A data center machine room lithium battery pack health state monitoring method and system

Abstract

This application relates to the field of energy storage power management technology, and discloses a method and system for monitoring the health status of lithium battery packs in data center computer rooms. The method includes: collecting data on the chemical reaction rate and electrolyte concentration inside the battery pack, extracting peak changes and gradient distribution differences to form an original index sequence; analyzing the sequence fluctuation trend and quantifying the degree of deviation from the health benchmark; if the deviation exceeds a threshold, extracting the coupling relationship between reaction rate and concentration, and judging potential degradation signals through pattern matching; performing degradation trend analysis and future trajectory simulation, calculating risk weights and generating an intervention command sequence, combining real-time data for feedback verification to make decision alarms, and adjusting the power scheduling strategy accordingly to determine the health maintenance configuration; and continuously monitoring and optimizing the battery pack's operating status based on this configuration. This invention achieves early degradation prediction and adaptive intelligent maintenance of lithium battery packs based on internal mechanisms, improving monitoring accuracy and operational safety.

Description

Technical Field

[0001] This application relates to the field of energy storage power management technology, and in particular to a method and system for monitoring the health status of lithium battery packs in data center computer rooms. Background Technology

[0002] To ensure the continuity of critical business operations, data center server rooms are typically equipped with uninterruptible power supplies (UPS) and energy storage battery banks as backup power systems. With the increasing advantages of lithium-ion batteries in terms of energy density, cycle life, and space utilization, the application of lithium battery banks in data center UPS / energy storage scenarios is constantly increasing. The health status of lithium battery banks directly affects backup power capacity and the operational safety of the server room; therefore, reliable and timely health monitoring is a crucial foundation for data center operation and maintenance management.

[0003] Current data center battery management technology has generally evolved from "external parameter monitoring" to "model-based evaluation" and then to "data-driven prediction": 1) In the early stages, basic safety monitoring mainly relies on the acquisition of external electrical parameters such as voltage, current, and temperature, combined with threshold alarms. 2) Subsequently, a battery management system (BMS) and online estimation methods are introduced, such as estimating SOC / SOH through coulomb measurement and equivalent circuit model, and health assessment is carried out in combination with indicators such as internal resistance and capacity decay; 3) In recent years, with the development of sensing, edge computing and data analysis technologies, multi-parameter fusion and trend prediction methods have been gradually adopted to identify and warn of degradation evolution, so as to support more refined operation and maintenance strategies.

[0004] Although the above-mentioned technologies have improved the monitoring capabilities of lithium battery packs to some extent, the following shortcomings still exist in data center scenarios: (1) Insufficient early detection capability for internal state deterioration. Existing monitoring relies mainly on external electrical parameters or macroscopic indicators (such as voltage, temperature, internal resistance, etc.), which have limited sensitivity to changes in the internal electrochemical processes of the battery and make it difficult to identify early signs of potential anomalies in a timely manner.

[0005] (2) Lack of characterization of the internal mechanism coupling of batteries. The degradation of lithium batteries is often related to the internal chemical reaction process and changes in electrolyte state. Especially under high-rate charge and discharge, temperature control fluctuations or long-term operation, changes in internal reaction rate and uneven electrolyte concentration distribution may interact and accelerate degradation. Existing methods usually cannot monitor and judge from the coupling perspective of "reaction process - electrolyte state" at the same time.

[0006] (3) Trend analysis and risk assessment are insufficient to support refined operation and maintenance. Some methods can achieve anomaly alarms, but they lack quantitative basis for the evolution trend of anomalies, future risk levels and priority handling order, making it difficult to support maintenance priority decision-making in a data center environment with multiple battery packs and multiple circuits running in parallel.

[0007] (4) Lack of closed-loop feedback in monitoring and policy linkage. Data center scenarios emphasize reliability and maintainability. Simple monitoring and alarms are insufficient to meet the requirements. If the monitoring results cannot be further linked to scheduling policies and maintenance configurations, and continuous calibration can be achieved through feedback updates, it will be difficult to achieve long-term stable health management.

[0008] Therefore, in the specific high-reliability scenario of data center computer rooms, it is urgent to break through the limitations of existing external monitoring and static management, and provide a lithium battery pack health status monitoring method that can perceive the core chemical state inside the battery in real time, analyze the dynamic coupling of multiple parameters, realize accurate prediction of early degradation, and support intelligent adaptive maintenance, so as to fundamentally improve the operational safety and service life of the battery system. Summary of the Invention

[0009] To address the shortcomings of existing data center lithium battery pack monitoring methods, such as reliance on external electrical parameters, insufficient perception of early degradation of the battery's internal electrochemical state, lack of trend prediction and risk prioritization decision-making basis, and difficulty in forming a closed-loop linkage between monitoring results and alarms and operation and maintenance strategies, this application provides a data center lithium battery pack health status monitoring method and system. By collecting internal chemical reaction data and electrolyte data of the lithium battery pack, it completes anomaly identification, trend prediction, risk assessment, and feedback updates, thereby achieving continuous monitoring and operation management of the lithium battery pack's health status.

[0010] Firstly, this application provides a method for monitoring the health status of lithium battery packs in a data center, the method comprising: S1. Collect chemical reaction data and electrolyte data inside the lithium battery pack, extract features from the chemical reaction data and electrolyte data, and obtain the original index sequence; S2. Analyze the fluctuation trend of the original indicator sequence and determine the degree of deviation of each indicator from the benchmark. S3. If the deviation exceeds the preset deviation threshold, extract the abnormal features and perform pattern matching to determine whether there is a potential degradation signal. S4. For potential degradation signals, analyze degradation trends and simulate future evolution trajectories, calculate and determine the risk weight allocation priority for each signal; S5. Generate an intervention action adjustment plan based on the priority allocation of risk weights, sort and optimize each intervention instruction, and generate an optimized instruction sequence; S6. Evaluate and optimize the instruction sequence. If the optimized instruction sequence meets the preset alarm conditions, determine whether to activate the alarm state through a feedback loop mechanism. S7. Adjust the power dispatch strategy according to the alarm status and determine the health maintenance configuration. The health maintenance configuration shall include at least power output regulation, charge and discharge cycle optimization and battery temperature control. S8. Continuously monitor the battery pack status and update data feedback according to the health maintenance configuration, thereby optimizing the battery pack operation status.

[0011] Secondly, this application provides a health status monitoring system for lithium battery packs in data center computer rooms, the system comprising: The data acquisition module is used to collect chemical reaction data and electrolyte data inside the lithium battery pack, extract features from the chemical reaction data and electrolyte data, and obtain the original index sequence. The trend analysis module is used to analyze the fluctuation trend of the original indicator sequence and determine the degree of deviation of each indicator from the benchmark. The anomaly detection module is used to extract abnormal features and perform pattern matching when the deviation exceeds a preset deviation threshold to determine whether there are potential degradation signals. The risk prediction module is used to analyze the degradation trend and simulate the future evolution trajectory for potential degradation signals, and calculate and determine the risk weight allocation priority for each signal. The intervention optimization module is used to generate intervention action adjustment plans based on risk weight priority allocation, and to sort and optimize various intervention instructions to generate an optimized instruction sequence. The alarm evaluation module is used to evaluate and optimize the command sequence. If the optimized command sequence meets the preset alarm conditions, the alarm state is activated through a feedback loop mechanism. The scheduling and adjustment module is used to adjust the power scheduling strategy according to the alarm status and determine the health maintenance configuration. The health maintenance configuration includes at least power output regulation, charge and discharge cycle optimization and battery temperature control. The feedback optimization module is used to continuously monitor the battery pack status and update the data feedback based on the health maintenance configuration, thereby optimizing the battery pack's operating status.

[0012] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows: 1. By directly collecting and fusing data on chemical reaction rates and electrolyte concentrations within lithium-ion battery packs, a sequence of microscopic indicators reflecting the core electrochemical mechanisms is constructed, achieving a leap from "external phenomenon monitoring" to "internal mechanism insight." This method can keenly identify early degradation signals characterized by abnormal internal coupling relationships, greatly improving monitoring sensitivity and solving the problem of severe early warning lag and missed detection caused by relying on external parameters such as voltage and temperature. This represents a fundamental breakthrough in monitoring depth and early warning foresight.

[0013] 2. After identifying potential degradation, qualitative anomalies are transformed into quantitative risk weights and priority rankings through inflection point analysis and future trajectory simulation. This provides a scientific data-driven decision-making basis for the precise scheduling and handling order of operation and maintenance resources in the parallel scenario of multiple battery groups in data centers, realizing the upgrade from "uniform processing" to "differentiated precise intervention", and moving risk decision-making from experience-based and static to quantitative and dynamic.

[0014] 3. By setting up a feedback loop mechanism based on real-time status verification, the intervention plan can be dynamically evaluated and calibrated, and key parameters can be fed back to the front-end analysis model. This forms a complete intelligent closed loop of "monitoring-diagnosis-decision-execution-optimization", enabling the entire health management system to continuously improve itself and adapt to changes in operating conditions, significantly improving the reliability and stability of long-term operation.

