysis results, while features with smaller dimensions but important physical significance will be overlooked. Therefore, it is necessary to standardize each column of the feature matrix (i.e., each feature) to transform it into a dimensionless standardized variable with a mean of 0 and a standard deviation of 1. Specifically, for the j-th feature (j=1,...,P), the mean of all N individuals in the population for that feature is calculated. and standard deviation : , . The elements of the standardized feature matrix Z are: In the standardized matrix Z, each column has a mean of 0 and a variance of 1. This process eliminates differences in dimensions and scales, ensuring that different features have equal weight in subsequent principal component analysis, and making the covariance matrix directly correspond to the correlation coefficient matrix, which is easier to interpret.

[0046] II. Covariance Matrix Calculation and Eigenvalue Decomposition Based on the standardized feature matrix Z, the covariance matrix C of the population is calculated. The covariance matrix is ​​a P×P symmetric matrix, where the element in the j-th row and k-th column represents the degree of linear correlation between the j-th feature and the k-th feature. The calculation formula is: , N-1 is used as the denominator here to obtain an unbiased estimate. The covariance matrix C contains the correlation structure information between all features. When two features are highly positively correlated, their corresponding covariance values ​​in C are large; when two features are negatively correlated, the covariance value is negative; when the features are independent, the covariance approaches 0.

[0047] Perform eigenvalue decomposition on the covariance matrix C. Eigenvalue decomposition is a standard decomposition method in linear algebra. For a real symmetric matrix C, there exists an orthogonal matrix U and a diagonal matrix Λ such that: , in, ; For a diagonal matrix, the elements on the diagonal are... These are called eigenvalues ​​and are arranged in descending order. U=[ , ,..., ] is a P×P orthogonal matrix, each column It is an eigenvector of length P, corresponding to the eigenvalues. The eigenvectors are orthogonal to each other (i.e., uncorrelated), and each eigenvector has a magnitude of 1.

[0048] The physical meaning of eigenvalue decomposition is that it decomposes the total variance (i.e., the total variation of all features) in the original P-dimensional feature space into variance components along P orthogonal directions. Eigenvalues The size represents the data in the corresponding feature vector. The magnitude of variance in a direction. A larger variance in a direction indicates a higher degree of dispersion of the data in that direction, meaning that the direction contains more information about group differences.

[0049] III. Definitions of Main Healthy Subspace and Singular Deviation Subspace Based on the magnitude of the eigenvalues, eigenvectors can be sorted and selected. The first few eigenvectors correspond to larger eigenvalues, and the subspace they span contains the main directions of variation in the data, representing the common health patterns of most battery cells in the population. Specifically, the top K eigenvectors are selected such that their cumulative contribution rate reaches a preset threshold η (typically 80% to 95%). The formula for calculating the cumulative contribution rate is: The cumulative contribution rate measures the proportion of total variance explained by the top K principal components. The choice of the threshold η depends on the application scenario: if more common information about the population is desired, a higher η (e.g., 95%) can be chosen, but in this case, M may be larger, resulting in a higher subspace dimension; if more noise and individual differences are desired, a lower η (e.g., 80%) can be chosen, but a small amount of normal variation information may be lost. In this embodiment of the invention, a balance value of 90% is recommended for η.

[0050] The subspace spanned by these first K eigenvectors is called the principal healthy subspace, denoted as: , The main health subspace characterizes the dominant health feature patterns shared by most battery cells within the current testing period. For example, if the internal resistance and polarization characteristics of most battery cells exhibit a linear correlation, this relationship will be captured by a certain direction in the main health subspace. The normalized feature vector of a normal battery cell, when projected onto the main health subspace, will fall within or near that subspace; while the projection of an abnormal cell into the subspace may deviate from the group center.

[0051] The subspace spanned by the remaining PK eigenvectors (corresponding to smaller eigenvalues) is called the singular deviation subspace, denoted as: , The singular deviation subspace mainly contains two types of information: first, individual random fluctuations caused by measurement noise, small random perturbations, etc.; and second, special variations that are inconsistent with the common pattern of the group, resulting from abnormal health states of certain battery cells (such as a sharp increase in internal resistance, abnormal polarization, etc.). Therefore, the magnitude of the projection component of a battery cell in the singular deviation subspace can directly reflect the degree of deviation of its health behavior from the mainstream pattern of the group.

[0052] IV. Calculation of Deviation: Singular Energy Index After defining the main healthy subspace and the singular deviation subspace Then, the deviation of each battery cell in the population health characterization space can be calculated. There are various ways to quantify the deviation; this embodiment uses the Singular Prediction Error (SPE) as the core deviation metric.

[0053] For the i-th battery cell, its standardized feature vector is: (A row vector of length P). First, [the following is a list of steps / methods]. Projected onto singular deviated subspace Above. The projection matrix is ​​composed of the remaining eigenvectors: , is a P×(PK) matrix. Projection vector. for: (here Consider it as a 1×P row vector. Given a P×(PK) matrix, the result is... (A 1×(PK) row vector).

[0054] Projection vector Each component represents the coordinates of the individual entity in the corresponding singular direction. Then, a singular energy index is defined. Projection vector The square of the Euclidean norm, i.e., the sum of the squares of its components: . The physical meaning is as follows: If the health characteristics of a single battery cell perfectly conform to the common pattern of the population (i.e., all its variations can be explained by the main health subspace), then its projected component in the singular deviation subspace should be close to zero. It's very small. Conversely, if the individual exhibits any health characteristic changes inconsistent with the mainstream pattern of the population (such as abnormally high internal resistance, shift in relaxation time constant, etc.), these abnormal variations cannot be contained in the main healthy subspace and will leak into the singular deviation subspace, leading to an increase in the magnitude of the projected components. Significantly increased. Therefore, It is an indicator that is highly sensitive to anomalies and can detect early, subtle, latent degradation.

[0055] In addition, as a supplement or auxiliary method, the projected coordinates of an individual in the main health subspace and its distance to the population health center (main modal distance) can also be calculated, but the deviation is explicitly defined as being quantitatively characterized by the singular energy index. In the technical solution, deviation refers to... .

[0056] The population health representation space constructed using the above methods has the following beneficial effects: First, self-referentiality. The entire construction process is based entirely on the current testing cycle data of all cells within the same battery pack, without relying on any external standard samples, manufacturer calibration curves, or historical data. Therefore, this method can automatically adapt to the reference drift caused by aging, environmental changes, etc., of the entire battery pack, avoiding the systematic deviations caused by reference mismatch in traditional methods.

[0057] Second, dimensionality reduction and denoising. Principal component analysis decomposes the original high-dimensional feature space into a low-dimensional main health subspace and a low-energy singular deviation subspace. The main health subspace extracts the most representative health patterns in the population, filtering out random noise and non-critical individual fluctuations; the singular deviation subspace, on the other hand, retains anomalous information. This decomposition makes anomaly detection more sensitive and robust.

[0058] Third, interpretability. Exotic energy indicators. It has a clear statistical significance: it equals the sum of the variances of the individual eigenvectors in directions orthogonal to the main healthy subspace. When the SPE exceeds an adaptive threshold determined by the population distribution, it can be considered an anomaly. Furthermore, by analyzing the projection vectors... The magnitude of each component can be used to trace which characteristic directions caused the deviation, thus providing a basis for tracing the cause of the anomaly.

[0059] Fourth, online adaptability. After each detection cycle, the feature matrix can be reconstructed and the covariance matrix updated based on the latest collected data, achieving dynamic updates to the population health representation space. Since battery aging in elevator energy storage devices is a slow process, a sliding time window weighted update method (such as...) is adopted. It can reflect recent changes in health status in a timely manner while maintaining historical information, avoiding excessive impact of short-term abnormal disturbances on the benchmark.

[0060] In another embodiment, such as Figure 3 As shown, abnormal battery cells are identified based on deviation, including: An adaptive anomaly threshold is determined based on the set of singular energy indicators of all individual battery cells. Calculate the projected coordinates of each battery cell in the main health subspace and its distance to the group health center, and use the distance as the main modal distance; If the singular energy index of a certain battery cell exceeds the adaptive anomaly threshold, or if the main mode distance of a certain battery cell exceeds the corresponding distance threshold, the battery cell is marked as an anomaly candidate. A comprehensive anomaly score is constructed, which is the sum of the ratios of the weighted combination of singular energy index and principal mode distance to the corresponding threshold. When the comprehensive abnormal score of an abnormal candidate battery cell is greater than 1, it is determined to be an abnormal battery cell.

[0061] It should be explained that the above implementation method has already defined the exotic energy index. This is used to quantitatively characterize the degree of deviation of the i-th battery cell from the mainstream health pattern of the population. However, relying solely on... The absolute value of the threshold cannot directly determine anomalies because the characteristic distribution range may differ across different detection cycles and battery packs. Therefore, an adaptive anomaly threshold determination method based on the statistical distribution within a group is needed. Meanwhile, The primary focus is on capturing anomalous variations orthogonal to the main health subspace. However, the location of an individual within the main health subspace—i.e., the offset of its overall health level relative to the population center—may also contain anomalous information. For example, if an individual's health feature vector lies entirely within the main health subspace but is far from the population center, it indicates that the individual's overall health level is significantly lower than the population average, and should also be considered an anomalous. Therefore, this embodiment introduces the principal modality distance and combines it with the singular energy index to construct a comprehensive anomaly score, thereby improving the accuracy and robustness of anomaly identification.

[0062] After completing the construction of the population health characterization space, the singular energy index of each of the N battery cells in the current detection period has been calculated. (i=1,2,...,N). These Form a set: , Due to different battery packs and different testing cycles The absolute value range may vary due to factors such as feature scale, number of batteries, and degree of aging, therefore a fixed numerical threshold cannot be used. This embodiment employs an adaptive threshold determination method based on population statistics, specifically including two optional schemes: statistical thresholding and quantile thresholding.

[0063] Statistical thresholding: Calculate all monomers mean and standard deviation : ; ; In the assumption Assuming the distribution approximately follows a unimodal pattern (such as a chi-square distribution or a log-normal distribution), an anomaly threshold can be set as follows: +c· , where c is a constant (usually taken as 2 or 3, corresponding to approximately 95% or 99% confidence levels, respectively).

[0064] Quantile thresholding method: defining outlier thresholds Let p be the p-quantile of the SPE set, i.e.: ,in This represents the quantile function, where p is the preset percentile, typically taken as 95% or 99%. For example, when p = 95%, The value makes 95% of the monomers Less than or equal to this value, while 5% of the monomers The value is greater than this. The quantile threshold method does not require the data to follow a specific distribution, is insensitive to outliers, and is more robust and reliable. In practical applications, the p-value can be adjusted according to the tolerance for false alarms and false negatives in the field: if you want to reduce false negatives (i.e., find as many potential anomalies as possible), you can decrease the p-value (e.g., 90%); if you want to reduce false alarms (i.e., reduce unnecessary maintenance checks), you can increase the p-value (e.g., 99%).

[0065] In addition to the singular energy index, this embodiment also introduces the principal modal distance to measure the degree of deviation of a single entity's position in the principal health subspace from the population health center.

[0066] First, define the group health center. In the main health subspace, the group health center can be defined as the origin (because the mean of the standardized feature matrix Z is 0, and the main health subspace passes through the origin), or it can be defined as the average of the projected coordinates of all individuals in the main health subspace. ,in It is the projection matrix (P×K) formed by the first K eigenvectors. To simplify the calculation, this embodiment uses the origin as the health center of the population, i.e., c=0. This is because in the standardized feature space, the natural center of the population is the origin.

[0067] For the i-th battery cell, its standardized eigenvector The Euclidean distance to the community health center c. Since c=0, we have:

[0068] Dominant mode distance The physical meaning is: the magnitude of the health feature vector of a single battery cell in the main health subspace reflects the degree to which the overall health level of that cell deviates from the average level of the population. A healthy cell, close to the average level of the population, has its projected coordinates close to the origin. Smaller; conversely, a monomer whose overall health has significantly deteriorated (e.g., a sharp decrease in volume and a general increase in internal resistance), even if its abnormal variation direction may still fall within the main health subspace (i.e., its abnormality is an extreme manifestation of the "mainstream direction"), its It will also be relatively large.

[0069] Similar to the singular energy metric, the dominant mode distance also requires an adaptive anomaly threshold. Define the distance threshold. The p-quantiles of the distances between all individual principal modes are given (p is also taken as 95% or 99%).

[0070] Based on the above two indicators and the corresponding adaptive thresholds, this embodiment first marks abnormal candidate cells. The marking rule is: if a certain battery cell i satisfies at least one of the following two conditions, it is marked as an abnormal candidate: (1) > (Exotic energy index exceeds threshold).

[0071] (2) > (The distance to the dominant mode exceeds the threshold).

[0072] The physical logic of this rule is: Exceeding the threshold indicates that there are "non-mainstream" abnormal variations in the health characteristics of the individual cell that cannot be explained by the common pattern of the group, such as abnormally increased internal resistance, abnormal relaxation behavior, etc. These abnormalities often indicate latent, localized battery degradation. Exceeding the threshold indicates that the overall health level of the individual is significantly lower than the population average, meaning that the individual has shown significant degradation in mainstream health dimensions. Both methods capture anomalies from the perspectives of "abnormal direction" and "extreme degree," respectively, complementing each other and effectively reducing missed diagnoses.

[0073] Furthermore, using any of the above indicators in isolation may lead to misjudgment or omission. For example, a single entity may... Slightly exceeding the threshold but The reading is far below the threshold, which may simply be due to measurement noise; conversely, a single cell may be... Slightly exceeding the threshold but This is normal and may simply be individual variation within the normal aging range. Therefore, this embodiment constructs a comprehensive anomaly score. The two indicators are normalized and then weighted and combined.

[0074] Specifically, the definition is: , in: / The ratio of singular energies after normalization is greater than 1. Exceeding the threshold; / The normalized principal mode distance is the relative value; a value greater than 1 indicates... Exceeding the threshold; and These are the weighting coefficients.

[0075] The choice of weighting coefficients reflects the emphasis placed on the two types of indicators. In typical energy-saving elevator energy storage devices, exotic energy indicators are more sensitive to early latent anomalies, and therefore can be assigned a higher weight, for example... =0.7, =0.3. Of course, the weighting coefficient can also be dynamically adjusted based on historical experience or through a manual judgment feedback mechanism.

[0076] Comprehensive Anomaly Score The physical meaning is: to perform a dimensionless weighted summation of the deviations between two dimensions. When A value >1 indicates that the overall deviation of this single cell, in a weighted sense, exceeds the weighted sum of the abnormal threshold combinations, and it is thus identified as an abnormal battery cell. This judgment rule is more stringent than the simple "any indicator exceeding the threshold" rule, effectively reducing false alarms caused by accidental fluctuations of a single indicator; simultaneously, because... It is a continuous value and can be further used for anomaly classification.

