Storage system bbu protection capability prediction method and system based on digital twinning

Abstract

This application discloses a method and system for predicting the protection capability of a storage system's Battery Backup Unit (BBU) based on digital twins. The method includes: receiving telemetry data locally reported by the battery backup unit; preprocessing the telemetry data and constructing a real-time operational status snapshot; using dual extended Kalman filtering, combined with the real-time operational status snapshot, to identify the health status parameters of the battery backup unit online, generating health status time-series data; based on the health status time-series data, using a time-series prediction model to predict the health status decay trend of the battery backup unit and calculating the predicted remaining protection time; sending the predicted remaining protection time to a hierarchical policy engine; and the hierarchical policy engine dynamically adjusting the three-level cache write policy according to the predicted remaining protection time. This application can maximize storage performance while ensuring data security, solving the problems of lack of prediction, policy decoupling, and insufficient hierarchical protection in existing technologies.

Description

Technical Field

[0001] This application relates to the field of data protection technology for enterprise-level storage systems, specifically to a method and system for predicting the protection capabilities of a storage system's BBU based on digital twins. Background Technology

[0002] Enterprise-level storage systems commonly employ write-back caching strategies to improve I / O (Input / Output) performance. Write operations first write to a high-speed DRAM (Dynamic Random Access Memory) cache, and then asynchronously flush the data to backend non-volatile storage media. In this scenario, if an unexpected power outage occurs, dirty data in the DRAM that has not yet been persisted risks being lost. To address this, storage systems are typically equipped with a BBU (Battery Backup Unit), which takes over the controller cache power supply during a power failure, ensuring that cached data can be safely written to the non-volatile storage media, thereby guaranteeing data integrity.

[0003] As a critical component in the data protection chain of a storage system, the health status of the BBU directly determines the reliability of power-loss protection. Existing BBU management technologies mainly cover the following types of solutions.

[0004] (1) Traditional BBU monitoring and threshold alarm solution

[0005] Existing enterprise-level storage system BBU solutions (such as Dell PowerStore and NetApp FAS / AFF series built-in BBU modules) generally provide basic battery monitoring functions, mainly including voltage detection, temperature monitoring, current sampling, and SOC (State of Charge) estimation. When the battery's SOH (State of Health) drops to a preset threshold (e.g., 80%), the BMS (Battery Management System) triggers a replacement alarm. This type of solution is essentially a passive management mode: the BMS only monitors the battery's real-time status and issues over-limit alarms, and cannot predict the degradation trend of the BBU's protection capability based on historical operating data. Maintenance personnel can only respond passively after a fault occurs or an alarm is triggered, making it difficult to plan maintenance windows in advance.

[0006] (2) Battery SOC / SOH estimation method

[0007] Numerous battery state estimation methods have been proposed by academia and industry to improve the accuracy of SOC and SOH estimations. For example, patent CN119511105B discloses a comprehensive battery health state assessment method that evaluates battery SOH through multi-dimensional parameter fusion. Existing technologies have also proposed joint SOC / SOH prediction models based on LSTM (Long Short-Term Memory) and extended Kalman filtering, utilizing temporal neural networks to capture battery aging characteristics. However, these methods only provide battery state estimation itself; their output is not further transformed into control inputs for storage system data protection strategies, and they do not form a closed loop with the write-back / write-through strategy switching mechanism of the storage system. A significant disconnect exists between state estimation and system decision-making.

[0008] (3) Power allocation method for energy storage system

[0009] Patent document CN115800342A discloses an active power allocation method for AGC (Automatic Generation Control) in energy storage power stations based on power allocation factors, which comprehensively considers factors such as SOC, charging and discharging power limits, and battery aging to extend battery life. Patent document WO2025097784A1 discloses a control method for hybrid energy storage systems, which achieves energy dispatch among multiple types of energy storage media by optimizing power allocation strategies. These solutions are geared towards power dispatch scenarios in grid-scale energy storage power stations, aiming to improve the energy conversion efficiency and economy of large-scale energy storage systems. They are not designed for power-down data protection scenarios in storage system BBUs, and their algorithm complexity and resource consumption are not suitable for the operating environment of embedded BBU controllers.

[0010] (4) Application of digital twins in the field of energy storage

[0011] Existing technologies have proposed digital twin modeling methods for energy storage power plants. These methods construct high-fidelity digital mirror models of physical entities to simulate and optimize the operation of large-scale energy storage power plants in real time. However, such solutions are designed for MWh-level large-scale energy storage power plants. The construction and operation of their digital twin models require substantial computing resources and data infrastructure support, making them unsuitable for direct deployment on resource-constrained embedded BBU controllers or BMC (Baseboard Management Controller) platforms. Furthermore, their optimization objectives differ fundamentally from the power failure protection requirements of the storage system's BBU.

[0012] Based on the above analysis of existing technologies, current management technologies for storage system BBUs have the following four shortcomings:

[0013] (1) Passive operation and maintenance, lack of predictive ability: The existing BBU management solution adopts a passive response mode, and problems can only be discovered after the SOH drops to the alarm threshold or the battery fails. Operation and maintenance personnel cannot know the decay trend of BBU protection capability, and it is difficult to predict the risk window in advance. This results in the maintenance response lagging behind the equipment aging process, and there is a hidden danger of data protection vacuum due to sudden BBU failure.