[0015] 4. The health maintenance configuration can be directly linked to power scheduling to implement proactive voltage regulation, cycle adjustment, and temperature control strategies. This not only effectively suppresses abnormal expansion and prevents safety accidents such as thermal runaway, ensuring the continuity of power supply for core businesses, but also maximizes the lifespan of battery packs through precise health status management, reducing the high operation and maintenance costs caused by excessive maintenance or sudden replacement. This provides key technical support for data centers to achieve the operational goals of high reliability and low total cost of ownership (TCO), improving the economy and security of data center power supply systems. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for monitoring the health status of lithium battery packs in a data center, as described in this application. Figure 2 This is a schematic diagram of the structure of a data center computer room lithium battery pack health status monitoring system according to this application. Detailed Implementation

[0018] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of a method for monitoring the health status of lithium battery packs in a data center, provided by the present invention. The flowchart specifically includes the following steps: S1. Collect chemical reaction data and electrolyte data inside the lithium battery pack, extract features from the chemical reaction data and electrolyte data, and obtain the original index sequence.

[0020] In one specific embodiment, the process of performing step S1 may specifically include the following steps: A sensor array is deployed in the key electrochemical region inside the lithium battery pack to acquire real-time data on chemical reaction rates and electrolyte concentrations within the lithium battery pack. Extract the peak variation characteristics of chemical reaction rate data over time, and extract the gradient distribution differences of electrolyte concentration data in the internal space of the battery. By correlating and integrating the peak variation characteristics and gradient distribution difference characteristics, the original micro-indicator sequence is generated; The original micro-indicator sequence is cleaned and formatted to obtain the original indicator sequence.

[0021] Specifically, in data center scenarios, lithium battery packs are typically installed in battery cabinets or racks as backup energy storage units for UPS systems. They operate under conditions such as float charging, periodic capacity checks, or load switching for extended periods. Traditional monitoring relies heavily on external quantities such as terminal voltage, current, and temperature, making it difficult to detect subtle fluctuations in the internal electrochemical processes of the battery in their early stages. Furthermore, the internal reaction changes are coupled with the electrolyte state distribution and exhibit spatial non-uniformity. Single external parameters or single-point data can easily average out local anomalies, leading to a delay in identifying potential degradation signals. Step S1 revolves around "converting internal process data into time-series indicators that can be used for subsequent fluctuation analysis." The core of this step is to establish a temporal and spatial correspondence between chemical reaction rate data and electrolyte concentration data, and to map the two types of raw data into a sequence of raw indicators on the same time axis after feature processing. To address this, when arranging the sensor array within the electrochemical region of the battery pack, the sensor positions are corresponding to the structural partitions of the cells or modules. This ensures that the sampling points cover both the reactive areas of the electrodes and the electrolyte channels or locations prone to concentration stratification. The array output forms two types of synchronous data streams: one is a chemical reaction rate sequence, forming a set of reaction rate data according to the sampling time; the other is an electrolyte concentration matrix, forming a set of multi-point concentration data by superimposing spatial positions according to the sampling time. The chemical reaction rate data is a kinetic index sequence estimated from voltage, current, temperature, and impedance spectrum data using an equivalent circuit model and a state observer. The electrolyte concentration data is a concentration gradient index sequence mapped from multi-point temperature, pressure difference, and impedance frequency band characteristics using a mass transfer model. The sensor array is arranged on the module surface and busbar nodes within the battery cabinet, without puncturing the cells.

[0022] Feature extraction of chemical reaction rate data targets peak changes because degradation-related side reactions, increased local polarization, or interfacial film growth can manifest as sharp peaks, peak width variations, or abnormal peak spacing in the reaction rate curve under charge / discharge or float charge disturbances. Peak features are more effective than mean values ​​in reflecting sudden and phased changes. In implementation, the reaction rate sequence is segmented into time windows. Within each window, peaks are identified through local extrema detection and threshold constraints. Parameters such as peak amplitude, peak width, peak location, and the difference between adjacent peaks are then calculated and mapped to the corresponding peak change feature set for that window. To avoid spurious peaks caused by electromagnetic interference from the data center, sampling jitter, or transient noise, smoothing and robustness constraints are introduced before peak identification. For example, median filtering or moving averages are applied to the reaction rate sequence to reduce isolated spikes, and minimum duration and minimum amplitude change thresholds are used to exclude short-term fluctuations. This ensures a stable data processing chain: "original reaction rate sequence—window—peak set—peak change feature set." At this point, each time window outputs a set of peak change characteristics, which correspond to the original reaction rate data in the time dimension. That is, the reaction rate samples within the window determine the peak change characteristics of the window.

[0023] Feature extraction of electrolyte concentration data targets gradient distribution differences because spatial non-uniformity of concentration can lead to limited local ion transport, accelerated local reactions, and hotspot formation, thereby causing capacity decay or safety risks. This spatial non-uniformity often precedes terminal voltage anomalies. In practice, at each sampling time, the concentration values ​​at each sensor location are organized geometrically into a spatial vector or spatial grid. Indicators reflecting distribution differences are then calculated; for example, the difference between the maximum and minimum values ​​characterizes the distribution span, the statistical measure of the concentration difference between locations characterizes the degree of dispersion, and the difference between adjacent locations characterizes the gradient strength. These calculation results are then aggregated into a gradient distribution difference feature set for that moment. Since concentration sensors may experience drift and response hysteresis, consistency correction is performed on the concentration sequence during data processing. For example, zero-drift compensation or baseline calibration based on stable sections under the same operating conditions ensures that concentration data from different locations are comparable on the same scale. Missing value interpolation maintains the integrity of the spatial vector, ensuring that gradient calculations are not interrupted by single-point missing measurements.

[0024] The integration of peak change features and gradient distribution difference features requires addressing the inconsistency in temporal granularity, as peak features are often output in units of windows, while gradient features are output at sampling times. Integration involves projecting both types of features onto the same time index. This can be achieved by using the center time of the window as the time label for the window features, while simultaneously aggregating gradient features within the window's coverage area to obtain gradient statistical features with the same granularity as the window; or by downsampling the peak features to the window label and then concatenating them with the gradient aggregation results of the same label, thus constructing a comprehensive micro-feature vector corresponding to each window. One component of this vector originates from the peak change in reaction rate, and another component from the difference in concentration gradient distribution. These two components are parallel under the same time label, ensuring that subsequent fluctuation trend analysis can simultaneously observe the synchronous changes of both "temporal abrupt changes" and "spatial inhomogeneity."

[0025] Data cleaning and formatting revolve around usability and consistency. In a data center environment, issues such as sampling packet loss, transient interference, and communication delays exist. During the cleaning phase, outlier removal and missing data repair are performed on reaction rate and concentration data respectively. Time alignment verification is then performed before fusion to ensure that feature vectors under the same time label originate from the same operational phase. In the formatting phase, feature vectors are stored in a structured manner by field, including time labels, peak change feature fields, gradient distribution difference feature fields, and quality marker fields. The quality marker records whether the data has undergone interpolation or strong correction, which is used for weighted processing in subsequent deviation calculations. After cleaning and formatting, the original index sequence is obtained. This sequence expresses the abrupt changes in internal reaction rates and the spatial non-uniformity of electrolyte concentration in a unified form, enabling subsequent steps to analyze fluctuation trends and deviations based on the same data. This addresses the problems of "external electrical parameters being insufficient to capture subtle internal degradation," "lack of characterization of the coupling changes between the reaction process and electrolyte state," and "noise and inconsistencies leading to misjudgments."

[0026] S2. Analyze the fluctuation trend of the original indicator sequence and determine the degree of deviation of each indicator from the benchmark.

[0027] In one specific embodiment, the process of performing step S2 may specifically include the following steps: Within a set time window, fluctuation frequency analysis is performed on the original index sequence to identify the locations of abrupt changes in chemical reaction rates and the periodic patterns of electrolyte concentration changes in the original index sequence. Based on the location of the mutation point and the periodic pattern of change, the abnormal fluctuation range in the original indicator sequence is determined; Calculate the deviation of the abnormal fluctuation range from the preset baseline to quantify the initial degree of trend deviation; The initial trend deviation is compared and verified with historical health data trends, and the initial trend deviation is calibrated and updated based on the verification results to obtain the final deviation.

[0028] Specifically, the goal of this step is to transform the original indicator sequence from the form of "time tag + multi-dimensional feature vector" into "fluctuation trend quantification result", so that each time period corresponds to a deviation degree from the benchmark, and ensure that the deviation degree can be aligned with the historical health status trend for verification, thereby reducing misjudgments caused by noise, communication jitter, and operating condition switching, and solving the problems of "lack of analysis of internal micro fluctuations, lack of quantification of coupling changes, and lack of reliability verification of trend judgment".