[0077] The anomaly identification method proposed in this embodiment has the following beneficial effects: First, it has strong adaptability. The anomaly threshold is dynamically calculated based on the actual data distribution of all battery cells in the current detection cycle (such as the quantile method). There is no need to preset a fixed threshold. It can automatically adapt to the characteristic distribution changes of different battery packs, different aging stages, and different operating conditions, avoiding false alarms or missed alarms caused by improper setting of fixed thresholds.

[0078] Second, a multi-dimensional comprehensive judgment is adopted. By simultaneously considering singular energy indicators (reflecting non-mainstream abnormal variations) and dominant mode distance (reflecting the overall degradation of mainstream health dimensions), a comprehensive anomaly score is constructed, making the identification results more comprehensive and reliable. For example, if a battery cell deviates from the center in the main health subspace merely due to manufacturing discreteness but without actual abnormal degradation, its... Typically, the value is very small, and the overall score will not exceed the threshold; conversely, a single cell with a slight but persistent tendency for internal micro-short circuits may have a significantly higher risk. Still within the normal range, but The score will increase significantly, but the overall score can still be correctly identified.

[0079] Through the above steps, this implementation method achieves the identification of abnormal battery cells based on population self-reference, adaptive threshold, and multi-index fusion under the condition of no external standard samples.

[0080] In one embodiment, identifying abnormal battery cells based on deviation further includes: Anomalies are classified into multiple levels based on anomaly scores or the degree of excess of exotic energy. If the same battery cell is determined to reach the preset abnormal level in multiple consecutive testing cycles, it is confirmed as a stable abnormality.

[0081] It should be explained that this implementation further defines the grading of anomalies and introduces a continuous confirmation mechanism. In practical engineering applications, the health degradation of a single battery cell is a gradual process, often taking months or even longer from early minor deviations to severe failure. An increase in anomaly scores during a single detection cycle may stem from various reasons: it could be due to early battery degradation, or it could be caused by short-term environmental disturbances (such as instantaneous temperature fluctuations), measurement noise, communication interference, or load fluctuations—factors not directly related to the battery itself. If maintenance alarms are triggered immediately every time a score exceeds a threshold, it will lead to a large number of false alarms, reducing maintenance personnel's trust in the system. Conversely, if alarms are only triggered when an anomaly reaches a certain severity, the optimal opportunity for early intervention may be missed. Therefore, this implementation introduces anomaly grading and continuous confirmation mechanisms. While maintaining sensitivity to early anomalies, it filters out transient disturbances through multi-cycle continuous observation and outputs different levels of maintenance recommendations based on the severity of the anomaly.

[0082] The aforementioned implementation method has calculated a comprehensive anomaly score for each battery cell i. Furthermore, it is possible to obtain the singular energy index of this single entity. Above relative to the threshold The extent to which limits are exceeded. This embodiment classifies anomalies into multiple levels based on these quantitative indicators, so that operations and maintenance personnel can take differentiated response measures according to the severity.

[0083] Specifically, the following hierarchical rules are defined: Level 1 Anomaly (Mild Deviation): Overall Anomaly Score Meets Standard Alternatively, if classification is based solely on singular energy indices (without using a comprehensive score), then when < ≤ ′· It was determined to be a Level 1 anomaly. This is the first grading threshold, typically ranging from 1.2 to 1.5. ′ represents the singular energy rate threshold, typically ranging from 1.2 to 1.5. A Level 1 anomaly indicates a detectable deviation in the health characteristics of the battery cell, but the deviation is not yet severe, possibly indicating early degradation or a one-off minor disturbance. For Level 1 anomalies, the system can output an "Operational Attention Notice," suggesting that maintenance personnel monitor the subsequent trends of the cell, but no immediate maintenance is required.

[0084] Level 2 abnormality (significant degradation): The overall abnormality score meets the requirements. Or the exotic energy index satisfies ′· < ≤ ′· .in This is the second-level threshold, typically ranging from 2.0 to 3.0. ′ represents the singular energy rate threshold, typically ranging from 2.0 to 3.0. A Level 2 anomaly indicates a significant deviation from the mainstream health pattern of the battery cell, likely signifying substantial battery degradation, such as a marked increase in internal resistance, intensified polarization, or capacity decay. For Level 2 anomalies, the system should output a "derating recommendation" or "recent inspection prompt," suggesting appropriate limitation of the energy storage device's charge and discharge power during subsequent operation and scheduling upcoming manual inspections or cell balancing maintenance.

[0085] Level 3 Abnormality (Severe Abnormality): The comprehensive abnormality score meets the requirements. > Or the exotic energy index satisfies > ′· A Level 3 anomaly indicates that the health characteristics of the individual cell deviate significantly from the normal range of the group, and there is a high probability of major safety hazards such as internal micro-short circuits, severe leakage, or precursors to thermal runaway. For a Level 3 anomaly, the system should immediately output a "maintenance or replacement recommendation" and suggest that the battery cell be immediately isolated from the energy storage device or replaced to prevent the fault from spreading or causing a safety accident.

[0086] The above-mentioned grading thresholds , as well as ′、 The specific value of ′ can be configured based on factors such as field experience, battery type, and safety requirements. For example, in situations with high safety requirements, it can be... Set it to 1.2. Setting it to 2.0 lowers the threshold for level 2 anomalies, allowing for earlier warnings; for situations where a certain number of false alarms are permissible, it can be... Set it to 1.5. Set to 3.0 to reduce unnecessary maintenance reminders. These thresholds can also be adaptively adjusted through a manual feedback mechanism.

[0087] It should be noted that anomaly classification can be based on comprehensive anomaly scores. It can also be based solely on the singular energy index. The degree to which limits are exceeded. In practical applications, due to... Since it already incorporates the main modality distance information, it can more comprehensively reflect the degree of anomaly. Therefore, it is recommended to prioritize the use of methods based on... The hierarchical method. However, in some simplified implementations, only the hierarchical method may be used. The classification is performed, and the threshold in the classification rules is adjusted accordingly to the singular energy multiplier threshold.

[0088] Since abnormal score increases during a single testing cycle may be caused by transient disturbances (such as electromagnetic interference, sampling noise, or communication packet loss), this embodiment introduces a continuous verification mechanism to avoid triggering unnecessary maintenance actions due to such non-continuous anomalies. This mechanism requires that the same battery cell must be determined to reach a preset anomaly level (e.g., level one or above, or a specific level) in multiple consecutive testing cycles before it is confirmed as a "stable anomaly," thereby triggering corresponding maintenance recommendations or alarms.

[0089] The specific implementation method is as follows: Define L as the threshold for the number of consecutive anomaly confirmation cycles, where L is a positive integer greater than or equal to 2, typically 3. Let the current detection cycle be t. For battery cell i, the system records the anomaly level determination results for its most recent L detection cycles (including the current cycle). If, in each of these L consecutive cycles, the cell is determined to have reached at least the preset anomaly level threshold (e.g., Level 1 anomaly or higher), then the cell is confirmed as a "stable anomaly," and the maintenance recommendation corresponding to the highest anomaly level currently reached by the cell is output.

[0090] If, within a consecutive L cycles, the anomaly level of a single entity is below a preset threshold (i.e., it is judged as normal or only below level one), the continuity counter is reset to zero, and it is not confirmed as a stable anomaly. Subsequently, if the single entity experiences an anomaly again, it needs to accumulate L consecutive cycles again before confirmation is triggered.

[0091] For example, assuming L=3, the preset anomaly level threshold is Level 1 (meaning that any level 1 or higher is counted in the continuous count). If battery cell A is determined to be Level 1, Level 2, and Level 1 in the three consecutive detection cycles t, t+1, and t+2 respectively, then since it reaches Level 1 or higher in each cycle, it is confirmed as a stable anomaly at the end of cycle t+2, and a derating recommendation is output according to Level 2 anomaly (the highest level). If it is Level 1 in cycle t, normal in cycle t+1 (not reaching Level 1), and again Level 1 in cycle t+2, then the continuity is interrupted, a stable anomaly is not confirmed, and only a one-time Level 1 anomaly warning is output in cycle t+2, but no maintenance recommendation is triggered.

[0092] Different consecutive confirmation period requirements can be set for anomalies of different levels. For example, a Level 1 anomaly (minor deviation) may require three consecutive periods of occurrence before it is confirmed as a stable anomaly, to avoid overreacting to short-term fluctuations; while Level 2 or 3 anomalies (significant or severe degradation), due to their higher risk, may require two consecutive periods or even just a single confirmation to trigger maintenance recommendations. This differentiated setting reflects the concept of risk-based hierarchical management: the higher the risk of the anomaly, the lower the confirmation threshold to ensure timely handling; the lower the risk of the anomaly, the higher the confirmation threshold to filter out noise.

[0093] It's easy to understand that the core principle of the continuity verification mechanism is to use information redundancy in the time dimension to suppress random errors in the spatial dimension and in single measurements. In engineering practice, abnormal data caused by factors such as electromagnetic interference, poor sampling line contact, and communication interruptions usually manifest as "isolated events," meaning they only appear in one or a few discontinuous detection cycles and do not repeat in multiple consecutive cycles. Conversely, abnormal characteristics caused by battery health degradation (such as continuously increasing internal resistance and gradually changing polarization time constant) have temporal continuity and monotonicity, and will continue to appear in consecutive detection cycles. Therefore, by requiring verification over multiple consecutive cycles, it is possible to effectively distinguish between battery-related abnormalities and external disturbances.

[0094] The continuity verification mechanism in this embodiment has the following beneficial effects: First, it significantly reduces the false alarm rate. Without continuous confirmation, a single, isolated false alarm... Exceeding the threshold will trigger an alarm, causing maintenance personnel to receive frequent invalid alarms and reducing their trust in the system. By requiring confirmation for three consecutive cycles, the false alarm probability of random disturbances can be reduced to the order of 1 / L (assuming the disturbances are independently and identically distributed in each cycle), greatly improving the reliability of alarms.

[0095] Second, avoid premature intervention. Battery health degradation is usually a slow process; early level 1 anomalies may take weeks or even months to develop into level 2 anomalies. If maintenance is triggered by a single level 1 anomaly, it may lead to unnecessary downtime and inspections. A continuous verification mechanism allows the system to issue recommendations only after confirming that the degradation trend is ongoing, ensuring that genuine degradation is not missed while avoiding overreaction.

[0096] Third, it supports a tiered maintenance strategy. Based on anomaly classification, the system can set different consecutive confirmation cycles for different levels, enabling refined health management. For example, Level 1 anomalies require three consecutive confirmation cycles before attention is recommended; Level 2 anomalies require two consecutive confirmation cycles before a reduction in risk level is recommended; and Level 3 anomalies require only a single confirmation before replacement is recommended. This strategy ensures both security (rapid response to severe anomalies) and reduced operational costs (continuous monitoring of minor anomalies).

[0097] In one embodiment, identifying abnormal battery cells based on deviation further includes: Calculate the contribution of each characteristic component of the abnormal battery cell in the singular deviated subspace; Based on the category of the feature component with the largest contribution, the anomaly is mapped to an internal resistance anomaly, polarization anomaly, relaxation anomaly, or consistency degradation anomaly.

[0098] It should be explained that the aforementioned implementation method determines whether a battery cell is abnormal by comprehensively scoring anomalies or identifying singular energy indicators, but it does not explain which specific health characteristics cause the anomaly. In practical engineering applications, maintenance personnel not only need to know "which cell is abnormal," but also want to understand "what the cause of the anomaly is" in order to take targeted maintenance measures (such as replacing cells with abnormal internal resistance, cells with large differences in equalization polarization, or handling sampling line faults). This implementation method utilizes the structural characteristics of the constructed singular deviation subspace, and by analyzing the contribution of each projection component of the abnormal cell in this space, maps the anomaly into an anomaly type with clear physical meaning, thereby providing actionable guidance for on-site maintenance.

[0099] After completing principal component analysis, singular deviation subspace It is spanned by the last PK feature vectors. For any battery cell i (including cells judged to be abnormal), its standardized feature vector is... The projection vector onto the singular deviated subspace is: , in Let P×(PK) be the projection matrix. It is a row vector of length (PK). Each component of the projection vector... (j=K+1,…,P) corresponds to the projected coordinates in the direction of the j-th eigenvector.

[0100] What needs to be understood is that each feature vector It is itself a P-dimensional vector, where each component (i.e., the p-th element of the feature vector) represents the weight of the original P health features in that singular direction. Therefore, yes and The inner product of these two factors reflects the "strength" of the health eigenvector of cell i in the j-th singular direction. However, The singularity itself can be positive or negative, and different singular directions have different physical meanings. To quantify the contribution of each original health feature to the anomalous deviation, a contribution decomposition is required.

[0101] The goal of contribution analysis is to identify the singular energy index that leads to monomer i. During the voltage rise process, which original health characteristics (such as initial voltage transition and relaxation time constant) play a dominant role? This embodiment can employ a contribution calculation method based on projection reconstruction.

[0102] First, the exotic energy index This can be expressed as the sum of squares of the projection components in each singular direction: Each squared term ^2 represents the j-th singular direction pair The contribution of the singularity direction is still not specific to the original feature dimension. To attribute anomalies to specific health features (such as internal resistance-related features, polarization-related features, etc.), it is necessary to further distribute the contribution of each singularity direction to each original feature according to the feature load in that direction.

[0103] The projected energy in each singular direction is distributed to each original feature according to the square of the feature load in that direction, and then the total contribution of each original feature is obtained by summing them. The larger the contribution, the greater the role of the p-th healthy feature in causing the individual to deviate from the mainstream pattern of the population.

[0104] In practical calculations, to facilitate comparison of the contributions of different features, the contributions of all P features can be normalized to obtain the relative contribution percentage.

[0105] Then, by sorting them from largest to smallest according to their relative contribution, the characteristics of the dominant anomaly can be identified.

[0106] After extracting the dominant health characteristics (i.e., the one or several characteristics with the greatest contribution) of the abnormal cells, these specific feature dimensions need to be mapped to anomaly types with engineering significance. To this end, this embodiment predefines the mapping rules between health characteristics and anomaly types. Based on the main failure modes of the batteries in the energy storage device of the energy-saving elevator, anomalies are divided into the following four types: (1) Internal resistance anomaly: This type of anomaly is mainly caused by an increase or decrease in the battery's ohmic internal resistance. The corresponding health characteristics include: initial voltage jump (the voltage jump at the moment of charging, which directly reflects the ohmic internal resistance) and the initial voltage rise rate during the charging stage (which is also related to internal resistance and polarization internal resistance). When the characteristic with the greatest contribution belongs to the above characteristics, the anomaly is mapped as "internal resistance anomaly". Internal resistance anomaly usually means that there are problems such as deterioration of the conductive network, loose connection pieces or electrolyte drying in the battery. It is recommended to retest the cell's internal resistance or replace it.