[0014] (2) Protection strategy decoupled from battery health status: The write-back / write-through strategy switching of the existing storage system is based solely on whether the remaining power of the BBU meets the single threshold condition of 24-hour data retention, without comprehensively considering the impact of aging factors such as battery SOH decay and ESR (Equivalent Series Resistance) growth on the actual protection capability of the BBU. When the battery is highly aged but the power has not yet reached the threshold, the system still maintains write-back mode, which poses a hidden risk that the actual protection capability is insufficient and the system fails to detect it in time.

[0015] (3) Lack of hierarchical protection mechanism and coarse granularity of safety and performance trade-off: The existing solution only performs binary strategy switching (Write-Back to Write-Through) after BBU aging, lacking intermediate transition levels and hierarchical degradation strategies based on battery health status. This either-or switching method cannot achieve a fine dynamic balance between safety and I / O performance, causing the system to either bear excessive data risk in exchange for performance, or sacrifice too much I / O performance due to overly conservative strategy switching.

[0016] (4) Lack of transparency in operation and maintenance information and lack of data support for maintenance plans: The existing solution fails to provide operation and maintenance personnel with key decision-making information such as the remaining available protection time of the BBU, the evolution trend of its health status, and the predictive replacement window. Operation and maintenance personnel rely solely on static threshold alarms to formulate maintenance plans, making it difficult to accurately assess the actual available lifespan of the BBU and the optimal replacement time. This results in unreasonable allocation of maintenance resources, increasing the risk of system downtime and maintenance costs. Summary of the Invention

[0017] To address this, this application provides a method and system for predicting the protection capability of a storage system's battery unit (BBU) based on digital twins, thereby solving the problems of insufficient prediction capability, decoupling of protection strategies from battery health status, and lack of graded protection mechanisms in existing technologies.

[0018] To achieve the above objectives, this application provides the following technical solution:

[0019] In a first aspect, a method for predicting the protection capability of a storage system's Base Unit (BBU) based on digital twins is provided, characterized in that the method is applied to a digital twin engine, which is deployed in the base unit management controller of the storage system, and the method includes:

[0020] Step 1: Receive telemetry data collected and reported locally by the battery backup unit; wherein, the battery backup unit is a hybrid topology structure in which battery energy storage unit and capacitor energy storage unit work together; the telemetry data includes: battery energy storage unit operation data, capacitor energy storage unit operation data and power distribution controller status word;

[0021] Step 2: Perform data preprocessing on the telemetry data, and construct a real-time operating status snapshot of the battery backup unit based on the preprocessed telemetry data; the data preprocessing includes: time alignment, outlier removal, temperature compensation, and auxiliary data fusion;

[0022] Step 3: Using a dual extended Kalman filter, combined with the real-time operating status snapshot, the health status parameters of the battery backup unit are identified online, and health status time series data is generated;

[0023] Step 4: Based on the health status time series data, use a time series prediction model to predict the health status degradation trend of the battery backup unit;

[0024] Step 5: Calculate and predict the remaining protection time based on the effective energy of the current battery energy storage unit and the capacitor energy storage unit, the current load information of the storage system, and the health status decay trend;

[0025] Step 6: Send the predicted remaining protection time to the hierarchical policy engine; the hierarchical policy engine dynamically adjusts the cache write policy according to the predicted remaining protection time; wherein, the dynamic adjustment of the cache write policy includes: a three-level hierarchical protection policy of write-back, write protection and direct write; the hierarchical policy engine is deployed in the I / O scheduling software stack of the storage controller.

[0026] As a preferred option, it also includes:

[0027] The predicted remaining protection time, the health status decay trend, and the health status parameters of the battery energy storage unit and the capacitor energy storage unit are sent to the operation and maintenance platform so that the operation and maintenance platform can generate and display the health status trend, remaining service life prediction, and replacement suggestion window of the battery backup unit, and push proactive warning information to the operation and maintenance platform when the predicted remaining protection time is lower than the preset warning threshold.

[0028] Preferably, in step 3, the dual extended Kalman filter includes: fast-timescale state of charge estimation and slow-timescale health state parameter identification, wherein the fast-timescale state of charge estimation and the slow-timescale health state parameter identification are coupled and closed-loop corrected; wherein, the fast-timescale state of charge estimation is updated once in each telemetry data reporting cycle, and the slow-timescale health state parameter identification is updated once after accumulating multiple reporting cycles, and the updated health state parameters are fed back into the model parameters of the fast-timescale state of charge estimation.

[0029] Preferably, step 4, which involves using a time-series prediction model to predict the health status degradation trend of the battery backup unit, specifically includes:

[0030] Using the daily feature values ​​in the health status time series data as input, the time series prediction model performs forward propagation prediction and outputs the health status decay trend for the next preset number of days. When the deviation between the predicted value of the time series prediction model and the health status parameters identified in real time by the dual extended Kalman filter exceeds a preset threshold, a calibration process is triggered. The input sequence of the time series prediction model is drift corrected based on the real-time identification result of the dual extended Kalman filter, and the prediction is re-executed.