[0029] The original index sequence consists of feature vectors corresponding to multiple time labels. One component corresponds to the peak variation characteristics of chemical reaction rates, while another component corresponds to the differences in electrolyte concentration gradient distribution. Therefore, fluctuation trend analysis requires processing both types of features simultaneously on the same time axis while maintaining their correspondence. Time window settings are used to segment the continuous sequence, allowing for comparison of local fluctuation structures. The window length matches the sampling period and aligns with the scale of common events in the data center. It can be set according to the typical operating cycle of the data center, for example, adapting to the adjustment cycle of the float charge control loop, the equalization charge cycle, or the duration of load switching events, thus ensuring that each window covers several feature vectors. Fluctuation frequency analysis performs time-series statistics on the feature vector components within each window. For chemical reaction rate-related components, abrupt changes are identified through the rate of change and spectral features; for electrolyte concentration-related components, periodic patterns are identified through autocorrelation or spectral peaks. In specific processing, the reaction rate-related component forms a time-varying sequence within the window. The slope sequence is obtained through differencing, and the extreme values ​​of the rate of change are calculated. When the rate of change exceeds a threshold multiple of the steady-state noise level within the window and the duration exceeds a minimum duration threshold, the location of the mutation point is recorded. Each mutation point location corresponds one-to-one with a time label, forming a mutation point set. The concentration gradient-related component forms another sequence within the window. The interval between repeating peaks is found through the autocorrelation function, or the period length corresponding to the frequency domain energy peak is used to obtain a set of periodic parameters. These periodic parameters correspond to the window labels and can be updated as the window scrolls. The mutation point set and the periodic parameter set maintain a correspondence with the window, allowing the same window to simultaneously describe both the "location of the mutation" and the "scale of periodic fluctuations." This establishes a connection between the two types of data in the algorithm processing: mutation points are used to characterize event-type anomalies, while periodic parameters are used to characterize slow evolution or periodic fluctuations introduced by control loops.

[0030] The determination of abnormal fluctuation intervals is based on the statistical analysis of fluctuation amplitudes within a window. Reaction rate-related components typically exhibit peak amplitude increases or disordered peak spacing near abrupt change points, while concentration gradient-related components may show periodic drift or increased amplitude during periodic fluctuations. Therefore, a joint rule can be used to determine abnormal intervals: calculate the fluctuation amplitude measures of both the reaction rate-related and concentration gradient-related components within the window, and then map these two amplitude measures to a joint amplitude sequence on the same time label. The amplitude measures can be composed of the difference between the maximum and minimum values ​​within the window, the root mean square fluctuation, or a robust amplitude measure based on percentile differences, resulting in a reaction rate amplitude sequence and a concentration gradient amplitude sequence. These two sequences are aligned using the same time label and weighted to synthesize a joint amplitude sequence. An abnormal interval is defined as a continuous time period during which the joint amplitude sequence exceeds the baseline level within the window, and this continuous time period includes abrupt change points or segments inconsistent with the main periodic peaks displayed by the periodic parameters. This incorporates both abrupt change structures and periodic structures into the abnormal interval determination logic, avoiding misjudging normal periodic fluctuations as abnormal based solely on amplitude thresholds, and also avoiding misjudging short-term noise as fault symptoms based solely on abrupt change points.

[0031] Deviation calculation quantifies the deviation of an abnormal interval from a preset baseline as a trend deviation. The baseline is derived from a health status reference, which can be obtained from statistics of similar operating conditions in historical healthy periods or provided by a calibration model. The baseline corresponds to a reference value that changes over time or with an operating condition index for each feature component, ensuring comparability across different operating conditions. Deviation calculation differs the feature vector of each time tag within the abnormal interval from the baseline to obtain a component deviation vector. This component deviation vector is then converted into a dimensionless deviation quantity using a normalized scale. The normalization scale can be the fluctuation range or standard deviation of that component in historical healthy periods, thus placing the reaction rate feature and concentration gradient feature within the same dimensional framework. To express the degree of deviation, an aggregation operator is used to aggregate the dimensionless deviation quantity into a single-value index. For example, the weighted absolute sum of the dimensionless deviation quantity is taken, and the average or peak value is taken within the abnormal interval to obtain a preliminary trend deviation degree. This processing chain maps multidimensional feature deviations to single-value deviation degrees, solving the problem that it is difficult to form a unified criterion for various internal data, and making the deviation degree reflect both the abnormality of the reaction process and the abnormality of the electrolyte spatial distribution.

[0032] Due to control strategy switching and load fluctuations in data center server rooms, the initial deviation of a single window may be affected by short-term operational disturbances. Therefore, historical health data trend comparison and verification are introduced. Historical health data trends consist of deviation sequences of similar battery packs during their health phases, stored hierarchically by operational condition label and time scale, allowing for the selection of matching historical segments during comparison. The comparison and verification aligns the current initial deviation sequence with the historical health trend sequence on the same time scale, calculating similarity or deviation residuals. If the current deviation falls within the confidence interval of the historical health trend, its anomaly weight is reduced; if it consistently exceeds the upper confidence bound or shows a monotonically increasing trend and significantly deviates from the historical health trend, its anomaly weight is increased. This verification result is used for calibration updates. Calibration updates can correct the initial deviation through a confidence factor, attenuating short-term, sudden deviations consistent with historical health trends, while retaining or amplifying persistent deviations inconsistent with historical trends. This results in an updated deviation output that maintains correspondence with the window label. For example, the confidence factor is a weighted coefficient characterizing the consistency between the initial trend deviation and the historical health data trend. Its value range can be set from 0 to 1. Calibration is achieved by multiplying the initial deviation by this coefficient. High consistency results in a coefficient approaching 0 to attenuate deviations caused by operational disturbances, while low consistency results in a coefficient approaching 1 to retain the deviation. The confidence factor is obtained by mapping the similarity between the current deviation sequence and the historical health sequence. High similarity reduces the deviation weight, while low similarity maintains the deviation weight. This calibration process links "current observation—historical health reference—confidence correction," ensuring that the deviation comes from real-time data and is constrained by health references, reducing error propagation caused by noise and operational disturbances.

[0033] S3. If the deviation exceeds the preset deviation threshold, extract the abnormal features and perform pattern matching to determine whether there are potential degradation signals.

[0034] In one specific embodiment, the process of performing step S3 may specifically include the following steps: Determine whether the degree of deviation exceeds the preset deviation threshold. If so, extract the coupling relationship between chemical reaction rate and electrolyte concentration from the abnormal fluctuation range. Based on the coupling relationship characteristics and gradient distribution difference characteristics, a comprehensive state feature vector is constructed. The comprehensive state feature vector is then matched and analyzed with a pre-set degradation mode feature library to obtain the mode matching results. Based on the pattern matching results, determine whether there are potential degradation signals in the current battery state; If potential degradation signals are present, record the characteristic parameters of the abnormal features. The characteristic parameters include at least the chemical reaction rate characteristics and the electrolyte concentration gradient characteristics. By correlating characteristic parameters with historical degradation data, a preliminary degradation signal report is generated.

[0035] Specifically, data center lithium battery packs operate under float charging and periodic discharge conditions. External electrical parameters, influenced by UPS control and bus voltage regulation, exhibit gradual changes. This leads to early-stage internal degradation often manifesting as coupled fluctuations in the chemical reaction process and electrolyte state, rather than significant shifts in port voltage. The degree of deviation corresponds one-to-one with time window labels. Preset deviation thresholds can be configured into a threshold table based on battery model, operating condition label, and ambient temperature level. Therefore, the degree of deviation within the same time window can be directly mapped to the corresponding threshold and trigger judgment. When the deviation exceeds the threshold, it indicates that the combined amplitude anomaly within that window has exceeded the healthy fluctuation boundary. However, this conclusion still needs to distinguish between "recoverable fluctuations caused by operating condition disturbances" and "persistent anomalies caused by degradation mechanisms." Therefore, anomaly feature extraction and pattern matching processes are initiated to identify potential degradation signals.

[0036] Anomaly feature extraction is limited to anomaly ranges within which fluctuation amplitudes are abnormal. Each anomaly range corresponds to a set of time labels and includes a multidimensional feature vector set of the original indicator sequences. The chemical reaction rate component reflects fluctuations in the reaction process, while the electrolyte concentration gradient component reflects spatial unevenness. The extraction of coupling relationship features revolves around the dynamic correlation between reaction rate and concentration. This is because battery degradation is often accompanied by limited mass transfer and enhanced side reactions, manifesting as increased synchronicity, lag, or nonlinearity between changes in peak reaction rate and concentration gradient fluctuations. Observing any single component alone makes it difficult to reliably distinguish between disturbances and degradation. During data processing, the reaction rate feature sequence and concentration gradient feature sequence corresponding to each time label within the anomaly range are aligned to form paired sample sequences. These paired sample sequences maintain a one-to-one correspondence; that is, the reaction rate feature value and concentration gradient feature value under the same time label form the same pair. The coupling relationship characteristics of this paired sample sequence can be calculated using a combination of correlation and time-lag correlation. For example, by using sliding correlation coefficients, cross-correlation peak positions, and nonlinear coupling measures based on regression residuals, the coupling relationship characteristics can include both synchronous correlation strength and response lag information. When the cross-correlation peak deviates from zero and exhibits a stable lag, it indicates that there is a mass transfer constraint path for the change in reaction rate due to the change in concentration distribution. When the correlation coefficient increases significantly in the abnormal interval or the correlation structure changes from weak to strong correlation, it indicates that the coupling between the reaction process and the concentration distribution is strengthened. This strengthening can correspond to local polarization or interfacial film growth. This yields a set of coupling relationship characteristics, which corresponds to the abnormal interval labels and can be aggregated into a coupling relationship feature vector by window.