[0107] (2) Polarization Anomaly Type: This type of anomaly reflects the abnormal intensification of electrochemical polarization or concentration polarization during battery charging and discharging. The corresponding health characteristics include: maximum slope and its dispersion (reflecting unstable behavior during polarization establishment), mean and maximum value of voltage curvature characteristics (characterizing the degree of concentration polarization), relaxation voltage decay amplitude (total cumulative polarization), and relaxation time constant (distinguishing between rapid polarization and slow diffusion). When the characteristic with the greatest contribution belongs to the above characteristics, it is mapped as "polarization anomaly type". Polarization anomalies may originate from the thickening of the SEI film on the electrode surface, the decrease in the utilization rate of active materials, or the obstruction of lithium-ion diffusion. It is recommended to perform low-rate charge-discharge activation or cell equalization.

[0108] (3) Relaxation Anomaly Type: This type of anomaly is mainly related to abnormal voltage recovery behavior after the pulse ends. The corresponding health characteristics include: relaxation fitting residual energy (the degree to which relaxation behavior deviates from the standard exponential model), abnormal relaxation voltage decay amplitude, and relaxation time constant deviating from the population mean. When the feature with the greatest contribution is the feature of the relaxation stage (especially the fitting residual energy), it is mapped to "relaxation anomaly type". Relaxation anomalies often indicate that there are hidden faults such as micro-short circuits, abnormal self-discharge rate, or electrolyte decomposition inside the battery. It is recommended to conduct isolation testing or capacity verification on the cell.

[0109] (4) Consistency Degradation Type: This type of anomaly is not dominated by a single feature, but rather by multiple features exhibiting deviations, or is mainly reflected in the relative deviation features of the group. Specifically, when the feature with a high contribution is a relative deviation feature of the group and no single physical feature is dominant, or when the principal mode distance increases significantly but the singular energy index is not prominent, it is mapped to the "consistency degradation type". The consistency degradation type reflects that the individual cell is lagging behind the average level of the group in multiple health dimensions. It is usually caused by manufacturing discreteness, long-term temperature unevenness, or the accumulation of charge and discharge imbalance. It is recommended to perform cell balancing or whole-group maintenance.

[0110] Traditional anomaly detection methods (such as threshold-based voltage monitoring) can only indicate "voltage anomaly" but cannot explain whether it is due to internal resistance issues, polarization problems, or other causes. This embodiment, through contribution analysis, transforms abstract statistical deviations into specific, physically meaningful anomaly types, greatly improving the usability of diagnostic results and facilitating rapid fault location by field engineers. For internal resistance anomalies, replacement or retesting of the internal resistance is recommended; for polarization anomalies, low-rate activation or equalization charging is recommended; for relaxation anomalies, immediate isolation and in-depth testing are recommended; and for consistency degradation, overall equalization or adjustment of thermal management strategies is recommended. This differentiated maintenance guidance can effectively reduce operation and maintenance costs and extend the overall lifespan of energy storage devices. Furthermore, all calculations are based on the already constructed population health characterization space and existing feature vectors, without adding additional detection steps or data acquisition, resulting in minimal computational load.

[0111] In one embodiment, constructing a population health representation space as a reference benchmark for population health status through multivariate statistical analysis methods further includes: At the end of each detection cycle, the feature matrix and covariance matrix are updated based on the newly acquired data; A sliding time window weighted update method is adopted. The covariance matrix saved in the previous detection period and the covariance matrix calculated in the current detection period are weighted and fused by the forgetting factor to obtain the updated covariance matrix, so as to suppress the impact of short-term abnormal interference on the population health representation space.

[0112] It should be explained that this embodiment further defines the online update method for the characterization space across multiple detection cycles. It is easy to understand that during long-term operation, the health status of individual battery cells in an energy-saving elevator system gradually deteriorates with the increase in charge-discharge cycles, while operating conditions such as ambient temperature and load mode may also change slowly. If the group health characterization space is constructed completely independently for each detection cycle (i.e., based solely on data from the current cycle), although it can reflect the current health distribution, historical information will be lost, leading to significant fluctuations in the baseline between different cycles, which is detrimental to the stability of anomaly identification and the tracking of degradation trends. Conversely, if a fixed historical baseline (e.g., the baseline at the time of initial commissioning) is always used, it cannot adapt to the overall aging of the battery pack, causing more and more normal cells to be misjudged as abnormal. Therefore, an online update method is needed that can both track changes in the group health status in a timely manner and suppress short-term anomaly interference.

[0113] This embodiment employs a sliding time window weighted update strategy, as detailed below.

[0114] I. Periodic Update Strategy At the end of each detection cycle (i.e., after obtaining the standardized feature matrix of all battery cells in the current cycle) and the corresponding covariance matrix Afterwards, the system initiates the update process. The update operation does not recalculate all historical data each time, but rather uses the covariance matrix saved from the previous period. (Or more generally, the covariance matrix updated in the previous period) and the covariance matrix calculated in the current period. Perform weighted fusion. The advantages of this approach are: no need to store all historical data, low computational cost, and suitability for real-time execution in the cloud or edge computing units.

[0115] II. Weighted Update Formula for Sliding Time Window The forgetting factor α is defined as a constant between 0 and 1, i.e., α∈(0,1). The forgetting factor controls the relative weight of historical information and current information in the update process. The update formula is: , in: This represents the weighted covariance matrix saved in the previous detection period (or the previous update period), which has incorporated historical information up to the previous period; This represents the covariance matrix calculated based on the standardized feature matrices of all battery cells within the current detection period (based only on data from the current period). This represents the updated covariance matrix, used for constructing the population health characterization space in the next cycle; α is the forgetting factor, typically ranging from 0.7 to 0.95, and the specific value can be adjusted according to the battery aging rate and the length of the detection cycle.

[0116] It should be noted that in the first detection period (t=1), since there is no historical covariance matrix, It can be initialized to a zero matrix, or you can directly set it to zero. = (That is, no weighting is applied). Starting from the second cycle, the weighted update formula described above is formally applied.

[0117] Furthermore, the physical meaning of the forgetting factor α lies in the fact that it determines the weight of the historical covariance matrix in the updated matrix. The closer α is to 1, the more historical information is retained, the less sensitive the updated matrix is ​​to short-term fluctuations, and the higher the stability of the population health representation space, but the slower the response to the actual aging trend of the battery pack. The closer α is to 0, the greater the weight of current information and the faster the response speed, but it is easily affected by accidental factors in a single detection cycle (such as temperature fluctuations, measurement noise, and interference from a few abnormal cells), leading to unnecessary drift of the baseline.

[0118] In this embodiment of the invention, considering that battery health degradation is a relatively slow process (typically occurring gradually over hundreds of days or even years), and the detection cycle is usually several days to a week, a large forgetting factor, such as α=0.9, is recommended. This means that in the updated covariance matrix, historical information accounts for 90%, while current information accounts for only 10%. This allows for slow tracking of the overall aging trend of the battery pack while effectively suppressing abrupt changes in the covariance matrix caused by individual abnormal cells or external disturbances during a particular detection.

[0119] To more accurately adapt to different scenarios, the forgetting factor can also be set to an adaptive form. For example, the difference between the covariance matrix of the current period and the previous period (such as the relative change after Frobenius norm normalization) can be calculated. When the difference is large (indicating possible anomalies or sudden changes in operating conditions), α is appropriately reduced (i.e., the weight of the current information is increased) to quickly respond to actual changes; when the difference is small, α is increased to maintain stability. However, as a basic implementation method, fixing α is sufficient to meet most engineering needs.

[0120] Obtain the updated covariance matrix Subsequently, the construction of the population health characterization space for the next testing cycle will be based on The process is as follows: Specifically, during the next detection cycle, the system first collects data for that cycle, extracts and standardizes features, then calculates the covariance of the standardized feature matrix, and then... That is, the covariance matrix saved in the previous period is weighted and fused to obtain the new one. (The subscripts are cyclic here). However, it should be noted that the weighted update updates the covariance matrix itself at the end of each period, while the construction of the population health representation space for the next period still requires recalculation of the feature matrix and projection direction based on the data collected in the previous period; If the covariance matrix calculated independently for the current period is used directly to construct the population health representation space, then when there are many anomalous individuals or poor data quality in a certain detection period, the covariance matrix will deviate significantly from the normal distribution, causing distortion in the main health subspace. This could lead to anomalous individuals being incorrectly included in the main health subspace, while normal individuals are misclassified as anomalous. By using a sliding time window for weighted updates, historical information (i.e., the normal distribution over multiple past periods) plays an "anchoring" role. Even if there are short-term anomalous disturbances in the current period, the matrix is ​​still mainly dominated by the historical normal distribution because the weight of historical information α is relatively large (e.g., 0.9), thus ensuring the stability and robustness of the population health representation space. At the same time, since the forgetting factor α is less than 1, historical information will gradually decay over time, so long-term aging trends can still be gradually absorbed into the covariance matrix. For example, when the average internal resistance of the entire battery pack slowly increases, the covariance matrix of each period will reflect this change. After weighted fusion, it will also drift slowly, allowing the population health representation space to track the overall health evolution.

[0121] Based on the above, the sliding time window weighted update method proposed in this embodiment has the following beneficial effects: First, smoothing filtering suppresses short-term fluctuations. By introducing a forgetting factor to weight and fuse the historical and current covariance matrices, the baseline disturbances caused by temperature fluctuations, measurement noise, communication anomalies, or individual extreme anomalies in a single detection cycle are effectively reduced, making the population health representation space more robust and significantly reducing the false alarm rate of anomaly identification.

[0122] Second, it adaptively tracks long-term degradation. The forgetting factor gives historical information a "limited memory," enabling the gradual elimination of outdated health distribution information, thereby tracking the overall gradual aging of the battery pack. In this way, even if all cells age to a certain extent simultaneously, the overall health representation space will adjust accordingly, preventing normal aging from being misjudged as abnormal.

[0123] Third, the parameters are adjustable and highly adaptable. The forgetting factor α can be flexibly set according to actual working conditions. For systems with a fast aging rate (such as those with frequent deep charge and discharge), α can be appropriately reduced to improve the response speed; for systems with a slow aging rate or a short detection cycle, α can be increased to enhance stability.

[0124] In one embodiment, the preset operating conditions include: the state of charge of the energy storage device reaches or exceeds a preset threshold, and the battery temperature is within a preset allowable range; Health management methods also include: When the state of charge is lower than the preset threshold, the energy storage device is charged until the preset threshold is met; When the temperature exceeds the allowable range, the step of suspending the application of controlled pulse excitation to the energy storage device for a preset duration is paused.

[0125] It is easy to understand that this embodiment further defines the preset operating conditions that must be met before applying controlled pulse excitation, and the handling measures to be taken when the conditions are not met. In actual operation, the state of charge (SOC) and battery temperature of the energy storage device in the energy-saving elevator system vary considerably depending on factors such as the elevator's workload, ambient temperature, and charging / discharging history. The battery's terminal voltage response characteristics are not only related to its health status but also strongly dependent on the current SOC and temperature. For example, in the low SOC range, the battery's equivalent internal resistance increases significantly, and the polarization effect is more pronounced. Under these conditions, even a battery in good health may exhibit a large voltage jump and a slow relaxation response when subjected to the same pulse excitation, thus being misjudged as abnormal. Conversely, under high temperature conditions, the battery's internal resistance decreases, polarization weakens, and a battery in poor health may exhibit a response close to that of a normal battery, leading to missed detection. Therefore, to ensure the comparability of health characteristics between different detection cycles and between different battery cells, the SOC and temperature must be controlled within a relatively consistent and suitable standard range before applying excitation. This implementation aims to eliminate the impact of SOC and temperature, the two main interfering factors, on the extraction of health features by setting a SOC threshold and an allowable temperature range, and performing corresponding preprocessing operations.

[0126] In this implementation, the SOC preset threshold is defined as follows: The typical value is 80% to 90%, with 90% being preferred in this embodiment. This threshold is chosen based on the following principle: In a higher SOC range (e.g., above 90%), the battery's terminal voltage is at the end of the charging platform or close to full charge. At this point, the battery's polarization state is relatively stable, and small changes in SOC have little impact on the voltage response. Simultaneously, the battery's equivalent internal resistance and polarization characteristics at high SOC better reflect its intrinsic health state because the utilization rate of the active material in the electrode material is close to saturation; any increase in internal resistance or loss of active material will be significantly reflected in the voltage response. Conversely, in a low SOC range (e.g., below 30%), the battery voltage changes drastically with SOC, and its polarization behavior is complex, making it unsuitable as a benchmark for health monitoring.

[0127] At the start of each detection cycle, when executing the detection trigger and operational preparation steps, the system first reads the current SOC value of the energy storage device. If the current SOC ≥ If the condition is met, the process can proceed directly to the subsequent controlled pulse excitation application step. If the current SOC < The system then charges the energy storage device through the elevator system's energy management unit or energy storage converter until the State of Charge (SOC) reaches or exceeds the specified value. The charging process should preferentially employ low-rate constant current or constant power charging to avoid introducing additional polarization differences due to rapid charging. After charging is complete, the battery should be allowed to rest for a period of time (e.g., 5 to 10 minutes) to allow the internal polarization to relax to some extent before controlled pulse excitation begins. The purpose of resting is to eliminate the polarization voltage accumulated during charging, ensuring that the initial state before pulse excitation is as consistent as possible.

[0128] It is important to note that the recharging process only occurs when the State of Charge (SOC) does not meet the threshold, and the goal of recharging is to bring the SOC to the threshold, not to fully charge it to 100%. This is because charging the SOC from a low value to 100% may take a long time, affecting the timeliness of the detection; charging to 90% provides sufficient comparability and avoids the impact of prolonged charging on battery life. Furthermore, if the SOC is too low (e.g., below 20%), the system can also choose to skip the current detection cycle and wait for the elevator system to operate normally and allow the SOC to naturally recover before resuming detection. However, this embodiment prioritizes active recharging to ensure the regularity of the detection cycle.

[0129] Furthermore, this embodiment defines the allowable battery temperature range as [T_min, T_max], where T_min is the minimum allowable temperature, typically ranging from 15°C to 20°C; and T_max is the maximum allowable temperature, typically ranging from 35°C to 40°C. This embodiment preferably uses a range of 20°C to 35°C. Temperature has a significant impact on the electrochemical behavior of the battery: at low temperatures, electrolyte viscosity increases, lithium-ion migration rate decreases, electrode reaction kinetics slow down, leading to a sharp increase in battery internal resistance and intensified polarization; at high temperatures, although internal resistance decreases, side reactions (such as SEI film growth and electrolyte decomposition) are accelerated, and response data at high temperatures cannot represent the health status at room temperature. Therefore, only by conducting tests within the allowable temperature range can the comparability of data from different periods and seasons be ensured.

[0130] Before applying the controlled pulse excitation, the system reads the battery temperature using temperature sensors located inside the energy storage device (typically multiple temperature measurement points within each battery module or battery box). If the temperatures at all measurement points are within the range [T_min, T_max], the condition is met, and subsequent steps can proceed. If the temperature at any measurement point exceeds this range (e.g., below T_min or above T_max), the system pauses the current detection and does not apply the controlled pulse excitation. Simultaneously, the system records the current temperature exceeding the range and sets up a delayed retry mechanism. For example, the system can re-detect the temperature every 10 minutes until it recovers to the allowable range before triggering the detection process again. If the temperature cannot recover for an extended period (e.g., due to excessively low ambient temperature and the absence of a heating device), the system can skip the current detection cycle and attempt again in the next predetermined detection cycle.