[0031] Preferably, the time series prediction model adopts a long short-term memory network model; the long short-term memory network model is a single-layer structure with a hidden layer dimension of 32 or 64, and the input time step is a preset number of days.

[0032] Preferably, in step 5, the formula for calculating the predicted remaining protection time is:

[0033] RPT = E_total / P_total_discharge - t_total

[0034] Wherein, E_total represents the total effective energy of the battery energy storage unit and the capacitor energy storage unit, P_total_discharge represents the total power consumption during the power outage discharge period, and t_total represents the total time from the occurrence of the power outage to the complete flushing of the cached data.

[0035] Preferably, in step 6, the hierarchical strategy engine dynamically adjusts the cache write strategy based on the predicted remaining protection time, specifically including:

[0036] Read the current policy status, which includes: write-back mode, write-protected mode, and write-through mode;

[0037] If the current policy status is write-back mode and the predicted remaining protection time is less than or equal to 72 hours, the decision is to switch to write protection mode.

[0038] If the current policy status is write protection mode and the predicted remaining protection time is greater than or equal to eighty hours, the decision is to switch to write-back mode.

[0039] If the current policy status is write protection mode and the predicted remaining protection time is less than or equal to 24 hours, the decision is to switch to write-through mode.

[0040] If the current policy status is write-through mode and the predicted remaining protection time is greater than or equal to 30 hours, the decision is to switch to write protection mode.

[0041] If the minimum dwell time check is triggered, the decision is to maintain the current state;

[0042] In all other cases, the decision is to maintain the current state.

[0043] Preferably, the digital twin engine uses a multi-trigger mode for updates, which includes: timed updates, event-triggered updates, and exception-triggered updates.

[0044] Preferably, the battery energy storage unit is a lithium iron phosphate battery pack, and the capacitor energy storage unit is a supercapacitor.

[0045] Secondly, a storage system BBU protection capability prediction system based on digital twins includes a local MCU of the battery backup unit, a digital twin engine, a hierarchical strategy engine, and an operation and maintenance platform. The digital twin engine is communicatively connected to the local MCU of the battery backup unit, the hierarchical strategy engine, and the operation and maintenance platform. The digital twin engine is deployed in the baseboard management controller of the storage system to implement the aforementioned storage system BBU protection capability prediction method based on digital twins. The hierarchical strategy engine is deployed in the I / O scheduling software stack of the storage controller to dynamically adjust the cache write strategy.

[0046] Compared with the prior art, this application has at least the following beneficial effects:

[0047] 1. This application provides a method for predicting the protection capability of a storage system's Battery Backup Unit (BBU) based on digital twins, comprising: receiving telemetry data collected and reported locally by the battery backup unit; preprocessing the telemetry data and constructing a real-time operating status snapshot; using dual extended Kalman filtering, combined with the real-time operating status snapshot, to identify the health status parameters of the battery backup unit online and generate health status time series data; based on the health status time series data, using a time-series prediction model to predict the health status decay trend of the battery backup unit; calculating and predicting the remaining protection time based on the effective energy of the current battery energy storage unit and capacitor energy storage unit, the current load information of the storage system, and the health status decay trend; sending the predicted remaining protection time to a hierarchical strategy engine; and the hierarchical strategy engine dynamically adjusting the cache write strategy according to the predicted remaining protection time. The dynamic adjustment of the cache write strategy includes a three-level hierarchical protection strategy: write-back, write protection, and direct write. This application achieves advance prediction of BBU protection capability through digital twins and dynamically links the three-level hierarchical write strategy based on the prediction results, thereby maximizing storage performance while ensuring data security, and solving the problems of lack of prediction, strategy decoupling, and insufficient hierarchical protection in existing technologies.

[0048] 2. This application will send the predicted remaining protection time, health status degradation trend, and health status parameters of the battery energy storage unit and the capacitor energy storage unit to the operation and maintenance platform, so that the operation and maintenance platform can generate and display the health status trend of the battery backup unit, the prediction of the remaining service life and the replacement suggestion window, and push proactive warning information to the operation and maintenance platform when the predicted remaining protection time is lower than the preset warning threshold, so that the operation and maintenance information is fully visible and the maintenance plan is based on evidence. Attached Figure Description

[0049] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0050] Figure 1 A flowchart of a method for predicting the protection capability of a storage system BBU based on digital twins, provided in Embodiment 1 of this application;

[0051] Figure 2 This is a schematic diagram of the structure of a storage system BBU protection capability prediction system based on digital twin provided in Embodiment 2 of this application. Detailed Implementation

[0052] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0054] The terms used in this application, such as "upper," "lower," "left," "right," and "middle," are generally used to indicate the general relative positional relationship for the purpose of intuitive understanding by referring to the accompanying drawings, and are not absolute limitations on the positional relationship in the actual product.