[0037] The construction of the comprehensive state feature vector uses coupling relationship features and gradient distribution difference features as inputs. The gradient distribution difference features can be extracted using the same time label within the abnormal interval. Since coupling relationship features reflect intervariate relationships and gradient distribution difference features reflect univariate spatial states, the two are physically complementary. The comprehensive vector is constructed by splicing and normalizing the components using the same time window label, bringing all components to the same scale. During the processing, the coupling relationship feature components and gradient distribution difference components are normalized using the baseline scale. The baseline scale can be the statistical range or standard deviation of the healthy segment. After normalization, the vectors are spliced ​​to obtain the comprehensive state feature vector. The correspondence between the vector elements and the original data is clear: the coupling components are calculated jointly by the reaction rate feature sequence and the concentration gradient feature sequence, and the gradient components are calculated from multi-point concentration data. This comprehensive vector retains both the intensity information of spatial non-uniformity of concentration and the coupling structure information of the reaction process and concentration state.

[0038] Pattern matching is performed using a pre-defined degradation pattern feature library as a reference. This library stores pattern identifiers, feature vector templates, similarity thresholds, and applicable operating condition labels for various degradation patterns, such as overcharge-related patterns, mass transfer-limited patterns, and thermal runaway precursor patterns. Each entry contains a vector template and an allowable fluctuation range, and is indexed with operating condition labels to support matching within the same operating condition. The templates are obtained from abnormal samples corresponding to historical maintenance records through feature extraction and clustering. The thresholds are set by the quantiles of the similarity distribution of samples within the same category, and the template parameters can be updated incrementally with new samples. Matching analysis inputs the comprehensive state feature vector into the matching algorithm. The matching algorithm can be implemented using a combination of distance and similarity metrics to control the contribution of different scale components to the matching. For example, a distance sequence with each template is calculated using weighted distance, and then the distance sequence is mapped to a similarity score sequence, with each similarity score corresponding one-to-one with a library entry. When the similarity score exceeds the threshold for the corresponding entry, the entry is considered a match, and the identifier information of the matched entry is recorded. If multiple entries match simultaneously, the entry with the highest similarity score is selected, or the entry requiring priority response is selected according to the risk level rules. The pattern matching result includes a matching flag, a matching pattern identifier, and a similarity score, corresponding to anomaly interval labels or window labels. When judging potential degradation signals based on pattern matching results, the similarity score is compared with the threshold and combined with the persistence condition of the deviation degree. The persistence condition can be defined by the number of consecutive window matches or the similarity maintenance time to exclude false alarms caused by occasional matches in a single window, thereby achieving the screening of persistent anomalies.

[0039] When a potential degradation signal is established, the characteristic parameters of the anomaly are recorded to support subsequent trend prediction and intervention decisions. These characteristic parameters include at least chemical reaction rate characteristics and electrolyte concentration gradient characteristics. Chemical reaction rate characteristics may include peak amplitude, peak width, and peak spacing variation rate, while electrolyte concentration gradient characteristics may include distribution span, gradient statistics, and differential peak values. Each parameter is labeled with a time tag and operating condition tag to ensure comparability with historical samples. Historical degradation data correlation analysis is based on a mapping of "characteristic parameters—historical samples—pattern identifiers." Historical degradation data stores characteristic parameter trajectories and known treatment results according to pattern identifiers. Correlation analysis aligns the current characteristic parameter sequence with historical trajectories of the same pattern and calculates trajectory similarity or evolution rate deviation, forming the content structure of a preliminary degradation signal report. The report includes matching pattern identifiers, similarity scores, deviation ranges, and correlation results between characteristic parameter fragments and historical trajectories. These correlation results are used as one of the inputs for subsequent risk weight calculations, thereby transforming internally coupled anomalies into traceable signal records in data center operation and maintenance scenarios.

[0040] S4. For potential degradation signals, analyze degradation trends and simulate future evolution trajectories, calculate and determine the risk weight allocation priority for each signal.

[0041] In one specific embodiment, the process of performing step S4 may specifically include the following steps: The characteristic parameters of potential degradation signals are obtained, degradation curves are constructed based on the characteristic parameters, and degradation curve inflection point identification technology is used to analyze the slope change characteristics of the degradation curve. The degradation threshold of the degradation trend is determined based on the slope change characteristics. The degradation threshold is used to distinguish the degree of degradation at different stages. Based on the degradation threshold and the preset prediction time window length, the future evolution trajectory of the degradation trend is simulated to obtain prediction trajectory data; Based on the predicted trajectory data, the degree of risk that potential degradation signals may trigger within the predicted time window is quantitatively calculated, and the corresponding degradation risk weight value is obtained. Based on the degradation risk weight values, multiple potential degradation signals are prioritized to determine the risk weight allocation priority of each signal, and the risk weight allocation result is generated.

[0042] Specifically, in parallel operation and group management modes, data center lithium battery packs have multiple modules, multiple series and parallel branches, and multiple sensing channels. They also alternate between floating charge maintenance, periodic equalization charge, core capacity discharge, and load switching. Once potential degradation signals appear concurrently on different branches, relying solely on alarm thresholds is insufficient to determine the appropriate handling order. Step S4 establishes a data processing link based on the evolution of potential degradation signals, from "characteristic parameters—degradation curve—stage threshold—future trajectory—risk weight—priority ranking." This assigns a comparable risk weight value to each potential degradation signal and establishes a ranking basis among multiple signals, thereby supporting the generation of intervention commands and the linkage of power scheduling.

[0043] In step S3, the potential degradation signal has been structured into a output containing a pattern identifier, similarity score, anomaly interval label, and a set of feature parameters. These feature parameters include at least chemical reaction rate features and electrolyte concentration gradient features, and each record includes a time label and branch identifier. Based on these feature parameters, a degradation curve is constructed. The construction method focuses on "mapping discrete feature parameters into a time-varying degradation index sequence," ensuring that feature parameters of the same signal at different time windows can form the same curve. During data processing, feature parameters under the same branch and pattern identifier are sorted by time label to form a parameter sequence. The reaction rate feature sequence and the concentration gradient feature sequence maintain the same time index and are fused after normalization to obtain the degradation index sequence. The fusion method can use weighted summation or vector norm mapping to ensure the output is a single-valued sequence sensitive to both types of parameters. This single-valued sequence corresponds one-to-one with the time label, thus obtaining the degradation curve. Each sampling point of the degradation curve originates from the reaction rate and concentration gradient features within the same time window, ensuring that the curve reflects the co-evolution of reaction process fluctuations and spatial distribution anomalies. For example, a degradation curve is constructed by using the change in peak reaction rate as the ordinate and the difference in electrolyte concentration gradient distribution as the abscissa.

[0044] Inflection point identification of degradation curves is used to capture transition points in the degradation evolution process. This is because degradation in data center environments often exhibits a phased acceleration, with the curve slope changing from gradual to steep at certain times, related to accelerated interfacial film growth, increased mass transfer limitation, or localized overheating. The inflection point identification process calculates the slope change characteristics on the degradation curve. The slope is obtained by differencing adjacent points and can be smoothed within a sliding window to reduce noise. The slope change characteristics are represented by the rate of change of the slope sequence, second-order difference, or piecewise fitting residuals, allowing the transition from one linear approximation segment to another to be identified as an inflection point. The inflection point location corresponds to a time stamp, and the slope difference before and after the inflection point is used to quantify the intensity of the stage transition, forming a slope change characteristic set. This set includes both the time of the inflection point and the slope levels before and after the inflection point.

[0045] The degradation threshold is used to distinguish the degree of degradation at different stages and to provide boundary conditions for future trajectory simulation. The degradation threshold is determined based on slope change characteristics and historical degradation sample statistics. During processing, the slope levels before and after the inflection point are mapped to stage labels. For example, a slope below a certain level corresponds to a gradual change stage, a slope between two levels corresponds to a development stage, and a slope above the upper level corresponds to an acceleration stage. Simultaneously, a stage switching threshold is set in conjunction with the inflection point intensity. The threshold can be represented as segmented boundary values ​​of the degradation index sequence or slope level boundary values, and is associated with and stored with mode identifiers, allowing different degradation modes to have different threshold sets. This threshold enables the same signal to be determined to be in which degradation stage at the current moment, and allows different signals to be compared by stage, thus transforming "slope change characteristics" into "stage-based criteria."

[0046] The future evolution trajectory simulation extrapolates the degradation trend and outputs predicted trajectory data within a predicted time window. The predicted time window length is related to the data center maintenance cycle, the frequency of backup power supply strategy adjustments, and alarm response time limits. It can be configured at the minute, hour, or day level and matched with the sampling period to generate the prediction step size. The simulation process uses the current state at the end of the degradation curve as the initial condition and selects evolution model parameters based on the stage characteristics reflected by the degradation threshold. For example, a smaller growth rate is used in the gradual change stage, and a larger growth rate or non-linear growth form is used in the acceleration stage, thus obtaining the index prediction sequence within the prediction window. The predicted trajectory data includes the degradation index value and stage prediction label corresponding to each prediction step, and can be accompanied by a confidence interval to characterize uncertainty. This uncertainty can be estimated from the fluctuation range of historical samples of the same pattern or obtained from model residual propagation. In this way, a single potential degradation signal is extended from the current state to a future evolution trajectory, solving the problem of "the inability to predict future risk trends leading to passive handling."