[0131] It should be noted that the setting of the allowable temperature range should take into account the battery type and the actual application environment. For lithium iron phosphate batteries, which have poor low-temperature performance, T_min can be appropriately increased to 20°C; for lithium titanate batteries, which have better wide-temperature performance, T_min can be relaxed to 10°C. Furthermore, for energy storage devices equipped with a thermal management system (such as a heating film or air-cooled / liquid-cooled system), the thermal management system can be actively activated before testing to adjust the battery temperature to the allowable range before testing. This embodiment does not exclude this active temperature control method, but the core requirement is that the battery temperature must be within the preset allowable range at the moment the pulse excitation is applied.

[0132] In practice, the following order can be followed when checking and handling operating conditions: First, check the State of Charge (SOC). If the SOC is insufficient, perform a charge-up operation (charge to the threshold and let it rest). This is because the charge-up process may cause the battery temperature to rise. If the temperature is checked before charge-up, the temperature may exceed the allowable range after charge-up, requiring another cooling period. Once the SOC meets the requirements, check the battery temperature again. If the temperature exceeds the allowable range, delay the detection and retry periodically until the temperature recovers or the maximum waiting time is reached, then skip the current detection.

[0133] Once both SOC and temperature meet the requirements, controlled pulse excitation is applied immediately. Furthermore, to improve data comparability, temperature can be continuously monitored during the detection process. If the temperature changes beyond the allowable range during pulse excitation (e.g., due to temperature rise caused by the pulse current itself), the detection can be interrupted and marked as invalid, and restarted after the temperature stabilizes.

[0134] This implementation achieves the following beneficial effects by setting a preset SOC threshold and a permissible temperature range, and performing supplementary power or delayed detection when the conditions are not met: First, it eliminates the impact of SOC differences on voltage response. Testing is conducted in the high SOC range (above 90%), where the battery's voltage response depends primarily on internal resistance and polarization characteristics, rather than minute fluctuations in SOC. This makes the health characteristics comparable across different testing cycles and between different individual battery cells, avoiding misjudgments caused by differences in SOC.

[0135] Second, it eliminates the interference of temperature differences on the extraction of health characteristics. By controlling the temperature within a suitable range, it ensures that key parameters such as battery internal resistance and polarization time constant are within the normal range, and will not mask the true changes in health status due to artificially high internal resistance caused by low temperature or artificially low internal resistance caused by high temperature.

[0136] Third, it balances detection timeliness and data quality. When SOC is insufficient, it proactively replenishes power to avoid delays caused by waiting for natural recovery; when temperature exceeds limits, it delays detection to avoid collecting invalid data under harsh conditions. This combination maximizes the regularity of the detection cycle while ensuring data quality.

[0137] In one example, applying a controlled pulse excitation of a preset duration to the energy storage device includes: Within a preset duration, controlled pulse excitation is applied multiple times at preset rest intervals. The controlled pulse excitation is a constant power charging pulse.

[0138] It should be explained that the energy storage device in the energy-saving elevator system is subjected to short-term, medium-to-high rate non-periodic pulse charging and discharging during actual operation. Therefore, the excitation applied for health monitoring should simulate or cover the typical stress characteristics in actual operating conditions as much as possible, while also being controllable and repeatable. A single pulse excitation may produce errors due to accidental factors (such as instantaneous grid fluctuations or sampling time deviations). By repeatedly applying the excitation and setting appropriate rest intervals, the robustness and reliability of feature extraction can be effectively improved. In addition, the constant power mode is chosen instead of the constant current mode because the actual energy feedback and release process of the elevator system is usually executed in the form of power commands, and the constant power pulse can more realistically reflect the thermal effects and polarization characteristics of the battery under different health states.

[0139] In constant power mode, the power remains constant, while the current automatically decreases as the voltage increases. Energy feedback in actual elevator systems typically manifests as power feedback (e.g., the motor's feedback power is essentially constant during regenerative braking), therefore, constant power pulses more closely reflect real-world operating conditions. Simultaneously, constant power pulses are more sensitive to changes in battery internal resistance: when the battery's internal resistance increases, the initial current will be larger at the same power, resulting in a more significant difference in voltage response. Under constant power charging, the current automatically increases when the battery terminal voltage is low and automatically decreases when the terminal voltage is high. This is equivalent to applying a larger current excitation to batteries with lower voltage (often those in poor health or with low SOC), thus amplifying their abnormal responses; and applying a smaller current to healthy batteries with higher voltage to avoid overvoltage. This "adaptive" characteristic makes the response differences of abnormal cells more pronounced.

[0140] The duration of a single pulse excitation should be sufficient to allow the battery voltage to transition from the initial transition to a stage of slow polarization increase, without significantly altering the state of charge (SOC).

[0141] The purpose of the resting interval is to allow the battery to fully relax after each pulse, that is, to allow the polarization voltage (including ohmic polarization, electrochemical polarization and concentration polarization) accumulated during the pulse charging process to gradually dissipate, so that the battery's terminal voltage returns to a level close to that before the pulse.

[0142] To reduce random errors in a single pulse, this embodiment repeatedly applies the same constant power charging pulse K times within a preset duration, where K is a positive integer greater than or equal to 2, typically three times. Specifically, after meeting the SOC and temperature conditions, the system executes according to the following timing sequence: First pulse: Apply a constant power charging pulse.

[0143] After the first pulse ends, a resting phase begins, with a resting interval of T_s.

[0144] Second pulse: Apply a constant power charging pulse with the same parameters as the first pulse.

[0145] After the second pulse ends, allow the pulse to rest for another T_s.

[0146] Third pulse: Apply a pulse with the same parameters.

[0147] After the third pulse ends, the pulse excitation for this detection cycle is applied.

[0148] In one example, the response time series data includes voltage data for both the pulse charging phase and the relaxation phase.

[0149] It's easy to understand that when a battery is excited by a constant-power charging pulse, its terminal voltage doesn't simply rise or fall monotonically over time; instead, it exhibits a complex process containing multiple characteristic sub-stages. Collecting data only from the pulse charging stage provides information related to ohmic internal resistance and rapid polarization, but it fails to reveal the diffusion behavior and relaxation characteristics during polarization dissipation. Conversely, collecting data only from the relaxation stage lacks information about the initial conditions for inducing polarization. Only by collecting and analyzing both stages as a complete response time series can the battery's health be fully characterized.

[0150] The pulse charging phase refers to the continuous time interval from the moment the controlled pulse excitation begins to be applied until the moment the pulse excitation stops. Let the pulse start time be t=0, and the pulse duration be... For example, 10 seconds), then the time interval corresponding to the pulse charging phase is [0, ... ].

[0151] During this phase, the energy storage converter injects electrical energy into the energy storage device at a constant power P0. For each battery cell, its terminal voltage V(t) exhibits a rapid jump at t=0 (the instant the pulse begins). This jump is mainly caused by the battery's ohmic internal resistance, as the instantaneous current flowing through the ohmic internal resistance generates a voltage drop (voltage rises during charging). Subsequently, during the pulse duration, the voltage continues to rise at a slower rate, a process primarily dominated by electrochemical polarization and concentration polarization: the insertion / extraction of lithium ions at the electrode surface requires overcoming activation energy, while diffusion within the electrode is limited by the concentration gradient, resulting in a slow voltage increase over time. Therefore, the voltage time series during the pulse charging phase contains the following key health information: Initial voltage transition: reflects the magnitude of the battery's equivalent ohmic internal resistance. An increase in internal resistance (such as due to electrode corrosion or poor connections) will lead to an increase in the transition.

[0152] Voltage rise rate: reflects the speed at which polarization is established. Increased polarization (such as due to SEI film thickening or loss of active material) will lead to an increased rise rate or nonlinear changes.

[0153] Voltage curvature or second-order difference: reflects the degree of accumulation of concentration polarization. When the diffusion of lithium ions in the solid phase is hindered, the voltage curve will show obvious curvature, and the curvature will increase.

[0154] Voltage change integral area: comprehensively reflects the total energy accumulated by polarization, and is correlated with irreversible capacity loss.

[0155] The relaxation phase refers to the period after the pulse excitation stops, during which the internal polarization of the battery gradually dissipates and the terminal voltage slowly decreases. Let t be the time when the pulse ends. The duration of the relaxation phase is (For example, 30 seconds), then the time interval corresponding to the relaxation phase is ( , + ].

[0156] In this stage, the external excitation has been removed, but the voltage generated by polarization inside the battery still exists. The polarization voltage consists of three parts: ohmic polarization (which disappears instantaneously after the excitation stops), electrochemical polarization (related to charge transfer processes on the electrode surface, with a small time constant, typically dissipating within seconds to tens of seconds), and concentration polarization (caused by the ion concentration gradient inside the electrode and in the electrolyte, with a larger time constant, potentially taking minutes or even longer to dissipate completely). During the relaxation stage, the terminal voltage gradually decreases from V(T_c), approaching the open-circuit voltage (or equilibrium potential). This decay process can typically be fitted using a multi-exponential model (e.g., the sum of two or three exponential terms), each corresponding to a relaxation process with a specific time constant.

[0157] The voltage time series during the relaxation phase contains the following key health information: relaxation voltage attenuation amplitude This reflects the total amount of accumulated polarization. The larger the attenuation amplitude, the more severe the polarization accumulated during the pulse process.

[0158] The relaxation time constant τ (obtained through multi-exponential fitting): a fast time constant (on the order of several seconds) is mainly related to electrochemical polarization, while a slow time constant (on the order of tens to hundreds of seconds) is mainly related to concentration diffusion. An increase or decrease in the time constant can indicate obstruction of ion transport pathways or changes in electrode structure.

[0159] Relaxation residual energy: When a battery has latent faults such as micro-short circuits, abnormal self-discharge, or electrolyte decomposition, its relaxation behavior deviates from the standard exponential decay model, leading to a significant increase in the fitting residual. Therefore, the fitting residual energy is a highly sensitive indicator to latent anomalies.

[0160] From the perspective of the completeness of battery health management, the pulse charging phase and the relaxation phase are two inseparable aspects of the same electrochemical excitation-response process. From the rapid voltage transition and slow rise to relaxation decay, the dynamic behavior of the battery under electrical excitation is fully recorded, providing a rich data foundation for extracting multidimensional health characteristics. Data from the pulse charging phase is sensitive to internal resistance and rapid polarization, while data from the relaxation phase is sensitive to diffusion and latent faults (such as micro-short circuits). Combining the two can effectively distinguish different failure modes such as internal resistance anomalies, polarization anomalies, and relaxation anomalies.

[0161] In the second embodiment, the health management method further includes systematic preprocessing of the collected response time series data to eliminate interference from non-battery-related factors such as measurement noise, communication jitter, and operating condition differences on subsequent feature extraction and health assessment, thereby improving the reliability and comparability of the data. This preprocessing process specifically includes five steps: integrity verification, abnormal sampling point identification and correction, time alignment, stage segmentation, and baseline correction.

[0162] Health management methods also include: Perform integrity checks on the response time series data, including checks on the number of sampling points, the continuity of timestamps, and the consistency of pulse numbers; Abnormal sampling points are identified and corrected. Abnormal sampling point identification includes amplitude mutation detection and sliding window statistical detection, while abnormal sampling point correction includes linear interpolation or sliding window mean value replacement. Time alignment is performed on the response time series data, with the pulse start time as the time zero point; The response time series data is segmented into charging stage subsequences and relaxation stage subsequences. Baseline correction is performed on the response time series data, and the data is shifted using the voltage value at the pulse start time as the baseline.

[0163] It should be explained that the raw voltage time series data inevitably suffers from various interferences in real engineering environments: sampling circuits may generate instantaneous spikes due to electromagnetic interference; communication transmission may experience packet loss leading to missing sampling points; data acquisition channels for different battery cells may have slight clock skew; and the starting voltage of the battery may differ between different detection cycles (due to differences in SOC or historical state). If these interferences are used directly for feature extraction without processing, they will seriously affect the authenticity and comparability of health features, and may even lead to misjudgments in anomaly identification. Therefore, this embodiment provides a systematic preprocessing workflow, including integrity verification, anomaly sampling point identification and correction, time alignment, stage segmentation, and baseline correction.

[0164] The purpose of integrity verification is to confirm whether the voltage-time series acquired for each battery cell under each pulse excitation meets the basic quality requirements for subsequent analysis. This embodiment defines three verification dimensions: (1) Verification of the number of sampling points. For each pulse, the theoretical number of sampling points is ( + The system checks if the actual number of received sampling points equals the theoretical value. If it is less than the theoretical value, it indicates data loss or acquisition interruption, and the time series is marked as an "incomplete sequence". Incomplete sequences can be handled in two ways: if the number of missing sampling points is small (e.g., less than 5% of the total points), interpolation can be attempted to complete the sequence; if the number of missing points is large, the data for that pulse is directly discarded and not used for subsequent feature extraction.

[0165] (2) Timestamp continuity verification. The system checks whether the timestamp interval between adjacent sampling points is equal to or approximately equal to the preset sampling period Δt (e.g., 1 second). Due to clock drift or communication delay, the actual interval may deviate slightly. A tolerance range is set, for example, Δt ± 10%. When the time interval between adjacent sampling points exceeds this tolerance range, the sampling point is marked as a "time-abnormal sampling point". If the time interval of multiple consecutive sampling points is abnormal, it may indicate a fault in the acquisition system, which should be recorded and a manual check should be triggered.

[0166] (3) Pulse Number Consistency Verification. Multiple pulses are executed within a detection cycle (e.g., K=3 times). Each pulse's data should include a pulse sequence number during encapsulation. The system verifies whether the pulse numbers corresponding to each time series are consecutive and consistent with the preset order to avoid feature extraction errors caused by data aliasing (e.g., incorrectly including some data from the first pulse in the second pulse). If a pulse number jump or duplication is found, the pulse data should be marked as abnormal and an attempt should be made to realign it. If this cannot be corrected, the pulse should be discarded.

[0167] The output of the integrity check is an integrity flag, which is used in subsequent steps to select, remove, or mark data that requires manual review.

[0168] For time series that pass integrity checks (or cases where the number of sampling points is complete but local outliers exist), it is necessary to identify and correct abnormal sampling points caused by electromagnetic interference, sampling circuit noise, transient poor contact, etc. This embodiment provides two complementary anomaly detection methods: (1) Amplitude change detection. Calculate the voltage difference between adjacent sampling points. Since the battery voltage changes continuously under pulse excitation, the rate of voltage change between adjacent sampling points has a physical upper limit (e.g., limited by the battery polarization rate and sampling rate). A threshold value δ_max for amplitude abrupt change is set, typically 2-3 times the normal maximum rate of change (e.g., 50mV / Δt). When >δ_max, the sampling points This point is identified as an anomaly. Amplitude change detection is highly sensitive to spike noise and can quickly identify isolated, large-amplitude jumps.