[0055] This application provides a method and system for predicting the protection capabilities of a storage system's Battery Backup Unit (BBU) based on digital twins, applicable to the management and maintenance of BBUs in enterprise-level storage systems. Enterprise-level storage systems commonly employ a write-back caching strategy to improve write performance. This strategy temporarily stores host-written data in a DRAM (Dynamic Random Access Memory) cache, and then batch-flushes it to the back-end storage media when the system is idle. Under this strategy, the BBU, as the core component for power-loss protection, directly determines the reliability of its power-loss data protection capabilities based on its health status. When the storage system experiences a sudden power outage, the BBU needs to provide sufficient energy to completely flush the dirty data in the DRAM cache to non-volatile storage media (such as SSDs or NAND Flash) to prevent data loss or corruption.

[0056] The BBU addressed in this application adopts a hybrid energy storage architecture, namely a hybrid topology structure in which LFP (Lithium Iron Phosphate) battery packs and supercapacitors (also known as electric double-layer capacitors, EDLCs) work together. In this hybrid architecture, the LFP battery pack undertakes the main energy supply function, providing a longer duration of discharge capability to support the complete cache flushing process; the supercapacitor pack undertakes the instantaneous power buffering function, utilizing its high power density and fast charge / discharge characteristics to provide millisecond-level instantaneous power support at the moment of power failure, and providing a stable voltage transition for the start-up buffer of the LFP battery pack. The LFP battery pack and the supercapacitor pack work in coordination through an intelligent power distribution circuit: in the initial stage of power failure (0~100ms), the supercapacitor pack provides instantaneous response and peak power support; in the transition stage (100ms~500ms), the LFP battery pack starts up and takes over the main discharge load; in the continuous discharge stage (>500ms), the LFP battery pack continuously provides energy until the cached data is completely flushed. This hybrid architecture combines the advantages of high energy density of LFP batteries and high power density of supercapacitors, significantly improving the reliability and response speed of power failure protection.

[0057] As the battery pack (BBU) operates and ages, LFP battery packs exhibit capacity decay and increased internal resistance, while supercapacitor packs also show signs of aging, such as increased equivalent series resistance (ESR) and decreased electrostatic capacity. This leads to a gradual shortening of the actual protection time provided by the BBU. Traditional solutions rely on fixed thresholds or periodic manual inspections, which cannot dynamically detect subtle changes in the BBU's health status, posing a risk of insufficient protection while the system continues to operate in high-performance mode.

[0058] This application constructs a lightweight digital twin model based on the BBU's local BMS (Battery Management System) to estimate the BBU's SOC (State of Charge), SOH (State of Health), Remaining Protection Time (RPT), and future degradation trend in real time. The storage controller dynamically adjusts its cache write strategy (Write-Back / Write-Guard / Write) based on the predictions output by the digital twin.

[0059] -Through three-level hierarchical protection), to achieve hierarchical power failure protection. At the same time, predictive maintenance suggestions are output to the operation and maintenance platform, including the best replacement window and RUL (Remaining Useful Life) prediction. The core innovations of this application are: (1) full-state digital twin modeling of LFP+supercapacitor hybrid BBU; (2) three-level hierarchical protection strategy based on predicted remaining protection time; (3) embedded deployment scheme of lightweight digital twin model.

[0060] Example 1

[0061] Please see Figure 1 This embodiment provides a method for predicting the protection capability of a storage system's BBU based on digital twins. This method is applied to a digital twin engine, which is deployed in the baseboard management controller of the storage system. The method includes:

[0062] S1: Receive telemetry data collected and reported locally by the battery backup unit; wherein, the battery backup unit is a hybrid topology structure in which battery energy storage unit and capacitor energy storage unit work together; the telemetry data includes: battery energy storage unit operation data, capacitor energy storage unit operation data and power distribution controller status word;

[0063] Specifically, the battery backup unit (BBU) uses lithium iron phosphate (LFP) battery packs for its battery energy storage unit and supercapacitors for its capacitor energy storage unit. The battery management system (BMS) of the battery backup unit is responsible for real-time monitoring, local protection, and data reporting of the hybrid energy storage system. It also monitors the operating status of the LFP battery pack and the supercapacitor pack and coordinates their collaborative operation through an intelligent power distribution controller.

[0064] a. LFP battery pack monitoring function

[0065] The local MCU of the battery backup unit acquires the following parameters of the LFP battery pack through multiple ADCs (Analog-to-Digital Converters): voltage of each series cell (typically 8~16 series, range 2.5~3.8V, resolution 12bit, sampling accuracy ±5mV), total pack voltage (range 20~60V, accuracy ±50mV), total charging and discharging current (using Hall effect sensor or precision shunt, range ±10A, accuracy ±50mA, sampling rate 100Hz), and multi-point temperature (one NTC thermistor for every 4 cells, a total of 4~8 temperature sampling points, range -20℃~+60℃, accuracy ±1℃).

[0066] The battery backup unit's local MCU runs a lightweight ECM (Equivalent Circuit Model), which employs a first-order RC (Resistance-Capacitance) network structure, including an ohmic internal resistance R0, a polarization internal resistance R1, and a polarization capacitor C1. Based on the ampere-hour integration method combined with open circuit voltage (OCV) lookup table correction, local SOC estimation is achieved with an accuracy within ±10%. The local ECM simultaneously executes hard protection logic for overvoltage, undervoltage, overtemperature, and overcurrent. When any parameter exceeds the safety threshold, the power distribution controller immediately cuts off the corresponding discharge path, with a protection response time of <10ms.