[0047] Risk quantification is based on predicted trajectory data, converting the "potential future degradation level" into comparable risk weight values. During quantification, the predicted trajectory is associated with risk criteria, which can be derived from a set of degradation thresholds. For example, the time it takes for the predicted trajectory to enter the acceleration phase within a window, the magnitude of exceeding the phase boundary, and the index value at the end of the window can all be considered risk factors. The risk weight value can be denoted as R_k, corresponding to the k-th potential degradation signal. The calculation can employ a weighted combination method. For instance, the proportion of time exceeding the threshold within the predicted window, the integral of the threshold exceedance magnitude, and the reciprocal of the margin between the predicted end and the safety boundary are used as inputs, normalized, and then weighted summed to obtain a single-value weight. Preferably, the risk weight is obtained by combining the proportion of threshold exceedance duration, the integral of the threshold exceedance magnitude, and the reciprocal of the margin between the predicted end and the safety boundary, and normalized to 0-1. When the same battery cluster bears a high proportion of the load, the risk weight is multiplied by a load impact coefficient to reflect the degree of power supply risk exposure. Since data centers have multiple branches running in parallel, risk weights can also be superimposed with mapping factors that affect business operations, such as the load ratio or redundancy level of the branch. This allows the weights to reflect both the severity of degradation and the degree of power supply risk exposure, supporting the selection of operation and maintenance resource allocation and scheduling strategies.

[0048] Prioritization of multiple potential degradation signals is performed on a set of risk weight values, which are independently calculated for each signal and are comparable. The ranking results are arranged in descending order of weight to form a priority list, while retaining pattern identifiers, branch identifiers, and stage labels as explanatory fields. This allows the subsequent intervention generation stage to extract the corresponding treatment strategies and triggering times in the order of the list. If signals with similar weights exist, secondary ranking can be performed based on the order in which the predicted trajectory enters the acceleration stage, or fine-tuning can be done by combining similarity scores as a matching reliability factor, thereby integrating "pattern recognition reliability" and "trend risk" into the same ranking framework. This risk weight allocation result maps each signal to a ranking position and weight value, solving the problem of "lack of treatment order for multiple concurrent signals".

[0049] S5. Generate an intervention action adjustment plan based on the priority allocation of risk weights, sort and optimize each intervention instruction, and generate an optimized instruction sequence.

[0050] In one specific embodiment, the process of performing step S5 may specifically include the following steps: Obtain the results of risk weight allocation priority, and extract the interval data corresponding to the high-risk level from the results; Based on the data from high-risk zones, the timing for triggering intervention instructions is determined. Based on the triggering timing and the degradation characteristics reflected by the high-risk interval data, the intervention instructions are prioritized to generate an initial intervention instruction sequence. Analyze the matching degree between the initial intervention instruction sequence and the high-risk interval data, and adjust the action magnitude of the corresponding instructions in the initial intervention instruction sequence based on the analysis results; The initial intervention instruction sequence is optimized based on the adjusted intervention action range to generate an optimized instruction sequence.

[0051] Specifically, lithium battery packs in data center computer rooms are typically connected to the UPS DC bus at the cabinet or cluster level. There is a linkage between battery management and power supply and distribution scheduling. Potential degradation signals, after risk weight allocation, still need to be converted into executable intervention instructions. Otherwise, they can only remain at the level of alarm prompts and cannot solve the problems of "lack of closed-loop linkage and lack of quantitative basis for handling sequence and action intensity". Step S5 takes the risk weight allocation priority as input, maps the degradation risk to the intervention action adjustment plan and forms an instruction sequence, so that each instruction has triggering conditions, execution timing and action amplitude parameters, and ensures that multiple instructions meet the sequence relationship and resource constraints in the parallel signal scenario.

[0052] The risk weight allocation priority result includes branch identifiers, pattern identifiers, risk weight values, and ranking positions for multiple potential degradation signals. It may also include predicted trajectory data or its derivatives, such as the time point of entering the acceleration phase and the duration of exceeding the threshold. The interval data corresponding to the high-risk level is extracted from this result. In the time dimension, the interval data corresponds to segments within the prediction window where risk elements are significant; in the numerical dimension, it corresponds to segments where the predicted degradation index sequence exceeds the stage threshold or where the risk weight value exceeds the level threshold. The extraction process scans the predicted trajectory sequence of each high-risk signal by time index, marks continuous time periods that meet the threshold exceeding condition, and binds these time periods with the branch identifier of the signal to form a high-risk interval dataset.

[0053] The timing of intervention commands is determined based on high-risk interval data. The trigger timing can be represented as the time stamp corresponding to the interval start point, a specific node within the interval, or a lead time before the interval. The selection method is related to the data center operation and maintenance strategy; for example, setting an early trigger before entering the acceleration phase can prevent risk accumulation. The processing aligns the start time of the high-risk interval with the stage stamp in the predicted trajectory and, combined with the redundancy configuration of the battery cluster in the UPS system, determines the acceptable handling delay, forming the trigger time parameters for each signal. If multiple high-risk signals exist within the same battery cluster, the trigger timing also needs to consider concurrent conflicts. For example, operations that cause bus power fluctuations are not allowed to be executed simultaneously within the same time period. Therefore, the trigger timing can be adjusted through window shifting or time slot allocation to ensure that the set of trigger timings satisfies scheduling constraints.

[0054] The priority of intervention instructions is determined by the triggering timing and the degradation characteristics reflected in the high-risk range. These degradation characteristics originate from the pattern identifier and feature parameters in step S3, and the stage label and threshold state in step S4. Therefore, each signal corresponds to a set of feature fields that can be used to select the intervention type. For example, the pattern identifier points to categories such as overcharge-related or mass transfer-limited related; the stage label points to slow change, development, or acceleration stages; and the feature parameter fragment points to the emphasis direction of abnormal reaction rate peaks or expanded concentration gradients. The intervention instruction library can be pre-set in the embodiments, including at least current limiting and load reduction instructions, charging cutoff target reduction instructions, balancing strategy adjustment instructions, and temperature control setpoint adjustment instructions. Each instruction includes applicable conditions, action objects, upper and lower limits of amplitude, duration, and mutual exclusion constraints. The action objects can correspond to maintenance configuration items such as power output adjustment, charge / discharge cycle adjustment, or temperature control. When generating the initial intervention instruction sequence, each high-risk signal is mapped to the applicable conditions in the instruction library, a candidate instruction set is obtained, and then a sequence is generated according to the triggering timing and risk weight ranking rules. The sequence elements include "instruction identifier—branch identifier—trigger time—initial amplitude parameter". If multiple candidate instructions exist at the same trigger time, they are sorted by risk weight and the instruction with the most direct impact on security risk is selected for priority execution based on degradation characteristics, thus forming an initial intervention instruction sequence.

[0055] The matching degree analysis between the initial intervention command sequence and high-risk interval data is used to verify whether the command covers the risk interval and to assess whether the intensity of the action is adapted to the risk evolution. The matching degree can be defined as the time difference between the command trigger time and the starting point of the high-risk interval, the overlap ratio between the command duration and the length of the risk interval, and the correspondence between the expected impact of the command and the change in risk elements. In implementation, the trigger time and duration parameters of each command are read, and interval overlap calculation is performed with the high-risk interval of the same signal to obtain the time coverage index. At the same time, the over-threshold amplitude or growth rate of the predicted trajectory within the risk interval is read as the risk intensity index. Then, the action amplitude parameter corresponding to the command in the command library is mapped to the expected suppression intensity index, and the intensity matching index is formed by the ratio or difference. The time coverage index and the intensity matching index are combined to obtain the matching degree score. For example, the matching degree score is obtained by weighted summation of the time coverage index and the intensity matching index according to preset weights, or by multiplying the two. The matching degree score corresponds one-to-one with the command. If the score is too low, it means that the trigger is too late or the action is too weak. If the score is too high and the corresponding risk intensity is low, it means that the action is too strong and may introduce bus disturbance or shorten the life. This process establishes a data link between "predicted risk range" and "instruction timing and intensity," providing a basis for adjusting the magnitude of actions and avoiding mismatches caused by setting based solely on experience.

[0056] The action amplitude adjustment updates the amplitude parameters of corresponding commands in the initial command sequence based on the matching degree analysis results. Amplitude parameters can be expressed as power output adjustment, charge / discharge cycle adjustment, or temperature control setpoint adjustment, etc., and each amplitude parameter has units and upper / lower limit constraints. The update rule can adopt a segmented mapping method, mapping the matching degree score and risk intensity index to amplitude coefficients, then multiplying the amplitude coefficient by the initial amplitude to obtain the adjustment amplitude, and trimming within the upper and lower limits to meet equipment constraints and safety specifications. The amplitude coefficient can be denoted as Q_m, corresponding to the m-th command, and the adjustment amplitude can be denoted as U_m, where U_m = Q_m × U_m0, and U_m0 represents the initial amplitude parameter. Q_m can increase with rising risk intensity, increase with insufficient matching degree, and decrease with excessively high matching degree, thereby enhancing "weak commands" and converging "strong commands." The updated amplitude parameters correspond to the command identifiers, forming the adjusted command sequence.