[0169] (2) Sliding window statistical detection. This method is suitable for identifying outliers that are relatively flat but deviate from local trends. A sliding window of width W (W is an odd number, such as 5 or 7) is constructed on the time series, with the current sampling point to be detected at the center of the window. The median M and median absolute deviation (MAD) of all sampling points within the window are calculated. If the current sampling point... If the deviation from the median M exceeds a preset percentage threshold (e.g., 3 times the MAD or 30% of M), it is marked as an outlier. The sliding window statistical detection is well-adapted to local trends and will not misjudge normal trends (such as voltage increases) as abnormal.

[0170] Once a sampling point is marked as abnormal, it can be corrected using any of the following methods: Linear interpolation: This method uses the two nearest normal sampling points before and after the outlier to obtain a replacement value through linear interpolation over time. It is suitable for isolated, discontinuous outliers.

[0171] The sliding window median substitution method directly replaces outliers with the median of normal sampling points within the sliding window. This method is more robust to outliers occurring in clusters (such as several consecutive outliers caused by brief communication interruptions).

[0172] After the correction is completed, the system records the original value of the anomaly and the correction method so that data quality can be traced later.

[0173] Because there may be slight clock skew in the data acquisition channels of different battery cells, or the actual start time of pulse excitation may deviate from the theoretical time (e.g., relay action delay), it is necessary to align the sequences of all cells in the time dimension.

[0174] (1) Pulse start point alignment. The actual pulse charging start time detected for each battery cell is taken as its time zero point. This can be achieved by detecting the voltage rising edge: when the voltage rise rate exceeds a preset threshold (e.g., more than 5 times the steady-state rise rate) within several consecutive sampling points, it is determined as the start of a pulse. The time series of all cells are resampled or indexed and aligned according to their actual start times, so that the aligned sequences all have their respective pulse start points at t=0. This process ensures that different cells correspond to the same physical stage under the same time index (e.g., t=1 second represents 1 second after the pulse start), avoiding feature misalignment caused by clock asynchrony.

[0175] (2) Stage segmentation. Based on the preset pulse parameters ( , The aligned time series is divided into two subsequences: the charging phase subsequence, corresponding to the time interval [0, ... ]; Relaxation phase subsequence, corresponding time interval ( , + After segmentation, subsequent feature extraction will be performed separately for the two sub-sequences. Because the battery characteristics reflected in the two stages are different, different feature extraction models are required (e.g., focusing on initial transitions and rise rates during the charging stage, and fitting the decay curve during the relaxation stage). The accuracy of the segmentation points depends on the precise alignment of the pulse start points.

[0176] After time alignment and stage segmentation are completed, a significant difference still exists in the voltage time series of different battery cells: the static voltage of each cell before the pulse begins. The results may differ, possibly due to subtle differences in SOC, varying historical charge / discharge states, or variations in battery self-discharge rates. If absolute voltage values ​​are used directly for feature extraction... Differences in [the data] can mask the true differences in dynamic response. For example, a A cell with a high voltage level will have a higher absolute voltage throughout the charging process, but this does not necessarily indicate better health. Therefore, baseline correction is necessary.

[0177] The specific method for baseline correction is as follows: for the voltage time series of each battery cell, the voltage value at the start of its pulse is used as the baseline. Using this as a baseline, subtracting the baseline from the entire sequence yields the relative voltage change sequence. .

[0178] After baseline correction, the relative voltage of all cells at t=0 is 0. This process ensures that the characteristics primarily reflect the dynamic response of the battery to pulse excitation, rather than the static voltage level, thereby eliminating baseline bias introduced by differences in SOC or historical state.

[0179] Optionally, the corrected sequence can be normalized in magnitude, for example, by dividing by the maximum value of the sequence or by the population mean, to further eliminate the influence of differences in nominal voltage between different cells. However, baseline correction is the most critical process, and magnitude normalization is optional.

[0180] To achieve the best results, the above preprocessing steps should be performed in the following order: 1. Integrity verification (number of sampling points, timestamp continuity, pulse number consistency). Unqualified sequences are marked and directly removed or enter the repair process.

[0181] 2. Abnormal sampling point identification and correction (amplitude mutation detection, sliding window detection, and replacement with linear interpolation or median). The corrected sequence proceeds to the next step.

[0182] 3. Pulse start point alignment (detect the actual start time, resample or index alignment).

[0183] 4. Stage segmentation (divided into charging stage subsequence and relaxation stage subsequence).

[0184] 5. Baseline correction (subtracting pulse initiation voltage).

[0185] After the above preprocessing, the resulting relative voltage change sequence will be used as input for feature extraction. This sequence has the following characteristics: no missing values ​​or anomalous jumps, consistent time indices, zero baseline, and clear phases. These characteristics are prerequisites for the accurate and comparable extraction of subsequent multidimensional health features.

[0186] In one example, the health management method also includes: The sampling points that are identified as abnormal and corrected are recorded. The recorded information includes the battery cell number, pulse number, timestamp, original voltage value, abnormality judgment type and correction method. When the number of abnormal sampling points exceeds a preset threshold or when abnormal sampling points are continuously distributed, an abnormal notification event is generated and sent to the analyst's terminal.

[0187] It's easy to understand that the occurrence of anomaly sampling points is itself an important event: it may stem from a genuine anomaly in the battery itself (e.g., a drastic voltage jump may indicate a loose internal connection), or it may originate from a problem in the data acquisition link (such as poor contact in the sampling line, electromagnetic interference, or packet loss). If all corrected anomalies are indiscriminately treated as normal preprocessing and ignored, two types of critical information may be lost: first, the health status information of the sampling link itself (sampling line failures require separate repair); and second, information on intermittent, drastic fluctuations in certain battery characteristics. Therefore, this example proposes detailed recording and statistical analysis of anomaly sampling points, triggering manual notifications when preset conditions are met, so that analysts can determine the root cause of the anomaly and take appropriate measures.

[0188] For each identified and corrected anomaly sampling point, the system records the following information to form a structured anomaly log: Battery cell number: Used to pinpoint which specific battery cell is experiencing the abnormal sampling point.

[0189] Pulse number: Indicates which pulse (e.g., the 1st, 2nd, or 3rd pulse) the anomaly appears on in the current detection cycle.

[0190] The timestamps corresponding to the abnormal sampling points are accurate to the sampling time so as to locate them in the time series.

[0191] Original sampled voltage value: The original measurement value before correction, retained for manual verification and fault analysis.

[0192] Anomaly detection type: Record the specific reason for triggering the anomaly flag, such as "amplitude change" or "sliding window outlier".

[0193] Correction method used: Record the actual correction method used, such as "linear interpolation" or "sliding window midpoint substitution".

[0194] This information should be persistently stored in the system database, which can be used for both real-time manual notification and judgment, as well as for subsequent data quality trend analysis. For example, if a battery cell repeatedly shows abnormal sampling points in multiple consecutive testing cycles, and the abnormality type is always "amplitude change", then it is highly suspected that there is a poor contact in the voltage sampling line of that cell.

[0195] After preprocessing all pulse data in each detection cycle, the system performs statistical analysis on the recorded abnormal sampling points, mainly examining them from two dimensions: (1) Statistics on the number of abnormal sampling points. The total number of abnormal sampling points identified and corrected for each battery cell within the detection cycle is counted and denoted as M_i. At the same time, the total number of abnormal sampling points for the entire battery pack and the abnormal distribution of each pulse can be counted. The absolute size of the number reflects the overall level of data quality.

[0196] (2) Temporal distribution characteristics of abnormal sampling points. Check whether the abnormal sampling points show a "continuous distribution" in the time series. Continuous distribution means that abnormal sampling points appear consecutively on the time index (e.g., more than 3 consecutive sampling points are marked as abnormal), or appear densely within a short time window (e.g., within 1 second). Continuous distribution of abnormalities is usually different from isolated random spikes: isolated spikes are often caused by electromagnetic interference, while continuous abnormalities may mean that the sampling channel has completely failed within that time period (e.g., signal line disconnection, amplifier saturation), or that the battery itself has experienced severe and abnormal voltage fluctuations (e.g., a micro-short circuit momentarily pulls down the voltage). Therefore, "continuous distribution" needs to be used as an independent judgment condition.

[0197] Specifically, a threshold value L_cont for the length of a continuous anomaly window can be defined (e.g., L_cont=3). If L_cont or more consecutive sampling points in the time series are all marked as anomalies, it is determined that "the anomaly sampling points show a continuous distribution". Alternatively, the proportion of anomalies in any 5 consecutive sampling points exceeding 80% can be defined as another criterion for continuous distribution.

[0198] To promptly alert analysts in cases of severe data quality degradation or potential drastic battery malfunctions, this embodiment sets the following triggering conditions; an anomaly notification event is generated when any one of these conditions is met: (1) The number of abnormal sampling points exceeds the preset threshold. For each battery cell, a maximum allowable number of abnormal sampling points θ_M is set in a single cycle. If the total number of abnormal sampling points M_i of a certain cell in a single detection cycle is greater than θ_M, ​​a notification is triggered. This condition is used to capture cells with poor overall data quality.

[0199] (2) Abnormal sampling points exhibit a continuous distribution. Regardless of whether the total number of abnormal points exceeds the threshold, a notification is triggered as long as L_cont or more consecutive abnormal sampling points are detected (e.g., 3 consecutive sampling points are all abnormal). Continuous abnormalities are often more serious than sporadic abnormalities: sporadic abnormalities can be effectively corrected by interpolation and have little impact on feature extraction; while continuous abnormalities mean that a segment of data is almost completely unusable, and even interpolation will introduce a large error, requiring manual intervention to determine whether the segment of data can be discarded or reflects a true battery abnormality.

[0200] (3) Persistent anomalies across cycles. If the number of abnormal sampling points exceeds the threshold or is continuously distributed in multiple consecutive detection cycles (e.g., 3 consecutive cycles) for the same battery cell, a notification should be generated again, even if no notification is triggered individually in each cycle (or the issue has been resolved after being triggered), to indicate that there may be a chronic fault in the sampling link of that battery cell.

[0201] When any of the above triggering conditions are met, the system generates an exception notification event. This event must contain at least the following information: Time of occurrence (detection cycle identifier); The battery cell numbers involved; Triggering condition type (quantity exceeds threshold / continuous distribution / sustainable across periods); Statistical summary (total number of outliers, maximum length of consecutive outliers, etc.); Access links or key data fragments in the exception log.

[0202] Then, the anomaly notification is sent to the analyst's terminal via the communication interface of the cloud service platform or local monitoring system. The terminal can take the form of a mobile app push notification, SMS, email, or a monitoring system pop-up. After receiving the notification, the analyst can log in to the system to view detailed anomaly logs, the original voltage waveform, and the corrected waveform, thereby determining the root cause of the anomaly.

[0203] In one example, the health management method also includes: When multiple pulse excitations occur within the same detection cycle, a consistency check is performed on the pre-processed time series of the same battery cell under different pulses, and a morphological similarity index is calculated. When the morphological similarity index of a pulse is lower than the preset similarity threshold, the pulse data is marked as an abnormal pulse and removed.

[0204] It should be explained that within one detection cycle of this invention, the same controlled pulse excitation is typically executed multiple times (K times, e.g., 3 times), with a rest interval set after each pulse to ensure the independence of each pulse. Theoretically, for the same battery cell, under the same operating conditions (SOC, temperature) and the same excitation parameters, the voltage response time sequence of each pulse should have a high degree of consistency, i.e., the waveform shape, initial transition amount, rise rate, relaxation decay, and other characteristics should be basically repeated. However, in actual engineering, the data of a certain pulse may be abnormal due to the following reasons: instantaneous fluctuations in the power grid cause the actual pulse power to deviate from the set value; the sampling system experiences brief communication packet loss or clock jitter during the pulse; intermittent micro-short circuits or poor contacts occur inside the battery, which only manifest as abnormalities in this pulse; or external electromagnetic interference happens to be superimposed within the sampling window of this pulse. If the data containing abnormal pulses is directly used for feature extraction (e.g., taking the average of multiple pulse features), the abnormal pulses will contaminate the final statistical results, causing the health characteristics to deviate from the true value, thereby affecting the accuracy of anomaly identification. Therefore, this example introduces an inter-pulse consistency check mechanism. By calculating the morphological similarity of the time series of the same individual under different pulses after preprocessing, abnormal pulses are identified and eliminated, ensuring that the data used for subsequent feature extraction are all high-quality, highly consistent, and effective data.

[0205] Morphological similarity metrics are used to quantify the similarity of two time series in waveform shape, without relying on the magnitude of absolute amplitude (because amplitude differences may be caused by small residuals after baseline correction). This embodiment provides two optional similarity measurement methods: the correlation coefficient method and the dynamic time warping distance method.

[0206] 1. For the same battery cell, within the same detection cycle, after each pulse is preprocessed, the time series can be used to calculate the Pearson correlation coefficient by taking any two pulses; or, in order to comprehensively evaluate the consistency of each pulse, the average correlation coefficient of each pulse with all other pulses can be calculated, or the minimum correlation coefficient between all pulses can be calculated.

[0207] Specifically, the average similarity index of the k-th pulse is defined as: Where k and l are two randomly selected pulses, The Pearson correlation coefficient; The closer it is to 1, the higher the consistency between this pulse and other pulses; The smaller the value, the more likely the pulse is to be an abnormal pulse.

[0208] 2. Dynamic Time Warping Distance. For situations where there may be slight time axis offsets (e.g., residual offset after pulse start point alignment), Dynamic Time Warping (DTW) distance can be used as a similarity measure. DTW can calculate the minimum cumulative distance between two time series, allowing for non-linear alignment. After normalizing the DTW distance, the smaller the distance, the more similar the waveforms are.

[0209] A preset similarity threshold θ_sim is defined to determine whether a pulse is consistent with other pulses. The selection of this threshold is based on the following considerations: for multiple pulse responses of the same battery cell under the same conditions, due to random noise and minor fluctuations in operating conditions, the correlation coefficient is usually above 0.95; if the correlation coefficient is below 0.90, it indicates that the waveform morphology has changed significantly, and it is very likely to be an abnormal pulse.

[0210] If the average similarity index is used As a criterion, then when When <θ_sim, the k-th pulse is marked as an abnormal pulse.

[0211] The core principle of this embodiment is that the battery's health status does not change significantly within a short period (approximately 10 minutes between pulses within the same detection cycle). Therefore, the voltage response of each pulse should be highly repeatable. Any factor that causes a pulse response to be significantly inconsistent with other pulses, whether external (grid fluctuations, sampling faults) or internal (intermittent micro-short circuits, momentary loosening of connectors), is considered an "abnormal event." Removing these abnormal pulses prevents a single abnormal event from contaminating the overall health characteristic assessment.