[0067] b. Supercapacitor bank monitoring function

[0068] Supercapacitor bank monitoring is a key component of this embodiment. The local MCU monitors the supercapacitor bank for the following: individual cell voltages (typically 6-12 cells in series, range 0-2.8V, 12-bit resolution, ±3mV accuracy, individual cell rated voltage 2.5V or 2.7V), total voltage (range 0-32.4V, ±30mV accuracy), charging and discharging current (using a precision shunt, range ±30A, ±100mA accuracy, 1kHz sampling rate to capture millisecond-level transient currents), and multi-point temperature (2-4 NTC thermistors, range -40℃ to +65℃, ±1.5℃ accuracy).

[0069] The core monitoring parameters for supercapacitor banks are ESR (Equivalent Series Resistance) and capacitance C. This embodiment employs an online ESR detection algorithm: During periodic self-testing (typically every 24 hours, or each time the system is powered on), the power distribution controller controls the supercapacitor bank to perform pulse charging and discharging with a fixed current I_pulse (typically 1~3A, pulse width 100ms). The current ESR value is calculated by measuring the voltage change ΔV before and after the pulse, combined with the formula ESR=ΔV / I_pulse. Simultaneously, the C value is periodically detected using a constant current discharge method (discharging the supercapacitor from full voltage to half voltage with a fixed current I_discharge, recording the time Δt, and calculating the capacitance using the formula C=I_discharge×Δt / ΔV_total). The ESR detection results are used to assess the aging status of the supercapacitors—an early warning is triggered when the ESR increases to 200% of the initial value or the capacitance decays to 70% of the nominal value.

[0070] The supercapacitor bank's function in power-down protection is as follows: During the instantaneous power failure (0-100ms), the supercapacitor bank, utilizing its high power density (typically 10-15kW / kg), instantaneously provides a peak discharge current (20-50A) to support the energy requirements of the storage controller for critical operations such as power failure detection confirmation and issuing cache state saving commands. During the 100ms-500ms transition phase, the supercapacitor bank operates in parallel with the LFP battery pack during startup, smoothing voltage fluctuations during LFP battery pack startup. After the LFP battery pack fully takes over, the supercapacitor bank switches to standby mode or continues to provide auxiliary power according to the power distribution controller's instructions. The supercapacitor bank's independent protection capability (not dependent on the LFP battery pack) provides 3-10 seconds of emergency discharge time, sufficient to support the complete overwrite of small caches (≤4GB).

[0071] c. Power distribution controller function

[0072] The power distribution controller is the core execution component of this hybrid BBU architecture, controlled by the local MCU via PWM (Pulse Width Modulation) signals. Its main functions include: power failure detection (monitoring the input bus voltage drop rate dV / dt; when dV / dt > 50V / ms and the bus voltage < 80% of the nominal value, a power failure event is determined, with a response time < 2ms); power supply path switching (controlling the MOSFET switching array to achieve power supply switching and power allocation between the LFP battery pack and the supercapacitor pack); current limiting protection (implementing independent overcurrent protection for each path, limiting the LFP path to 10A and the supercapacitor path to 50A); and status feedback (encoding the current power supply mode, path status, and fault flag bits into a status word, which is reported along with telemetry data).

[0073] d. Data reporting function

[0074] The local MCU packages the collected telemetry data, including LFP battery pack data, supercapacitor pack data, and power distribution controller status words, into PMBus standard format data frames. These frames are then reported to the digital twin engine deployed in the BMC (Baseboard Management Controller) via the SMBus / I2C bus. The data frame format conforms to the PMBus 1.3 protocol specification and includes: READ_VIN (input voltage), READ_VOUT (output voltage), READ_IOUT (output current), READ_TEMPERATURE (temperature), and MFR_SPECIFIC (manufacturer-defined data area used to transmit extended data such as SOC, ESR, and power distribution status). The reporting cycle is 1 second in BBU operating mode (during charging and discharging) and 60 seconds in idle mode (BBU standby float charging state) to balance communication load and real-time requirements.

[0075] In summary, in this embodiment, the ADC timer of the BBU local MCU triggers data acquisition every 1 second, sequentially acquiring analog signals such as the voltage of each string of the LFP battery pack, total current, multi-point temperature, and the voltage, ESR, total current, and temperature of each individual cell of the supercapacitor pack. The MCU runs the local ECM model, performing ampere-hour integral SOC estimation and overvoltage / undervoltage / overtemperature hard protection checks. If a hard protection trigger condition is detected, the MCU immediately cuts off the power path and sends an emergency power-down protection signal to the storage controller via an independent hardware signal line (this path does not rely on BMC communication, and the response time is <10ms). Under normal conditions, the MCU packages the acquired data into PMBus data frames and reports them to the BMC via the SMBus / I2C bus. The time for a single data acquisition and reporting is <5ms.