[0057] Command sequence optimization is performed after action amplitude adjustment to meet resource constraints, reduce command conflicts, and improve coverage of high-risk areas. During optimization, consistency checks are performed on the command sequence, duplicate commands from the same branch in adjacent time periods are merged, and commands from different branches that may cause bus power superposition in the same time period are staggered. Simultaneously, temperature control commands and charge / discharge adjustment commands are rearranged according to their dependencies, ensuring that temperature control adjustments take effect before high-load discharge or high-current charge / discharge. In implementations without a cost function, commands are rearranged in descending order of risk weight and ascending order of trigger timing, and conflicting commands are allocated by time slot to achieve peak shifting, resulting in an optimized command sequence. Optionally, a cost function is introduced to measure the cost of bus disturbance and the benefit of risk reduction. Rule iteration or local search is performed based on amplitude parameters, trigger time, and risk weight to achieve sequence rearrangement and amplitude fine-tuning. Preferably, only the adjustment amplitude can be used to reorder and merge the initial sequence, for example, placing commands with enhanced amplitude at the beginning of the sequence to generate an optimized command sequence for optimization of high-risk areas. The optimized instruction sequence retains the branch identifier, trigger time, and amplitude parameters of each instruction, and the sequence order reflects the comprehensive result of risk weight priority and trigger timing constraints.

[0058] S6. Evaluate and optimize the instruction sequence. If the optimized instruction sequence meets the preset alarm conditions, determine whether to activate the alarm state through a feedback loop mechanism.

[0059] In one specific embodiment, the process of performing step S6 may specifically include the following steps: Obtain the specific content of the optimized instruction sequence and evaluate the optimized instruction sequence through preset alarm conditions; If the optimized instruction sequence meets the preset alarm conditions, the intervention magnitude of the optimized instruction sequence will be compared and analyzed with the current peak change data of the chemical reaction rate. Determine whether to activate the alarm state based on the comparison and analysis results. If so, record the key parameters that triggered the alarm. Key parameters are fed back to the instruction execution loop mechanism to update subsequent intervention plans.

[0060] Specifically, after the lithium battery packs in the data center are connected to the UPS DC bus, changes in the battery pack status are coupled with the power supply and distribution scheduling strategy. The execution of intervention commands will change the charging and discharging rhythm, power output, and temperature control settings, thereby having a feedback impact on the chemical reaction rate and electrolyte state. If there is a lack of evaluation of the command sequence and alarm triggering mechanism, two types of problems are likely to occur: first, potential degradation continues to develop but only remains at the control and adjustment level without entering the alarm handling process; second, short-term anomalies caused by operating condition disturbances are misjudged and trigger unnecessary alarms, resulting in operational and maintenance pressure and frequent switching of power supply strategies. Taking the optimization of command sequence as the object, the alarm status is evaluated by preset alarm conditions and combined with feedback loop to form a closed loop relationship between alarm triggering and command execution. The trigger information is written back to update the subsequent intervention plan, realizing the corresponding handling of the problem of insufficient linkage between monitoring results, alarms, and operation and maintenance.

[0061] The elements of the optimized instruction sequence include branch identifier, trigger time, duration, action object, and action amplitude parameters. The action object can correspond to power output adjustment, charge / discharge cycle adjustment, or temperature control setpoint adjustment, etc. After obtaining the specific content of the sequence, it is converted into an evaluable data structure, so that each instruction corresponds to a set of evaluation fields, such as action amplitude, expected impact direction, execution window, and resource occupancy flag, and is aligned with the risk weight, mode identifier, and current operating condition label of the branch. Preset alarm conditions are stored in the form of rule sets. The rule set includes at least action amplitude threshold, risk weight threshold, instruction density threshold, key indicator anomaly threshold, and persistence threshold, which are used to determine whether the instruction sequence enters the alarm evaluation branch. The persistence threshold is the number of consecutive windows or the duration. The evaluation process traverses the optimized instruction sequence and calculates sequence-level features, such as the total cumulative action amplitude within a given evaluation window, the trigger frequency of instructions in the same branch, and the potential disturbance level to the bus power, etc. At the same time, instruction-level features are extracted, such as whether the amplitude of a single instruction exceeds the allowable range of the equipment, and whether a single instruction is marked as "forced suppression", etc. Sequence-level and instruction-level features are compared with corresponding thresholds in the alarm conditions to obtain a set of satisfied indicators. When the set of satisfied indicators includes conditions such as "high-risk trigger" or "strong intervention trigger," the optimized instruction sequence is determined to meet the preset alarm conditions and enters the comparison and analysis process. This design makes alarm judgment no longer solely dependent on original monitoring indicators, but incorporates the "intervention intensity driven by risk assessment results" into the alarm triggering logic, solving the problem of missed alarms or delayed responses caused by alarms based solely on external thresholds.

[0062] The peak change data of the chemical reaction rate comes from real-time sampling of the sensor array and the feature extraction link in step S1. At the current moment, it forms a parameter sequence of changes in peak amplitude, peak width, or peak spacing, and is aligned to the command trigger time axis with time labels. The comparative analysis uses "command amplitude - peak change of reaction rate" as the input pair. Commands are mapped to corresponding sampling channels according to branch identifiers, and data segments within the same time period are extracted from the peak change sequence according to the trigger time, forming a paired dataset so that each command corresponds to a segment of peak change data. The comparison logic focuses on two relationships: first, if the amplitude is large enough but the peak change continues to increase, it means that the intervention has not suppressed the abnormal reaction; second, if the amplitude is small but the peak change is close to the abnormal threshold, it means that the treatment level needs to be upgraded. In implementation, comparison indicators can be constructed, such as calculating the attenuation or increase of the peak change within a preset response delay after command triggering, and comparing its relationship with the command amplitude. The attenuation can be obtained from the difference in peak change parameters before and after triggering, and the increase can be obtained from the trend slope. Both, along with the amplitude parameter, constitute the judgment input. The comparison output can be denoted as G_r, where G_r is the comparison score for the r-th instruction. The score is obtained by mapping the amplitude parameter and the peak change response according to rules. A higher score indicates a higher degree of "strong intervention still cannot suppress" or "weak intervention has approached danger". This comparison establishes a link between control actions and internal response, so that alarms are no longer triggered by static thresholds, but are based on a linkage judgment of "intervention effectiveness" and "abnormal intensity", thereby suppressing false alarms caused by operating condition disturbances and improving sensitivity to continuous degradation.

[0063] The activation of the alarm state is based on the comparison score and feedback loop conditions. The feedback loop uses a rolling window to statistically analyze the comparison score and peak change trend. When the persistence threshold is met, the alarm state is activated, and the trigger parameters are written into the intervention strategy table to update the subsequent instruction amplitude and trigger lead. The feedback loop mechanism maintains the alarm criteria in a time-rolling manner, including continuous window statistics of the comparison score, persistence indicators of peak change data, and instruction execution status feedback. In implementation, the comparison score sequence is aggregated within the evaluation window, for example, taking the maximum value, average value, or the number of times the score threshold is exceeded, and the aggregation result is compared with the alarm trigger threshold. At the same time, the duration of continuous exceeding the threshold or the number of consecutive rising windows for peak change data are calculated as persistence evidence. The alarm state is activated when both the aggregation result and persistence evidence meet the conditions. If only a single high score occurs but persistence evidence is insufficient, the inactive state is maintained and observation continues, thereby avoiding alarms caused by short-term spikes or sampling jitter. When an alarm is activated, key parameters that trigger the alarm are recorded. These key parameters correspond to the aforementioned data links and may include branch identifiers, pattern identifiers, risk weight values, instruction identifiers, instruction amplitude parameters, comparison scores, peak change parameter segments, and trigger timestamps. Each parameter maintains a traceable mapping to the original data source, ensuring that subsequent intervention updates can pinpoint the specific object and cause. When key parameters are fed back to the instruction execution loop mechanism, they are written into the control strategy module or BMS strategy table in the form of feedback packets. These feedback packets are used to update subsequent intervention plans, such as improving the amplitude coefficient mapping of signals of the same pattern, shortening the trigger lead, adjusting peak shifting time slots, or upgrading the instruction category to a higher handling level, enabling subsequently generated instruction sequences to adaptively correct using alarm trigger information.

[0064] S7. Adjust the power dispatch strategy according to the alarm status and determine the health maintenance configuration. The health maintenance configuration includes at least power output regulation, charge and discharge cycle optimization and battery temperature control.