[0212] The first embodiment achieves automatic identification of individual battery cell anomalies through comprehensive anomaly scoring and adaptive thresholds. However, any automated algorithm may misjudge or miss anomalies in complex field environments. For example, certain special operating conditions (such as grid frequency fluctuations or instantaneous load shocks) may cause abnormal voltage responses, leading to normal batteries being misjudged as abnormal; conversely, some early latent degradation may not have reached the threshold and thus be missed. Furthermore, the distribution characteristics of health features may differ among energy storage devices of different battery types and service lives, making it difficult to maintain optimal thresholds and weighting coefficients over a long period. The third embodiment introduces a manual judgment step: when the system detects a specific anomaly, it automatically presents a detailed diagnostic data package to analysts (such as maintenance engineers or battery experts), allowing for a final judgment based on more comprehensive information. The system dynamically adjusts the anomaly judgment threshold, the weighting coefficients in the comprehensive anomaly score, and corrects the group health representation space based on the manual feedback, thereby achieving continuous optimization of algorithm performance.

[0213] In the third embodiment, the health management method for the energy storage device in the energy-saving elevator system further includes: When an abnormal battery cell meets the preset triggering conditions, a prompt message is presented to the analyst. The prompt message includes the original voltage time series, the preprocessed time series, key health characteristic curves, the group distribution location, and the anomaly score and level change trend. Receive the manual judgment result corresponding to the prompt information, and modify the threshold for anomaly judgment or the weight coefficient in the comprehensive anomaly score based on the manual judgment result, and / or modify the population health representation space.

[0214] To achieve a balance between manual workload and system accuracy, this invention sets the following four trigger conditions, which initiate the manual judgment process when any one of them is met: (1) A single battery cell is identified as a Level 2 or Level 3 anomaly. Level 2 anomalies (significant degradation) and Level 3 anomalies (serious anomalies) have a significant impact on system safety and require manual confirmation to avoid unnecessary shutdowns caused by false alarms, while ensuring that genuine serious anomalies are dealt with in a timely manner.

[0215] (2) The same monomer is identified as a Level 1 anomaly within L consecutive detection cycles. Where L is a preset threshold for the number of consecutive cycles, typically 3. This condition is used to capture chronic degradation that does not deviate significantly in a single cycle but persists.

[0216] (3) Abnormal score S_i shows a sudden change within adjacent cycles. A sudden change refers to the overall abnormal score of the same battery cell between two adjacent detection cycles. The magnitude of the change exceeds the preset mutation threshold ,For example =0.5. A sudden change could indicate a sudden internal battery failure (such as a separator puncture or internal short circuit), or it could be caused by a sampling link failure or external interference. In either case, rapid human intervention is required for assessment.

[0217] (4) The system detects a sampling anomaly or a data quality anomaly. When the number of abnormal sampling points exceeds the threshold or shows a continuous distribution, the system will generate an anomaly notification event. At this time, even if no battery cell is judged to be abnormal, a manual judgment process should be triggered to distinguish whether it is a sampling link problem or a battery cell anomaly. If sampling anomalies and battery anomalies occur simultaneously, manual judgment can help determine whether there is a causal relationship (e.g., severe battery polarization causing sampling point jumps).

[0218] When any of the above triggering conditions are met, the system automatically constructs a complete diagnostic data package for the relevant battery cell (and optionally, a normal control group cell), and presents it to the analysts through the cloud service platform or local monitoring interface. This data package contains at least the following five pieces of information: (1) Raw voltage time series. This is the voltage value sequence read directly from the data acquisition module without any preprocessing, containing complete waveforms of the charging and relaxation phases. The raw data retains the most accurate measurement values, including noise, spikes, etc. Analysts can observe the raw waveforms to determine whether there are obvious measurement anomalies (such as jumps or breaks), or whether there are abnormal waveforms that conform to the physical laws of the battery (such as an abnormally rapid drop in the relaxation phase indicating a micro-short circuit).

[0219] (2) Preprocessed time series. This refers to the voltage relative change series after processing steps such as integrity verification, outlier correction, time alignment, and baseline correction. The preprocessed series eliminates obvious measurement noise and baseline shifts, and better reflects the intrinsic dynamic response of the battery. Analysts can compare the original series with the preprocessed series to determine whether the preprocessing process is reasonable (e.g., whether the correction of outliers is excessive).

[0220] (3) Key health characteristic curves: These curves show the changes in key health characteristics of the battery cell over the most recent testing periods, such as initial voltage transition, relaxation voltage decay amplitude, and singular energy index. Main mode distance And so on, while also indicating the range of the group mean.

[0221] (4) Population distribution location. In the constructed population health representation space, the system displays the projected locations of all battery cells in the form of a two-dimensional or three-dimensional scatter plot. Typically, the scores of the first two principal components are used as the horizontal and vertical axes, and color or symbol size is used to represent the distribution. The size of the anomalous individual. Analysts can visually see the relative position of the anomalous individual within the cluster: whether it is an isolated point far from the main cluster, or located at the edge of the cluster but... Larger, or located at the center of the cluster but Anomalies. This visualization helps determine the nature of the anomalies.

[0222] (5) Anomaly score and grade change trend. The system displays the comprehensive anomaly score of the battery cell in the past multiple detection cycles (e.g., the most recent 10 cycles) using a line graph. The chart also includes the corresponding anomaly level (normal, Level 1, Level 2, Level 3). Additionally, if there are consecutive confirmed anomalies, the number of consecutive periods is marked on the chart. This trend chart helps analysts determine whether the anomaly is sudden or gradual, and whether it has persisted for multiple periods.

[0223] The above information should be presented in the form of a web dashboard or a dedicated client, supporting functions such as zooming and hovering over data points to view details, so that analysts can make quick judgments.

[0224] Based on the aforementioned data package, and considering their professional knowledge and on-site conditions (such as recent maintenance or abnormal ambient temperature), analysts make a judgment decision. This embodiment defines the following four types of judgment results, from which analysts select one: (1) TruePositive Confirmation: The analyst believes that the anomaly identified by the system is real, that is, the battery cell does have a deterioration in health or a failure. In this case, the sample is used as a positive sample for subsequent model optimization.

[0225] (2) False Positive: The analyst believes that the abnormality identified by the system is actually a normal battery, and the false positive is due to fluctuations in operating conditions, measurement noise, or other factors not related to the battery itself. This sample is used as a negative sample to correct the threshold or weight, so as to reduce the false positive rate of similar situations in the future.

[0226] (3) Data anomalies (sampling / communication issues): Analysts determine that the root cause of the anomaly is a data acquisition link failure (such as poor contact of the sampling line, communication packet loss, sensor drift), rather than a problem with the battery itself. In this case, the data from that individual cell should not be used in the construction of the population health characterization space, and a hardware check should be triggered.

[0227] (4) Deferred confirmation (continued observation required): If the analyst is temporarily unable to determine whether it is a real anomaly, it may be due to insufficient information or a slight degree of anomaly. It is recommended to continue to observe subsequent cycles without making immediate corrections.

[0228] The system receives analysts' selections through a human-computer interaction interface (such as buttons and drop-down menus), and also allows analysts to add text notes. The manual judgment results, along with the data for the corresponding period, anomaly scores, and other information, are stored in the database as input for subsequent model corrections.

[0229] It's important to clarify that when an analyst classifies a battery cell as a "TruePositive," the singular energy index SPSPE_i and principal mode distance D_i for that sample are considered representative values ​​of a genuine anomaly. Conversely, when an analyst classifies a cell as a "FalsePositive," the corresponding indices for that sample, although exceeding the automatic threshold, are actually within the normal range, indicating that the current threshold may be set too low.

[0230] Based on the received judgment result, the system automatically performs one or more of the following correction operations: (1) Adjust the adaptive anomaly threshold and distance threshold based on the set of abnormal samples confirmed by humans.

[0231] The abnormal threshold θ_E is corrected based on a set of manually confirmed abnormal samples.

[0232] ,in, The corrected anomaly threshold. The outlier threshold before correction. Let β be the average singular energy of manually confirmed anomalies, and β be the smoothing coefficient (ranging from 0 to 1). This formula makes the threshold converge towards the distribution center of true anomalies, reducing false positives. Similarly, the principal mode distance threshold θ_D can be corrected in the same way.

[0233] (2) Update the weight coefficients in the comprehensive anomaly score based on the results of manual judgment.

[0234] The comprehensive anomaly score is defined as follows: ,in, and Weighting coefficients ( 1+ 2=1).

[0235] Based on the results of manual judgment, the weights are updated according to the formula: Where η is the learning rate. The weights before the update. For the updated weights, i,t represents the confidence score added by the analyst, and Δj represents the misjudgment direction correction term (e.g., reducing it if the proportion of SPSPE_i exceeding the threshold in the misjudged samples is too high). 1). Renormalize after adjustment.

[0236] (3) For abnormal battery cells that are manually identified, their weight is reduced or they are removed in the subsequent construction of the population health characterization space.

[0237] For manually identified abnormal battery cells or data anomalies, their weights are reduced or they are temporarily removed when constructing the population covariance matrix. The formula for calculating the weighted covariance matrix is ​​as follows: , Among them, normal monomers ; Abnormal monomers Abnormal data samples were directly removed. 0). This correction ensures that the population health characterization space primarily reflects the characteristic distribution of truly healthy individuals, avoiding contamination of the benchmark by anomalous samples.

[0238] (4) Data anomalies and model isolation Samples manually identified as "data anomalies" are not used in model training (i.e., not for anomaly threshold updates and population health representation space construction), but are recorded separately and have their sampling lines or communication checked. When the proportion of data anomalies exceeds a preset threshold, model updates are paused (model freeze mechanism), and only monitoring output is performed. When the proportion of data anomalies exceeds the preset threshold, it indicates that the overall data quality of the current detection period or several recent periods is poor, which may be due to communication network failure, large-scale anomalies in the acquisition system, or strong environmental interference.

[0239] This embodiment achieves the following beneficial effects by introducing a closed loop of manual judgment and model correction: First, it significantly reduces the false alarm rate and false negative rate, enabling the system to adapt to different battery types and operating conditions; Second, it enhances the credibility of diagnostic results by feeding back the "gold standard" confirmed by humans to the model, allowing the model to continuously approach expert experience; Third, the population health characterization space achieves self-purification by eliminating abnormal samples, ensuring the purity of the benchmark.

[0240] In the fourth embodiment, the relative health state parameter ranges from [0,1]. The relative health state parameter for each battery cell is calculated based on the deviation, including: The normalized principal modality distance is mapped to the principal health component; Calculate the exotic health component based on exotic energy index and abnormal threshold; A stability correction factor is introduced, which is calculated based on the number of consecutive abnormal periods. The relative health state parameters of a single battery cell are obtained by weighting and combining the main health component, the singular health component, and the stability correction factor.

[0241] It should be explained that in actual operation and maintenance, it is not only necessary to know which individual cells are abnormal, but also to quantify the health level of each individual cell (e.g., health level decreases from 1 to 0) in order to perform health trend prediction, remaining life estimation, and overall battery pack performance evaluation. However, in energy-saving elevator systems, due to the lack of manufacturer-calibrated SOH curves and standard test conditions, it is impossible to directly calculate the absolute health state. Therefore, this embodiment adopts a group self-referenced relative health state parameter, whose value range is defined as [0,1], where 1 indicates that the cell has the best health level in the current group, and 0 indicates a severely degraded or failed state. This relative health state parameter is composed of a weighted combination of three components: the main health component (reflecting the cell's position in the main health subspace), the singular health component (reflecting the degree of deviation of the cell in the singular deviation subspace), and the stability correction factor (reflecting the penalty for the duration of the abnormality).

[0242] The relative health status parameter of battery cell i in the current testing cycle is defined as SOH_i, with a value range of [0,1]. SOH_i=1 indicates that the health status of this cell is at the optimal level in the group (e.g., the minimum dominant mode distance and the lowest singular energy index), while SOH_i=0 indicates that the cell has severely degraded or failed (e.g., the dominant mode distance has reached the historical maximum and the singular energy index far exceeds the threshold). It is important to emphasize that SOH_i is not an absolute health status relative to the manufacturer's rated capacity in the traditional sense, but rather a reference level relative to the healthy batteries in the current group. This relative definition method does not rely on external calibration data and can adaptively track the gradual aging of the entire battery pack.

[0243] The master health component is used to quantify the deviation of a single cell's position in the master health subspace from the population health center. (Main mode distance) This represents the distance from the projected coordinates of an individual in the main health subspace to the center of the population health. The smaller the value, the closer the overall health level of the individual is to the population center (i.e., the population average level), and the closer its main health component should be to 1. The larger the value, the further the overall health level of the individual is from the population center (which usually means performance degradation), and the smaller the main health component should be.

[0244] Mapping formula for the main health component: ,in, The maximum healthy distance within the historical or current period. The distance to the primary health subspace.

[0245] The singular health component is used to quantify the degree of deviation of a single cell in the singular deviation subspace. A larger singular energy index (SPE_i) indicates that more health characteristics of the cell cannot be explained by the main health subspace of the population, i.e., abnormal behavior exists. The singular health component should decay exponentially with increasing SPE_i, so that the health component remains relatively high when SPE_i slightly exceeds a threshold, while rapidly approaching 0 when SPE_i far exceeds the threshold.

[0246] Mapping formula for unusual health components: ,in, For unusual health benefits, This refers to the abnormal threshold of the singular energy index.

[0247] In energy-saving elevator systems, deviations occurring within a single testing cycle may be caused by transient disturbances (such as temperature fluctuations or electromagnetic interference) rather than continuous degradation of the battery itself. If only a single cycle is considered... and Calculating health status parameters can lead to significant fluctuations in health values ​​between periods, which is detrimental to trend prediction. Therefore, this embodiment introduces a stability correction factor to impose an additional penalty on monomers that exhibit abnormalities for multiple consecutive periods, in order to reflect the cumulative impact of persistent abnormalities on health status.

[0248] The formula for the stability correction factor is: ,in, κ represents the number of consecutive abnormal cycles for battery cell i (i.e., the number of detection cycles in which it is consecutively judged as abnormal (e.g., level 1 or above) since the most recent judgment as normal); κ is the attenuation coefficient, a preset positive number (typically 0.1~0.5). The value range of this correction factor is (0,1). The more consecutive abnormal cycles, the smaller the correction factor, and the heavier the penalty to the health status. Even if a single cell deviates only slightly in one instance (SPE_i is close to...), the penalty will still apply. However, if deviations occur consecutively across multiple cycles, it indicates that the degradation is persistent, and the individual should receive a lower health score than a single, occasional deviation. Conversely, if an individual experiences a large single deviation but subsequently returns to normal (e.g., due to external disturbances), then... It will be reset to 0. Resetting to 1 will not result in continued penalties.