[0076] S2: Perform data preprocessing on the telemetry data and construct a real-time operating status snapshot of the battery backup unit based on the preprocessed telemetry data; data preprocessing includes: time alignment, outlier removal, temperature compensation, and auxiliary data fusion;

[0077] Specifically, timing alignment: Since there may be microsecond-level time deviations in the data acquisition of LFP battery packs and supercapacitor packs, this embodiment uses the BMC system clock as a reference, timestamps the received data frames, and uses linear interpolation to align different data sources to the same time (accuracy ±1ms).

[0078] Outlier removal: A dynamic threshold detection algorithm based on the 3σ criterion (Raida criterion) is employed. For each sensor's continuous historical data, the mean μ and standard deviation σ of a sliding window (window length 60 sampling points) are calculated. When a new sampled value exceeds the range [μ-3σ, μ+3σ], it is marked as an outlier and replaced with the previous valid value. For abrupt changes in current and voltage, an additional rate-of-change constraint test (|Δx / Δt|>threshold) is used to identify physically impossible values.

[0079] Temperature Compensation: The OCV-SOC relationship of LFP batteries is temperature-dependent. In this example, based on real-time temperature data, a lookup table interpolation method is used to obtain the reference OCV value at the current temperature from the OCV-SOC-T three-dimensional mapping table. The temperature compensation table is discretized in 5℃ increments within the range of -20℃ to +60℃, and the SOC dimension is in 1% increments.

[0080] Auxiliary data fusion: The module fuses host write load information from the storage controller (including average write bandwidth, write IOPS, write cache utilization, etc. for the most recent 1 minute / 5 minutes / 15 minutes), as well as ambient temperature sensor data from the BMC. This data serves as an important input for subsequent protection time calculations.

[0081] In summary, this step, upon receiving the PMBus data frame, performs timing alignment (timestamping based on the BMC system clock), outlier removal (3σ criterion verification), temperature compensation (OCV-SOC-T lookup interpolation), and auxiliary data fusion (merging write load information and ambient temperature data reported by the storage controller). After processing, a real-time state snapshot of the BBU is constructed and written to the shared memory circular buffer. This step takes 3-8 ms.

[0082] S3: Using dual extended Kalman filtering and real-time operating status snapshots, the health status parameters of the battery backup unit are identified online, and health status time series data is generated.

[0083] Specifically, the Dual Extended Kalman Filter (Dual EKF) operates in parallel on two nested timescales, enabling fast tracking of the State of Charge (SOC) and online identification of SOH-related parameters (maximum available capacity Q_max, ohmic internal resistance R0). The Dual Extended Kalman Filter includes:

[0084] Fast timescale state of charge estimation (EKF-SOC): The sampling period is 1 s (synchronized with the telemetry data reporting period). The state vector is x_soc = [SOC; U1], where SOC is the state of charge and U1 is the polarization voltage of the RC network in the ECM. The state transition equation is constructed based on the ampere-hour integral and the ECM discharge equation. The observation equation uses the terminal voltage V_terminal as the observation. The SOC estimation is updated once every time telemetry data arrives, with a convergence time of <30 s (from initial guess to error <5%) and a steady-state estimation accuracy of ≤3% (under normal temperature of 25℃).

[0085] Slow-scale health status parameter identification (EKF-SOH): The sampling period is N×1s (N is the slow-scale coefficient, typically N=60, meaning SOH parameters are updated every 60s). The state vector x_soh=[Q_max; R0], where Q_max is the current maximum available capacity (Ah, initial value is the nominal capacity), and R0 is the ohmic internal resistance (initial value is the factory nominal value, in mΩ). Parameter identification uses the fast-scale SOC estimation results as known inputs. By comparing the residuals between the model-predicted terminal voltage and the actual measured terminal voltage, the estimated values ​​of Q_max and R0 are adaptively corrected using Kalman gain. The convergence time for Q_max identification is approximately 5~10 complete charge-discharge cycles, and the convergence time for R0 identification is approximately 1~2 cycles. SOH is calculated as SOH=Q_max_current / Q_max_nominal×100%, with an online identification accuracy ≤5%.

[0086] The collaborative working mechanism of Dual EKF is as follows: EKF-SOC performs a state update every 1 second, outputting the current SOC and predicted terminal voltage values; after accumulating N SOC estimation results, EKF-SOH triggers a parameter update, using the joint information of N SOC-voltage pairs to correct Q_max and R0; the updated Q_max and R0 are immediately fed back into the model parameters of EKF-SOC for SOC estimation in the next cycle, forming a closed-loop adaptive mechanism. This coupling structure ensures that the SOC estimation is always based on the latest battery characteristic parameters, while the SOH parameter identification fully utilizes the timing information of the SOC estimation.

[0087] In this embodiment, the fast-timescale state of charge estimation and the slow-timescale health state parameter identification are coupled and corrected in a closed loop. The fast-timescale state of charge estimation is updated once in each telemetry data reporting cycle, and the slow-timescale health state parameter identification is updated once after accumulating multiple reporting cycles, and the updated health state parameters are fed back into the model parameters of the fast-timescale state of charge estimation.

[0088] In summary, this step involves reading the latest state snapshot and performing Dual EKF joint estimation: the fast-scale EKF-SOC updates the state at 1-second intervals, outputting the current SOC estimate and terminal voltage prediction; every 60 seconds, the slow-scale EKF-SOH triggers parameter identification, updates the Q_max and R0 estimates, and calculates:

[0089] SOH=Q_max_current / Q_max_nominal×100%.