[0065] Specifically, the alarm status is output from the preceding evaluation and feedback loop, carrying data such as branch identifier, mode identifier, risk weight value, comparison score, trigger timestamp, and key parameter fragments. Power scheduling strategy adjustments use this alarm status as input and are bound to the data center's power distribution topology, causing scheduling actions to apply to the corresponding battery cluster, UPS channel, or DC bus segment. The processing maps the alarm status to a scheduling level label, which is determined by the risk weight range, the aggregation of the comparison score, and the duration of continuous over-threshold, and is associated with a scheduling rule table. The rule table provides a selectable set of power output adjustment parameters, charge / discharge cycle parameters, and temperature control parameters for different levels. Power output regulation is centered on bus power distribution and battery cluster output limits. The scheduling module reads UPS load power, bus power change rate, available capacity of redundant channels, and temperature control equipment capabilities as scheduling constraints. Combined with the battery cluster's allowable current range, it lowers the output limit corresponding to the alarm branch and distributes the load to redundant branches or bypass channels. Simultaneously, the adjusted output limit is written into the power output control instruction set, resolving the issue of degraded branches continuing to bear high loads, which could lead to increased risks. Charge and discharge cycle optimization aims to reduce the degradation acceleration factor. The scheduling module selects a cycle strategy based on the mode identifier. If the mode identifier points to overcharge, the target charging cutoff voltage is lowered and the trickle phase is extended. If the mode identifier points to mass transfer limitation, the charge and discharge rate is lowered and a rest phase is added. Cycle parameters are represented by cycle length, charging phase ratio, discharge window length, and equalization frequency, and are updated synchronously with the battery cluster control strategy table, addressing the issue of relying solely on alarms without actionable maintenance actions. Battery temperature control is achieved through temperature control setpoint and air / liquid cooling power adjustment. The scheduling module reads temperature sensor data, temperature control device status and ambient temperature, and adjusts the temperature control parameters to a lower temperature limit and increases heat dissipation power in combination with alarm level. At the same time, temperature constraint triggering conditions are added to the charging and discharging strategy to link current limit with temperature change, so as to avoid the risk spread caused by abnormal reaction rate peak superposition temperature rise.

[0066] The health maintenance configuration consists of power output adjustment parameters, charge / discharge cycle parameters, and temperature control parameters encapsulated under the same branch identifier and effective time label to form a configuration record, which is written into the health maintenance configuration library for subsequent continuous monitoring and retrieval. The configuration record is associated with key parameters of the alarm status, so that subsequent monitoring data can be aligned and verified according to the configuration version, thereby linking alarm triggering with scheduling actions and converting the scheduling results into traceable maintenance configurations.

[0067] Preferably, the health maintenance configuration, in addition to power output regulation, charge / discharge cycle optimization, and battery temperature control, may also include balancing strategy configuration, including balancing start / stop conditions, balancing current / power, balancing duration, balancing interval, and thresholds triggered by individual cell voltage difference or SOC difference, to suppress the risk of local overcharging and over-discharging caused by inconsistencies in series-connected cells; charging cutoff and voltage / current limiting configuration, including charging cutoff voltage target, trickle stage threshold, charging current upper limit, discharging current upper limit, pulse current limiting curve, etc., to reduce stress and control abnormal peak response rates when degradation signals appear; SOC working window and reserve capacity reservation configuration, including SOC upper / lower limits, reserve capacity reservation ratio, and core capacity trigger threshold, to maintain redundancy during risk phases and avoid accelerated degradation caused by deep discharge; isolation and bypass strategy configuration, including fault branch isolation conditions, contactor / circuit breaker operation conditions, cluster-level bypass switching conditions, and grid restoration conditions, to achieve segmented management of battery clusters when local risks increase.

[0068] S8. Continuously monitor the battery pack status and update data feedback according to the health maintenance configuration, thereby optimizing the battery pack operation status.

[0069] In one specific embodiment, the process of executing step S8 may specifically include the following steps: Acquire and execute health maintenance configurations, and continuously monitor and analyze changes in electrolyte concentration in lithium battery packs; The newly identified inflection point information of the degradation curve is integrated with the periodic pattern of electrolyte concentration change obtained from continuous monitoring and analysis to update the monitoring dataset; The length of the prediction time window is dynamically adjusted based on the rate of change of battery state reflected in the updated monitoring dataset. Use the adjusted forecast time window length to obtain new real-time feedback data; The operating parameters of the lithium battery pack are optimized and adjusted based on the new real-time feedback data, and an operating status report reflecting the latest operating status and optimization effect is generated.

[0070] Specifically, the health maintenance configuration is generated and distributed to the power scheduling and battery management execution link driven by alarm status. The configuration content includes power output adjustment parameters, charge / discharge cycle parameters, and temperature control parameters, and is bound to branch identifiers, configuration version numbers, and effective time tags. In data center scenarios, the same battery cluster may be in a state of current-limited discharge, reduced charging cutoff target, or increased heat dissipation power after the configuration takes effect. If there is a lack of continuous monitoring and feedback updates, the configuration may become mismatched due to changes in operating conditions. Step S8 continuously collects status data with the health maintenance configuration as a constraint and feeds the feedback back to the prediction and control parameter updates, so that the configuration execution effect can be verified at the data level and drive the adjustment of operating parameters.

[0071] After the execution link is initiated, the health maintenance configuration is written into the scheduling control table and the BMS strategy table, triggering an execution status receipt. The receipt includes the configuration version number, the instruction effective time, and key execution quantities, such as current limit value, charging stage ratio, and temperature control setpoint. The receipt and real-time sampled data are aligned using the same time tag to distinguish between the "pre-configuration status segment" and the "post-configuration status segment." Continuous monitoring and analysis focus on changes in electrolyte concentration because electrolyte concentration and its spatial distribution are sensitive to mass transfer limitations, local polarization, and the evolution of side reactions, and exhibit periodic patterns and amplitude changes after configuration adjustments. The sensor array continuously outputs multi-point concentration data. The data processing link organizes the multi-point concentration into a time-space vector according to location and calculates the gradient distribution difference characteristics. Simultaneously, it calculates the periodic parameters of the concentration sequence over time. The periodic parameters can be obtained from the autocorrelation peak spacing or the frequency domain main peak period and are recorded synchronously with the configuration version number. This allows the periodic changes in concentration under the same configuration to be observed and correlated with changes in configuration parameters, solving the problem that it is difficult to evaluate the effectiveness of maintenance actions based solely on external electrical parameters.

[0072] The inflection point information of the degradation curve comes from the degradation curve analysis link. This information includes the inflection point time label, the slope level before and after the inflection point, and stage labels, reflecting where the degradation evolution rate undergoes structural changes. Integrating the inflection point information with the periodic patterns of concentration changes is a process of aligning "degradation evolution events" with "concentration periodic responses." During integration, the time label is used as the primary key. Periodic parameter segments near the inflection point time are extracted, and the periodic changes before and after the inflection point are calculated, such as period length drift, period amplitude changes, or changes in the mean gradient difference. These changes are written into the monitoring dataset as event annotation fields, expanding the monitoring dataset from a simple concentration time series record to a structured dataset of "concentration time series + periodic features + inflection point event annotations." Updates to the monitoring dataset are incremental. New records carry configuration version numbers and event annotations, enabling subsequent analysis to distinguish response differences under different configurations and locate the correspondence between accelerated degradation phases and concentration periodic anomalies, thus supporting the continuous characterization of the coupling between degradation evolution and mass transfer states.

[0073] The dynamic adjustment of the prediction time window length is based on the rate of state change reflected in the monitoring dataset. This rate of change can be derived from the drift rate of concentration periodic parameters, the growth rate of gradient differences, or the absolute value of the slope difference before and after an inflection point. These derived quantities correspond to time labels to form a rate sequence. The dynamic adjustment logic maps the rate sequence to a window length. When the rate increases, the window is shortened to improve the response to rapid changes; when the rate decreases, the window is lengthened to smooth short-term fluctuations and improve trend stability. The mapping can use a piecewise rule table or a continuous function mapping. The rule table stores the window length interval corresponding to the rate threshold interval, while the continuous mapping is constructed with the window length monotonically decreasing with the rate, ensuring that the change in window length is consistent with the strength of state changes. The adjusted window length is bound to the configuration version number and written to the prediction parameter table, so that subsequent feedback sampling and analysis are aggregated using this window length. This directly feeds back the changes in the monitoring dataset to the prediction scale, solving the problems of lag in the rapid change phase and noise amplification in the slow change phase caused by a fixed window.

[0074] When acquiring real-time feedback data based on the adjusted prediction time window length, the real-time sampled data is aggregated by window to form a set of feedback indicators. This set may include updated values ​​of concentration cycle parameters, statistical values ​​of gradient differences, peak chemical reaction rate changes, and temperature control execution deviations, and is aligned with the configured execution quantities within the same window. This provides a data basis for optimizing and adjusting operating parameters. During the optimization process, the set of feedback indicators is read and compared with the configuration target or safety boundary. If the concentration gradient difference continues to increase, the charge / discharge rate is further reduced or the rest period is extended. If the concentration cycle amplitude is abnormal and the temperature is too high, the temperature control setpoint is lowered and the heat dissipation power is increased. If the peak reaction rate change remains high after current limiting, the power output allocation is adjusted and the load ratio of that cluster is reduced. The adjustment results are written to the control strategy table in the form of updated operating parameters and a derived version number of the configuration version is generated, ensuring that parameter updates are traceable and rollbackable. The operational status report is generated in each feedback window or preset period. The report content consists of configuration version number, execution quantity receipt, feedback indicator set, status change rate sequence, window length adjustment record and operational parameter update record, and is summarized by branch identifier. The report data and the monitoring dataset share time tag index, which facilitates subsequent steps to continue to perform degradation trend analysis and risk weight update, and realizes the closed-loop association of monitoring, maintenance configuration and feedback calibration.