[0249] The three components are weighted and combined according to their respective weighting coefficients to obtain the relative health state parameters of cell i. The formula for calculating the SOH of a single cell is as follows: ,in, and These are the weighting coefficients. ,and All are greater than zero. Mainly healthy portions, For unusual health benefits. This is a stability correction factor.

[0250] The default values ​​for the weighting coefficients can be set according to the actual application scenario. In typical applications, the singular health component can be given a higher weight because it is more sensitive to early anomalies. Of course, the weighting coefficients can also be dynamically adjusted through manual judgment and model correction mechanisms. If the dominant modal distance in a system better reflects the actual health status, it can be increased. Conversely, it increases. .

[0251] Furthermore, in this formula, the main health components are summed independently, while the exotic health components need to be multiplied by a stability correction factor. Then sum them up. The reason for this design is that the main health component reflects the deviation of the overall health level, and even if the deviation is large in a single instance, it should be directly reflected in the health value; while the unusual health component reflects non-mainstream anomalies, and if it only occurs once by chance, it should not be over-punished, so a stability correction factor is used for discounting; but if it occurs continuously, the discount is more severe, and the health value drops significantly.

[0252] In one example, predicting the health trend of individual battery cells includes: Construct time series of relative health status parameters across multiple testing cycles; The degradation rate is estimated using linear regression or nonlinear fitting methods based on the time series of relative health state parameters. The health trend of individual battery cells is predicted based on the degradation rate and a preset failure threshold.

[0253] It should be explained that, based on the fourth embodiment, this example further defines a method for predicting the health trend of individual battery cells. The fourth embodiment provides the relative health state parameter SOH_i(t) (range [0,1]) of battery cell i at each detection cycle t. As the energy storage device continues to operate, the health state of individual battery cells usually shows a gradual downward trend, but the degradation rate may differ for different cells and at different stages. By performing time series analysis on SOH_i(t) over multiple detection cycles, the degradation rate can be estimated, and the future failure threshold can be predicted accordingly, thus providing a basis for predictive maintenance. This embodiment uses both linear regression and nonlinear fitting methods to estimate the degradation rate and calculate the remaining usable life (RUL).

[0254] For each battery cell i, the system calculates its relative health state parameter SOH_i(t) in each detection cycle t (t=1,2,3,…, in the order of detection cycle numbers). These values ​​arranged in chronological order form a time series: SOH_i(1), SOH_i(2), SOH_i(3),...... To reduce the impact of random fluctuations in a single detection, the time series can be smoothed (e.g., by moving average), but smoothing is not a necessary step. The alignment of the time series uses the actual calendar time or cumulative running time of the detection period as the horizontal axis. For simplicity, this embodiment uses the period number t as the time variable, which can be converted to real time (day or week) in practical applications.

[0255] In stages where health degradation exhibits an approximately linear trend (e.g., mid-life of a battery), a linear regression model can be used to estimate the degradation rate. Within a time window of length W, the method of estimating the degradation rate using linear regression is as follows: , where t is the detection cycle number (or the corresponding time). The regression coefficient represents the linear rate of decline in health status for each cycle (usually a negative value, indicating a decline in health status). The intercept is used. Linear regression employs the least squares method, fitting data based on the most recent W periods. The typical value for the sliding window length W is 5 to 10 periods; if W is too small, it becomes sensitive to noise, and if W is too large, it cannot respond quickly to changes in the degradation rate.

[0256] The larger the absolute value of the degradation rate a_i, the faster the decline in health status. When A value less than 0 indicates that the battery is gradually degrading; if... A value close to 0 or positive indicates a stable health status or fluctuations in measurement.

[0257] For some battery cells, the degradation of their health may exhibit a non-linear accelerating trend (e.g., a significant capacity drop at the end of their lifespan). In such cases, linear models cannot accurately describe the degradation pattern. Non-linear trend fitting methods can be used, such as the exponential model calculation formula: In this model, A, B, and λ are the parameters to be fitted. B represents the lower asymptote of the healthy state (i.e., the long-term stable value or failure level), A+B is the initial healthy value (at t=0), and λ is the degradation rate parameter (λ>0, the larger λ is, the faster the degradation). This model can capture the typical aging characteristics of a healthy state that first declines slowly and then accelerates. Parameter estimation can be performed using nonlinear least squares methods (such as the Levenberg-Marquardt algorithm). During fitting, it is necessary to ensure that the value of B is reasonable (usually B≥0) and that the fitting residuals are within an acceptable range. If the goodness of fit of the exponential model (e.g., R²) is significantly higher than that of the linear model, it indicates an accelerated degradation trend, and the exponential model should be preferred for prediction.

[0258] The failure threshold refers to the value at which the relative health status parameter of a battery cell drops to a preset value, at which point the battery is considered unable to meet the normal operation requirements of the energy storage device and needs to be replaced or repaired. In one example, the range of values ​​for the State of Health (SOH) failure threshold is: The specific value can be determined based on the application requirements, safety redundancy, and field experience of the energy storage device. For example, for applications with high requirements, SOH_fail can be set to 0.8; for applications where a certain degree of performance degradation is acceptable, 0.6 can be used. It should be noted that since the SOH of this invention is a relative value relative to the overall health level of the population, the failure threshold also needs to be calibrated in conjunction with historical data and actual operating results. Initially, 0.7 can be used as the default value, and subsequent adjustments can be made through a manual feedback mechanism.

[0259] Based on the degradation rate and current health status, the remaining number of detection cycles (or time) required for a battery cell to reach the failure threshold can be predicted, i.e., the Remaining Useful Life (RUL) can be calculated. The RUL estimation formula is: ,in, The health status parameter for the current detection period t. The linear degradation rate is negative (so the numerator and denominator are both negative, and RUL is positive). This formula means: assuming the future degradation rate remains constant, from the current healthy state... Drop to failure threshold The required number of cycles.

[0260] It should be noted that the uncertainty of the RUL estimate increases with the extension of the forecast period. The forecasting method for calculating the uncertainty assessment is as follows: δ is obtained from historical residual statistics. This uncertainty assessment can be used as an optional output to help operations personnel determine the reliability of the prediction.

[0261] It's easy to understand that linear regression is suitable for the mid-stage of a steady decline in health, as it is simple to calculate and has good stability; exponential fitting is suitable for the accelerated degradation stage at the end of life, and can more accurately predict the time of failure. The system can automatically select the optimal model based on the goodness of fit.

[0262] In the fifth embodiment, after acquiring the response time series data of each battery cell at a fixed sampling frequency, the method further includes: The collected response time series data is packaged according to a preset format; Check the communication status with the cloud service platform or remote analysis server; When communication is normal, the encapsulated response time series data is sent to the cloud service platform or remote analysis server through the communication interface; In the event of a communication failure, the data will be temporarily stored in the local storage unit and then resent.

[0263] It should be explained that this embodiment further defines the specific steps for data encapsulation, transmission, and anomaly handling after acquiring the response time series data of each battery cell at a fixed sampling frequency. Energy storage health management in energy-saving elevator systems typically involves a large number of battery cells (e.g., dozens or even hundreds of cells in a single battery box), with each cell generating multiple voltage time series data points under various pulses within each detection cycle. Performing complex feature extraction, group modeling, and anomaly identification on all this raw data in a local control unit (such as an elevator controller or energy storage converter) would significantly increase the burden on local computing resources and make it difficult to achieve unified algorithm upgrades and maintenance. Therefore, this embodiment adopts a "local acquisition + cloud analysis" architecture: the local unit is only responsible for data acquisition, preliminary encapsulation, and transmission, uploading all raw time series data to a cloud service platform or remote analysis server; the cloud utilizes its powerful computing capabilities and storage resources to perform feature extraction, group modeling, anomaly identification, and health assessment. This architecture not only reduces the computational pressure on the local control unit but also ensures the scalability and consistency of the algorithm.

[0264] After acquiring data from all K pulse excitations within a single detection cycle (i.e., acquiring the voltage time series of all battery cells during the charging and relaxation phases), the system first encapsulates the raw data. The purpose of encapsulation is to organize the discrete sampled data into a structured format, facilitating cloud parsing and storage. The encapsulated data format must contain at least the following information: 1. Battery cell number: Used to uniquely identify each battery cell in the energy storage device (e.g., B001, B002, ..., B0N).

[0265] 2. Pulse number: Indicates which pulse in the current detection cycle this data belongs to (e.g., 1, 2, ..., K).

[0266] 3. Sampling timestamp: The absolute or relative time (relative to the pulse start time) of each sampling point, used for time axis alignment.

[0267] 4. Corresponding sampling voltage value: The terminal voltage value measured at each sampling point (the unit is usually millivolt or volt).

[0268] 5. Current testing cycle identification information: such as the serial number or date and time of the testing cycle, used to distinguish different batches of testing data.

[0269] Encapsulation can use common data exchange formats such as JSON, CSV, or Protobuf. The encapsulation process is performed in a local control unit (such as the energy management unit of an elevator system or a dedicated data acquisition terminal).

[0270] After encapsulation, the system checks the communication status with the cloud service platform or remote analysis server via wired or wireless communication modules (e.g., Ethernet, 4G / 5G, Wi-Fi). A specific check method could be to attempt to send a short handshake signal (e.g., Ping or HTTPHEAD request) to the server's preset communication interface and wait for a response. If a normal response is received from the server within the timeout period (e.g., 5 seconds), communication is considered normal; otherwise, communication is considered abnormal. When communication is normal, the system calls the preset communication interface (e.g., HTTPPOST, MQTTPUBLISH) and sends the encapsulated data as the message body to the cloud service platform or remote analysis server. A reliable transmission protocol (e.g., TCP) should be used during transmission, and the transmission timestamp should be recorded.

[0271] To ensure reliable data delivery, this embodiment introduces a send acknowledgment and local caching mechanism. After successfully receiving and parsing the data, the cloud service platform should return an acknowledgment (ACK) to the local unit. Upon receiving the acknowledgment, the local unit considers the data to have been successfully delivered and clears the detection data from its local cache.

[0272] Communication is considered abnormal or transmission has failed if any of the following occurs: failure to establish a connection with the server; connection timeout; server returns an error status code (such as 500 Internal Error); or no acknowledgment is received within the specified time. In this case, the local unit does not delete the data, but temporarily stores the encapsulated data in the local storage unit. The local storage unit can be a built-in flash memory, SD card, or external USB flash drive, and its capacity should be sufficient to store data for at least several detection cycles (e.g., storing data from the most recent 10 detection cycles in case of prolonged network outages).

[0273] For temporarily stored data, the system needs to resend it at a later time. The resending strategy can employ exponential backoff or fixed-interval retries: for example, retrying every 5 minutes, extending the retry interval to 30 minutes after 3 consecutive failures, until successful transmission or the maximum number of retries (e.g., 10). Once communication is restored and the data has been successfully sent and acknowledged, it is then deleted from local storage. If local storage space is about to run out (e.g., remaining space less than 10%), the system should trigger an alarm, prompting maintenance personnel to check the communication link or manually export the data.

[0274] After the data is successfully sent to the cloud service platform, the cloud will perform steps such as preprocessing, feature extraction, group modeling, anomaly identification, and health assessment. After the cloud completes the analysis, it can push the results (such as a list of abnormal individuals, health status parameters, and trend predictions) back to the local monitoring system or the terminal of the operation and maintenance personnel, forming a complete health management closed loop.

[0275] This invention also proposes a health management system based on an energy storage device in an energy-saving elevator system, comprising: A controlled excitation unit is used to apply a controlled pulse excitation of a preset duration to the energy storage device; The data acquisition unit is used to acquire the response time series data of each battery cell at a fixed sampling frequency during the controlled pulse excitation process and the relaxation phase after the excitation ends. The data preprocessing unit is used to preprocess the response time series data to remove outliers and standardize and align the data. The feature extraction and modeling unit is used to extract multidimensional health features from the preprocessed response time series data to characterize the battery health status, and to construct a group health representation space as a reference benchmark for the group health status based on the multidimensional health features of all battery cells in the same battery pack. An anomaly identification unit is used to calculate the deviation of the health characteristics of each battery cell in the group health characterization space, and to identify abnormal battery cells based on the deviation. The health assessment unit is used to calculate the relative health status parameters of each battery cell based on the deviation, and to predict the health trend of the battery cells by combining the health status parameters of historical testing cycles. Each unit is used to implement a health management method for energy storage devices in an energy-saving elevator system.

[0276] It should be explained that energy storage devices consist of multiple battery cells, typically arranged in series or a series-to-parallel configuration to form battery boxes or battery packs. This system can be deployed in an architecture combining a local control unit and a cloud service platform, where some units (such as controlled excitation units and data acquisition units) are located locally, while others (such as data preprocessing units, feature extraction and modeling units, anomaly detection units, and health assessment units) are located on the cloud service platform; alternatively, it can be deployed entirely in a local controller with sufficient computing power. The functions of each unit are described in detail below: The controlled excitation unit applies controlled pulse excitation of a preset duration to the energy storage device. This unit communicates with the energy management unit or energy storage converter of the elevator system. When the elevator system is not in operation and preset operating conditions are met (such as SOC reaching or exceeding a preset threshold, and battery temperature within allowable range), it outputs constant power charging pulses according to set parameters. The controlled excitation unit supports multiple pulse application: applying constant power charging pulses multiple times within a preset duration at preset rest intervals. The power value, duration, and rest interval between adjacent pulses are all configurable. This unit also has a linkage function with the operating condition detection module: when SOC is insufficient, it first instructs the energy storage converter to replenish the energy storage device; when the temperature exceeds the allowable range, it delays or suspends the excitation application.

[0277] The data acquisition unit is used to acquire the terminal voltage response time series data of each battery cell at a fixed sampling frequency (e.g., once per second) during the controlled pulse excitation process and the relaxation phase after excitation. This unit typically includes multiple voltage sampling channels (e.g., an analog front-end chip (AFE), an analog-to-digital converter, and a sampling timing controller. The data acquisition unit can synchronously acquire the voltage of all battery cells and associate the sampled values ​​with the corresponding timestamp, pulse number, and battery cell number to form the original data record. The acquisition time range at least covers the pulse charging phase [0, ...]. and relaxation phase ( , + This ensures the integrity of the data required for subsequent feature extraction. After acquisition, the data acquisition unit passes the raw data to the data encapsulation module (which can be integrated with the data acquisition unit or operate independently) for encapsulation and transmission.

[0278] The data preprocessing unit preprocesses the response time series data to remove outliers and standardize and align the data. This unit receives raw voltage time series data from the data acquisition unit and performs preprocessing operations, including: integrity verification (sampling point count verification, timestamp continuity verification, pulse number consistency verification), outlier sampling point identification and correction (amplitude mutation detection, sliding window statistical detection, and linear interpolation or sliding window midpoint replacement), time alignment (using the pulse start time as the time zero point), stage segmentation (dividing into charging stage subsequences and relaxation stage subsequences), and baseline correction (shifting based on the pulse start voltage value). The preprocessed data has a consistent time index, a unified baseline (relative voltage change value), and clear stage segmentation, providing high-quality input for subsequent feature extraction. In addition, the data preprocessing unit can also perform outlier sampling point recording and notification functions, as well as inter-pulse consistency checks and outlier pulse removal functions.