[0090] The output of Dual EKF is simultaneously updated to the input buffer of the LSTM model. This step takes 3~8ms (single EKF-SOC) + 5~15ms (EKF-SOH, once every 60s).

[0091] S4: Based on health status time series data, use a time series prediction model to predict the health status degradation trend of the battery backup unit;

[0092] Specifically, this step receives the SOH time series data output from the previous step (typically one SOH sampling point per day, obtained by averaging and filtering multiple Q_max identification results for that day) and performs medium- to long-term decay trend prediction. A dual-mode prediction strategy is used: short-term prediction (30-day rolling window) uses exponential smoothing (Holt two-parameter exponential smoothing, horizontal component α=0.3, trend component β=0.1); long-term prediction (365-day trend) uses a lightweight time series prediction model.

[0093] The time series prediction model employs a lightweight Long Short-Term Memory (LSTM) network model. Specific lightweighting measures for the LSTM model include: using a single-layer LSTM structure, limiting the hidden layer dimension to 32 or 64, and a time step of 30 days (corresponding to 30 SOH sampling points); limiting the number of model parameters to ≤1M (1 million parameters), where the input-to-hidden layer weight matrix dimension is (input feature dimension 5 × hidden dimension 64 = 320 parameters), the hidden-to-hidden layer recurrent weight matrix dimension is (64 × 64 = 4096 parameters), the bias vector dimension is 64, and the output layer fully connected dimension is (64 × 1 = 64 parameters), for a total of approximately 4544 floating-point parameters, far less than the 1M limit; and using an INT8 (8-bit integer) quantization strategy to map floating-point weights to [-128, ...

[127] The model size is compressed to about 4.5KB (about 18KB without quantization) within the integer range, and the inference speed is improved by 2 to 3 times. The optimal scaling factor is determined by minimizing the KL divergence based on the historical SOH dataset.

[0094] This embodiment accelerates LSTM forward propagation computation based on the ARM NEON SIMD (Single Instruction Multiple Data) instruction set. Typical inference latency is as follows: a single LSTM forward propagation (30 time steps) takes 5-15ms; a single Dual EKF iteration (including SOC and SOH dual filter updates) takes 3-8ms; and a single protection time calculation takes 2-5ms. The total time for a single complete inference cycle of the digital twin engine (from receiving telemetry data to outputting RPT) is 10-50ms, meeting a total latency budget of <100ms.

[0095] In this step, the complete collaborative data flow between Dual EKF and LSTM is as follows:

[0096] Step 401: EKF-SOH outputs a set of SOH-related parameter estimates (Q_max_k, R0_k) at each slow-scale period (60s), and these parameters are written to the shared memory historical parameter queue along with timestamps.

[0097] Step 402: At 0:00 every day (or the daily settlement time set by the system), the parameter smoothing submodule performs outlier removal and moving average filtering on the Q_max and R0 estimated sequences of the past 24 hours (window length 24, i.e. the parameter mean of 24 hours) to generate the daily SOH feature value SOH_day=Q_max_daily_avg / Q_max_nominal×100%.

[0098] Step 403: The daily SOH feature values ​​are fed into the input sequence buffer of the LSTM model. Every 30 daily SOH values ​​(i.e., 30 days of historical data) accumulated by the LSTM model, a forward propagation prediction is performed, outputting the SOH prediction curve for the next 365 days. When there are fewer than 30 days of historical data, the model reverts to exponential smoothing for short-term prediction.

[0099] Step 404: The consistency of the SOH curve predicted by LSTM is checked with the latest SOH value currently identified by EKF-SOH in real time. If the deviation between the two exceeds 10%, the model calibration process is triggered: the input sequence of the LSTM model is drift corrected based on the real-time result of EKF-SOH, and the prediction is re-executed.

[0100] Step 405: Transfer the SOH trend curve to the next step to deduce the protection capability of the BBU at any future time. Simultaneously, the SOH trend data is used for RUL prediction: when the predicted SOH curve intersects with a preset replacement threshold (e.g., SOH = 60%), the date corresponding to the intersection point is the predicted RUL value.

[0101] In this example, when using a time-series prediction model to predict the health status degradation trend of a battery backup unit, the daily feature values ​​in the health status time series data are used as input. The time-series prediction model performs forward propagation prediction and outputs the health status degradation trend for the next preset number of days. When the deviation between the predicted value of the time-series prediction model and the health status parameters identified in real time by the dual extended Kalman filter exceeds a preset threshold, a calibration process is triggered. The real-time identification result of the dual extended Kalman filter is used as the standard to perform drift correction on the input sequence of the time-series prediction model and re-execute the prediction.

[0102] The core advantages of this collaborative mechanism are: Dual EKF provides real-time, high-frequency SOH parameter identification capabilities, reflecting the current actual state of the battery; LSTM provides long-term trend extrapolation capabilities based on historical degradation patterns, capturing the nonlinear characteristics of battery aging; and the closed-loop verification mechanism of both ensures the reliability and consistency of the prediction results.