[0075] The above describes a method for monitoring the health status of lithium battery packs in a data center according to an embodiment of this application. The following describes a system for monitoring the health status of lithium battery packs in a data center according to an embodiment of this application. Please refer to [link / reference]. Figure 2 This application provides a structural schematic diagram of a data center computer room lithium battery pack health status monitoring system, which includes: The data acquisition module 10 is used to collect chemical reaction data and electrolyte data inside the lithium battery pack, extract features from the chemical reaction data and electrolyte data, and obtain the original index sequence. Trend analysis module 20 is used to analyze the fluctuation trend of the original indicator sequence and determine the degree of deviation of each indicator from the benchmark. The anomaly identification module 30 is used to extract abnormal features and perform pattern matching when the deviation exceeds a preset deviation threshold to determine whether there is a potential degradation signal. The risk prediction module 40 is used to analyze the degradation trend and simulate the future evolution trajectory for potential degradation signals, and to calculate and determine the risk weight allocation priority of each signal. The intervention optimization module 50 is used to generate an intervention action adjustment plan based on the priority allocation of risk weights, and to sort and optimize the various intervention instructions to generate an optimized instruction sequence. The alarm evaluation module 60 is used to evaluate the optimized instruction sequence. If the optimized instruction sequence meets the preset alarm conditions, the alarm state is activated through a feedback loop mechanism. The scheduling and adjustment module 70 is used to adjust the power scheduling strategy according to the alarm status and determine the health maintenance configuration. The health maintenance configuration includes at least power output regulation, charge and discharge cycle optimization and battery temperature control. The feedback optimization module 80 is used to continuously monitor the battery pack status and update the data feedback according to the health maintenance configuration, thereby optimizing the battery pack operation status.

[0076] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for monitoring the health status of lithium battery packs in a data center, characterized in that, The method includes: S1. Collect chemical reaction data and electrolyte data inside the lithium battery pack, extract features from the chemical reaction data and electrolyte data, and obtain the original index sequence; S2. Analyze the fluctuation trend of the original indicator sequence and determine the degree of deviation of each indicator from the benchmark; S3. If the deviation exceeds the preset deviation threshold, extract abnormal features and perform pattern matching to determine whether there is a potential degradation signal. S4. For the potential degradation signals, analyze the degradation trend and simulate the future evolution trajectory, calculate and determine the risk weight allocation priority of each signal; S5. Generate an intervention action adjustment plan based on the risk weight allocation priority, and sort and optimize each intervention instruction to generate an optimized instruction sequence; S6. Evaluate the optimized instruction sequence. If the optimized instruction sequence meets the preset alarm conditions, determine whether to activate the alarm state through a feedback loop mechanism. S7. Adjust the power scheduling strategy according to the alarm status and determine the health maintenance configuration, which includes at least power output regulation, charge and discharge cycle optimization and battery temperature control. S8. Continuously monitor the battery pack status and update data feedback according to the health maintenance configuration, thereby optimizing the battery pack operation status.

2. The method according to claim 1, characterized in that, S1 includes: A sensor array is deployed in a key electrochemical region inside the lithium battery pack to acquire real-time data on chemical reaction rates and electrolyte concentrations within the pack. Extract the peak change characteristics of the chemical reaction rate data over time, and extract the gradient distribution difference characteristics of the electrolyte concentration data in the internal space of the battery; The peak variation characteristics and the gradient distribution difference characteristics are correlated and integrated to generate the original micro-indicator sequence; The original micro-indicator sequence is cleaned and formatted to obtain the original indicator sequence.

3. The method according to claim 2, characterized in that, S2 include: Within a set time window, fluctuation frequency analysis is performed on the original index sequence to identify the locations of abrupt changes in chemical reaction rates and the periodic patterns of electrolyte concentration changes in the original index sequence. Based on the location of the mutation point and the periodic pattern of change, the abnormal fluctuation range in the original indicator sequence is determined; Calculate the deviation of the abnormal fluctuation range from the preset baseline to quantify the initial degree of trend deviation; The preliminary trend deviation is compared and verified with historical health data trends, and the preliminary trend deviation is calibrated and updated based on the verification results to obtain the final deviation.

4. The method according to claim 3, characterized in that, S3 includes: Determine whether the degree of deviation exceeds a preset deviation threshold. If so, extract the coupling relationship features between chemical reaction rate and electrolyte concentration from the abnormal fluctuation range. Based on the coupling relationship features and the gradient distribution difference features, a comprehensive state feature vector is constructed. The comprehensive state feature vector is then matched and analyzed with a preset degradation mode feature library to obtain the mode matching results. Based on the pattern matching results, determine whether there are potential degradation signals in the current battery state; If a potential degradation signal is present, the characteristic parameters of the abnormal feature are recorded. The characteristic parameters include at least the chemical reaction rate characteristics and the electrolyte concentration gradient characteristics. The characteristic parameters are correlated with historical degradation data to generate a preliminary degradation signal report.

5. The method according to claim 1, characterized in that, S4 include: The characteristic parameters of the potential degradation signal are obtained, a degradation curve is constructed based on the characteristic parameters, and the slope change characteristics of the degradation curve are analyzed by degradation curve inflection point identification technology. The degradation threshold of the degradation trend is determined based on the slope change characteristics, and the degradation threshold is used to distinguish the degree of degradation at different stages. Based on the degradation threshold and the preset prediction time window length, the future evolution trajectory of the degradation trend is simulated to obtain prediction trajectory data. Based on the predicted trajectory data, the degree of risk that the potential degradation signal may cause within the predicted time window is quantitatively calculated, and the corresponding degradation risk weight value is obtained. Based on the degradation risk weight values, multiple potential degradation signals are prioritized and sorted to determine the risk weight allocation priority of each signal, and a risk weight allocation result is generated.

6. The method according to claim 1, characterized in that, S5 include: Obtain the results of the risk weight allocation priority, and extract the interval data corresponding to the high-risk level from the results; Based on the data from high-risk zones, the timing for triggering intervention instructions is determined. Based on the triggering timing and the degradation characteristics reflected by the high-risk interval data, the intervention instructions are prioritized to generate an initial intervention instruction sequence; Analyze the matching degree between the initial intervention instruction sequence and the high-risk interval data, and adjust the action amplitude of the corresponding instructions in the initial intervention instruction sequence based on the analysis results; The initial intervention instruction sequence is optimized based on the adjusted intervention action amplitude to generate the optimized instruction sequence.

7. The method according to claim 1, characterized in that, S6 include: Obtain the specific content of the optimized instruction sequence, and evaluate the optimized instruction sequence based on the preset alarm conditions; If the optimized instruction sequence meets the preset alarm conditions, the intervention magnitude of the optimized instruction sequence is compared and analyzed with the current peak change data of the chemical reaction rate. Determine whether to activate the alarm state based on the comparison and analysis results. If so, record the key parameters that triggered the alarm. The key parameters are fed back to the instruction execution loop mechanism to update subsequent intervention plans.

8. The method according to claim 5, characterized in that, S8 includes: The health maintenance configuration is acquired and executed, and the changes in electrolyte concentration of the lithium battery pack are continuously monitored and analyzed. The newly identified inflection point information of the degradation curve is integrated with the periodic pattern of electrolyte concentration change obtained from continuous monitoring and analysis to update the monitoring dataset; The length of the prediction time window is dynamically adjusted based on the rate of change of battery state reflected in the updated monitoring dataset. Use the adjusted forecast time window length to obtain new real-time feedback data; The operating parameters of the lithium battery pack are optimized and adjusted based on the new real-time feedback data, and an operating status report reflecting the latest operating status and optimization effect is generated.

9. A data center server room lithium battery pack health status monitoring system, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The data acquisition module is used to collect chemical reaction data and electrolyte data inside the lithium battery pack, extract features from the chemical reaction data and electrolyte data, and obtain the original index sequence. The trend analysis module is used to analyze the fluctuation trend of the original indicator sequence and determine the degree of deviation of each indicator from the benchmark. An anomaly detection module is used to extract abnormal features and perform pattern matching when the deviation exceeds a preset deviation threshold to determine whether there is a potential degradation signal. The risk prediction module is used to analyze the degradation trend and simulate the future evolution trajectory for the potential degradation signals, and calculate and determine the risk weight allocation priority of each signal. The intervention optimization module is used to generate an intervention action adjustment plan based on the priority allocation of the risk weights, and to sort and optimize the various intervention instructions to generate an optimized instruction sequence. An alarm evaluation module is used to evaluate the optimized instruction sequence. If the optimized instruction sequence meets the preset alarm conditions, a feedback loop mechanism is used to determine whether to activate the alarm state. The scheduling and adjustment module is used to adjust the power scheduling strategy according to the alarm status and determine the health maintenance configuration, which includes at least power output regulation, charge and discharge cycle optimization and battery temperature control. The feedback optimization module is used to continuously monitor the battery pack status and update the data feedback according to the health maintenance configuration, thereby optimizing the battery pack operating status.

Images