[0279] The feature extraction and modeling unit is used to extract multidimensional health features from the preprocessed response time series data to characterize the battery health status, and to construct a group health representation space as a reference benchmark for the group health status based on the multidimensional health features of all battery cells in the same battery pack.

[0280] In terms of feature extraction, this unit extracts multidimensional features from the charging and relaxation stages, including but not limited to: initial voltage transition magnitude, initial voltage rise rate, maximum slope and slope dispersion, voltage curvature, relaxation voltage decay amplitude, relaxation time constant (fitted by a multi-exponential model), relaxation fitting residual energy, integral features, repetitive pulse stability features (mean, standard deviation, coefficient of variation), and population relative deviation features. The extracted features are combined to form a health feature vector for each battery cell.

[0281] In terms of population modeling, this unit constructs a population health representation space based on the health feature vectors of all individuals. Specifically, this includes: standardizing the feature matrix, calculating the covariance matrix, performing eigenvalue decomposition, selecting the top K principal components whose cumulative contribution rate reaches a preset threshold to form the principal health subspace, and the remaining feature vectors forming the singular deviation subspace. Simultaneously, this unit supports an online update method, periodically updating the covariance matrix using a sliding time window weighted update approach (forgetting factor α) to suppress the impact of short-term anomalies on the baseline.

[0282] The anomaly detection unit calculates the deviation of each cell's health characteristics from the population health characterization space and identifies anomalous cells based on this deviation. This unit executes the anomaly detection method: calculating the singular energy index for each cell. Distance to the dominant mode Determining adaptive anomaly thresholds based on population distribution and Construct a comprehensive anomaly score When S_i > 1, it is determined to be an abnormal single entity. Simultaneously, this unit supports anomaly grading and continuity confirmation mechanisms: based on anomaly scores or the degree of singular energy exceeding limits, anomalies are classified into Level 1 (mild deviation), Level 2 (significant degradation), and Level 3 (severe anomaly), and it determines whether the same single entity remains abnormal over multiple consecutive detection cycles to confirm stable anomalies. Furthermore, the anomaly identification unit can also execute anomaly cause tracing methods: calculating the contribution of each feature component of the abnormal single entity in the singular deviation subspace, mapping the anomaly to internal resistance anomaly, polarization anomaly, relaxation anomaly, or consistency degradation type.

[0283] The health assessment unit calculates the relative health state parameters for each battery cell based on deviation and predicts the health trend of the battery cells by combining health state parameters from historical testing cycles. This unit performs the relative health state parameter calculation method: [The text abruptly shifts to a different topic] After normalization, the mapping is based on the primary health component. , will strange energy index The SOH_i is obtained by mapping the SOH_i to a singular health component using an exponential function, introducing a stability correction factor, and weighting the components. This unit also performs a health trend prediction method: constructing a SOH_i(t) time series, estimating the degradation rate through linear regression or exponential fitting, and calculating the remaining usable life (RUL_i) based on a preset failure threshold SOH_fail (values ​​ranging from 0.6 to 0.8). The output of the health assessment unit includes the current SOH, degradation rate, remaining usable life, and overall health status of the battery pack for each individual cell.

[0284] The aforementioned units work collaboratively according to the following process: the controlled excitation unit triggers detection during off-peak hours, applying multiple constant power pulses; the data acquisition unit synchronously acquires the voltage time series of all battery cells and sends the packaged data to the cloud service platform; the data preprocessing unit cleans, aligns, and corrects the raw data; the feature extraction and modeling unit extracts multi-dimensional features and constructs a population health representation space; the anomaly identification unit calculates the deviation and identifies abnormal cells; and the health assessment unit calculates relative health state parameters and performs trend prediction. The system also supports manual interaction and model correction functions: when preset conditions are triggered, a diagnostic data package is presented to analysts, the results of manual judgment are received, and thresholds, weighting coefficients, and the population health representation space are adjusted accordingly.

[0285] Through the aforementioned system, this invention achieves fully automated, online, and external-reference-free health management of energy storage devices in energy-saving elevator systems. The system offers the following advantages: First, its modular design ensures clear functionality for each unit, facilitating independent upgrades and maintenance; second, it supports a deployment method combining local data acquisition and cloud analysis, reducing the computational requirements of on-site equipment; third, through a group self-reference and online update mechanism, it can adaptively track the gradual aging of battery packs; fourth, it integrates a manual judgment feedback loop, enabling continuous system optimization in complex on-site environments; and fifth, the output results include abnormal cell identification, health status quantification, and trend prediction, providing direct evidence for predictive maintenance.

[0286] The above are merely optional embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A health management method for an energy storage device in an energy-saving elevator system, characterized in that, The energy storage device includes multiple battery cells; the health management method includes: When the elevator system is in a non-operational period and meets the preset operating conditions, a controlled pulse excitation of preset duration is applied to the energy storage device. During the controlled pulse excitation process and the relaxation phase after the excitation ends, response time series data of each battery cell are collected at a fixed sampling frequency. Extract multidimensional health features to characterize the battery health status from the response time series data; Based on the multidimensional health characteristics of all individual cells within the same battery pack, a group health characterization space is constructed using multivariate statistical analysis methods to serve as a reference benchmark for the group's health status. Calculate the deviation of the health characteristics of each battery cell in the population health characterization space, and identify abnormal battery cells based on the deviation. Based on the deviation, the relative health status parameters of each battery cell are calculated, and the health trend of the battery cells is predicted by combining the health status parameters of historical detection cycles.

2. The health management method for energy storage devices in an energy-saving elevator system as described in claim 1, characterized in that, The population health representation space includes a main health subspace and a singular deviation subspace; The construction of a population health representation space as a benchmark for population health status using multivariate statistical analysis methods includes: The multidimensional health characteristics of all battery cells are used to construct a feature matrix and then standardized. Calculate the covariance matrix based on the standardized feature matrix; The covariance matrix is ​​subjected to eigenvalue decomposition, and principal component analysis is used to extract principal components that reflect the common health patterns of the group. The main health subspace and the singular deviation subspace are determined. The main health subspace is determined by the feature vectors corresponding to the top K principal components whose cumulative contribution rate reaches a preset threshold, and is used to characterize the common health pattern of the group. The singular deviation subspace is determined by the remaining feature vectors and is used to characterize the deviation of individual battery cells from the common pattern of the group. K is a positive integer. The calculation of the deviation of the health characteristics of each individual battery cell in the population health characterization space includes: Calculate the projection vector of each battery cell in the singular deviation subspace, and calculate the singular energy index based on the projection vector. The singular energy index is used to quantitatively characterize the deviation.

3. The health management method for energy storage devices in an energy-saving elevator system as described in claim 2, characterized in that, The step of identifying abnormal battery cells based on the deviation includes: An adaptive anomaly threshold is determined based on the set of singular energy indicators of all individual battery cells. Calculate the projected coordinates of each battery cell in the main health subspace and its distance to the group health center, and use the distance as the main modal distance; If the singular energy index of a certain battery cell exceeds the adaptive anomaly threshold, or if the main mode distance of a certain battery cell exceeds the corresponding distance threshold, the battery cell is marked as an anomaly candidate. A comprehensive anomaly score is constructed, which is the sum of the ratios of a weighted combination of singular energy index and principal mode distance to the corresponding threshold. When the comprehensive abnormal score of an abnormal candidate battery cell is greater than 1, it is determined to be an abnormal battery cell.

4. The health management method for an energy storage device in an energy-saving elevator system as described in claim 3, characterized in that, The method of identifying abnormal battery cells based on the deviation also includes: Anomalies are classified into multiple levels based on anomaly scores or the degree of excess of exotic energy. If the same battery cell is determined to reach the preset abnormal level in multiple consecutive testing cycles, it is confirmed as a stable abnormality.

5. The health management method for an energy storage device in an energy-saving elevator system as described in claim 3, characterized in that, The method of identifying abnormal battery cells based on the deviation also includes: Calculate the contribution of each feature component of the abnormal battery cell in the singular deviated subspace; Based on the category of the feature component with the largest contribution, the anomaly is mapped to an internal resistance anomaly, polarization anomaly, relaxation anomaly, or consistency degradation anomaly.

6. The health management method for an energy storage device in an energy-saving elevator system as described in claim 5, characterized in that, The construction of a population health representation space as a reference benchmark for population health status through multivariate statistical analysis methods also includes: After each detection cycle, the feature matrix and covariance matrix are updated based on the newly acquired data; A sliding time window weighted update method is adopted. The covariance matrix saved in the previous detection period and the covariance matrix calculated in the current detection period are weighted and fused by the forgetting factor to obtain the updated covariance matrix, so as to suppress the impact of short-term abnormal interference on the population health representation space.

7. The health management method for an energy storage device in an energy-saving elevator system as described in claim 1, characterized in that, The preset operating conditions include: the state of charge of the energy storage device reaches or exceeds a preset threshold, and the battery temperature is within a preset allowable range; The health management method also includes: When the state of charge is lower than the preset threshold, the energy storage device is charged until the preset threshold is met; When the temperature exceeds the allowable range, the step of applying controlled pulse excitation to the energy storage device for a preset duration is suspended.

8. The health management method for an energy storage device in an energy-saving elevator system as described in claim 1, characterized in that, The application of a controlled pulse excitation of a preset duration to the energy storage device includes: The controlled pulse excitation is applied multiple times within a preset time period according to a preset rest interval. The controlled pulse excitation is a constant power charging pulse.

9. The health management method for an energy storage device in an energy-saving elevator system as described in claim 1, characterized in that, The response time series data includes voltage data for the pulse charging phase and the relaxation phase.

10. The health management method for an energy storage device in an energy-saving elevator system as described in claim 1, characterized in that, The health management method also includes: The integrity of the response time series data is verified, including verification of the number of sampling points, verification of the continuity of timestamps, and verification of the consistency of pulse numbers. Abnormal sampling points are identified and corrected. Abnormal sampling point identification includes amplitude mutation detection and sliding window statistical detection, while abnormal sampling point correction includes linear interpolation or sliding window mean value replacement. The response time series data is time-aligned, with the pulse start time as the time zero point; The response time series data is segmented into charging stage subsequences and relaxation stage subsequences. The response time series data is baseline corrected by shifting the data with the voltage value at the pulse start time as the baseline.

11. The health management method for an energy storage device in an energy-saving elevator system as described in claim 10, characterized in that, The health management method also includes: The sampling points that are identified as abnormal and corrected are recorded. The recorded information includes the battery cell number, pulse number, timestamp, original voltage value, abnormality judgment type and correction method. When the number of abnormal sampling points exceeds a preset threshold or when abnormal sampling points are continuously distributed, an abnormal notification event is generated and sent to the analyst's terminal.

12. The health management method for an energy storage device in an energy-saving elevator system as described in claim 11, characterized in that, The health management method also includes: When multiple pulse excitations occur within the same detection cycle, a consistency check is performed on the pre-processed time series of the same battery cell under different pulses, and a morphological similarity index is calculated. When the morphological similarity index of a pulse is lower than the preset similarity threshold, the pulse data is marked as an abnormal pulse and removed.

13. The health management method for an energy storage device in an energy-saving elevator system as described in claim 3, characterized in that, The health management method for energy storage devices in energy-saving elevator systems also includes: When the abnormal battery cell meets the preset triggering conditions, a prompt message is presented to the analyst. The prompt message includes the original voltage time series, the preprocessed time series, key health characteristic curves, the group distribution location, and the abnormal score and level change trend. Receive the manual judgment result corresponding to the prompt information, and modify the threshold for anomaly judgment or the weight coefficient in the comprehensive anomaly score according to the manual judgment result, and / or modify the population health representation space.

14. The health management method for an energy storage device in an energy-saving elevator system as described in claim 13, characterized in that, The process of receiving manual judgment results and adjusting the threshold for anomaly determination or the weighting coefficients in the comprehensive anomaly score based on these results, and / or adjusting the population health representation space, includes: Based on a set of manually confirmed abnormal samples, the adaptive anomaly threshold and distance threshold are adjusted. The weighting coefficients in the comprehensive anomaly score are updated based on the results of manual judgment. For abnormal battery cells identified manually, their weight is reduced or they are removed in the subsequent construction of the population health characterization space. For samples that are manually identified as data anomalies, they will not participate in updating the anomaly detection threshold or constructing the population health representation space. When the proportion of abnormal data exceeds the threshold, the updating of the threshold for anomaly detection is paused.

15. The health management method for an energy storage device in an energy-saving elevator system as described in claim 3, characterized in that, The relative health state parameter ranges from [0,1]. The calculation of the relative health state parameter for each battery cell based on the deviation includes: The normalized principal modality distance is mapped to the principal health component; Calculate the exotic health component based on exotic energy index and abnormal threshold; A stability correction factor is introduced, which is calculated based on the number of consecutive abnormal periods; The relative health state parameters of a single battery cell are obtained by weighting and combining the main health component, the singular health component, and the stability correction factor.

16. The health management method for an energy storage device in an energy-saving elevator system as described in claim 15, characterized in that, The prediction of the health trend of individual battery cells includes: Construct time series of relative health status parameters across multiple testing cycles; The degradation rate is estimated using linear regression or nonlinear fitting methods based on the time series of the relative health state parameters. The health trend of individual battery cells is predicted based on the degradation rate and the preset failure threshold.

17. The health management method for an energy storage device in an energy-saving elevator system as described in claim 1, characterized in that, After acquiring the response time series data of each battery cell at a fixed sampling frequency, the following is also included: The collected response time series data is packaged according to a preset format; Check the communication status with the cloud service platform or remote analysis server; When communication is normal, the encapsulated response time series data is sent to the cloud service platform or remote analysis server through the communication interface; In the event of a communication failure, the data will be temporarily stored in the local storage unit and then resent.

18. A health management system for an energy storage device in an energy-saving elevator system, characterized in that, include: A controlled excitation unit is used to apply a controlled pulse excitation of a preset duration to the energy storage device; The data acquisition unit is used to acquire the response time series data of each battery cell at a fixed sampling frequency during the controlled pulse excitation process and the relaxation phase after the excitation ends. A data preprocessing unit is used to preprocess the response time series data to remove outlier data and standardize and align the data. The feature extraction and modeling unit is used to extract multidimensional health features from the preprocessed response time series data to characterize the battery health status, and to construct a group health representation space as a reference benchmark for the group health status based on the multidimensional health features of all battery cells in the same battery pack. An anomaly identification unit is used to calculate the deviation of the health characteristics of each battery cell in the group health characterization space, and to identify abnormal battery cells based on the deviation. The health assessment unit is used to calculate the relative health status parameters of each battery cell based on the deviation, and to predict the health trend of the battery cells by combining the health status parameters of historical detection cycles. Each of the units is used to implement the health management method for energy storage devices in an energy-saving elevator system as described in any one of claims 1 to 17.

Images