[0103] In summary, after obtaining the latest daily SOH feature value in this step, if it is the daily settlement time (0 o'clock) and the LSTM model input buffer is full of 30 days of data, then LSTM forward propagation prediction is performed to output the SOH prediction curve for the next 365 days; otherwise, the exponential smoothing method is used to output a 30-day rolling prediction.

[0104] S5: Calculate and predict the remaining protection time based on the current effective energy of the battery energy storage unit and the capacitor energy storage unit, the current load information of the storage system and the health status degradation trend;

[0105] Specifically, this step is responsible for converting the output of the previous step (SOC, SOH, ESR, etc.) into a "Predicted Remaining Protection Time (RPT)" metric that can be directly used in engineering. RPT is defined as: under the current BBU energy storage status, cache dirty data volume, and host write load conditions, if a power failure occurs, the minimum margin required for the BBU to safely flush all dirty data in the DRAM cache to non-volatile memory.

[0106] The formula for calculating RPT is as follows:

[0107] First, calculate the effective energy E_batt (in Wh) that the LFP battery pack can release under the current SOC and SOH:

[0108] E_batt = V_avg × Q_max_current × SOC_current × η_discharge,

[0109] Where V_avg is the average discharge voltage of the LFP battery pack (typically 3.2V × number of series), Q_max_current is the current maximum available capacity (Ah, identified by Dual EKF), SOC_current is the current SOC (%), and η_discharge is the discharge efficiency (taken as 0.90~0.95).

[0110] Secondly, calculate the effective energy E_scap (in Wh) that the supercapacitor bank can release:

[0111] E_scap = 0.5 × C_current × ( ) × η_scap,

[0112] Where C_current is the current capacitance (F), V_scap is the current supercapacitor bank voltage (V), V_scap_min is the minimum operating voltage (typically 50% of the rated voltage), and η_scap is the discharge efficiency (taken as 0.95).

[0113] Then, the total effective energy of the combined BBU:

[0114] E_total = E_batt + E_scap.

[0115] Next, calculate the total energy consumption during the power-down brushing process:

[0116] E_drain = P_cache_flush × t_flush + P_controller × t_total + P_ssd× t_flush;

[0117] A simplified model is used to model t_total in the total energy consumption formula for the power-down brush process:

[0118] t_total = t_overhead + t_flush,

[0119] Where t_overhead is the fixed overhead time (typically 30~60s), t_flush=D_dirty / W_ssd is the actual cache flushing time, and t_total represents the total time from the occurrence of power failure to the complete flushing of cache data.

[0120] Dirty data volume: D_dirty = DRAM_cache_size × write_ratio × dirty_ratio.

[0121] Therefore, the formula for calculating the Remaining Protection Time (RPT) is:

[0122] RPT = (E_total - E_drain_min) / P_avg_standby.

[0123] A more conservative, simplified engineering formula is:

[0124] RPT = E_total / P_total_discharge - t_total,

[0125] Wherein, P_total_discharge is the total power consumption during the power-down discharge period;

[0126] P_total_discharge=P_controller+P_ssd×duty_cycle+P_peripheral,

[0127] Where duty_cycle is the SSD write duty cycle (0.7~0.9).

[0128] Taking a typical enterprise-level storage system as an example (16GB DRAM cache, write cache ratio 50%, average dirty data rate 40%, SSD write speed 1GB / s, controller power consumption 50W, SSD array power consumption 100W), t_flush = (16 × 0.5 × 0.4) / 1 = 3.2s. Adding the fixed overhead of 60s, t_total ≈ 63.2s, and the total discharge energy consumption E_drain = (50 + 100) × 63.2 / 3600 ≈ 2.63Wh. If the total effective energy of the BBU E_total = 72Wh (nominal 80Wh × SOH 90%), then the RPT corresponds to 72 / 2.63 ≈ 27.4 power-down events. Converting RPT to equivalent continuous discharge time:

[0129] RPT_hours=E_total / P_total_discharge=72 / 150=0.48h≈28.8min, which is the equivalent protection time margin of the current BBU status.

[0130] In summary, this step calculates the predicted remaining protection time (RPT) based on the current E_total (LFP + supercapacitor effective energy), the D_dirty reported by the storage controller, and the write load, according to the RPT calculation formula. This step takes 5~15ms (LSTM mode) or 2~5ms (exponential smoothing mode).

[0131] S6: Send the predicted remaining protection time to the hierarchical policy engine; the hierarchical policy engine dynamically adjusts the cache write policy based on the predicted remaining protection time; the dynamic adjustment of the cache write policy includes: write-back, write protection and direct write three-level hierarchical protection policy; the hierarchical policy engine is deployed in the I / O scheduling software stack of the storage controller.

[0132] Specifically, this step performs a three-tiered protection strategy switch based on the Predicted Remaining Protection Time (RPT) output from the previous step. The strategy engine uses RPT as the sole decision metric and does not directly use SOH as the strategy switching condition (SOH is only reported to the operation and maintenance platform as auxiliary reference information).

[0133] When the hierarchical policy engine dynamically adjusts the cache write policy based on the predicted remaining protection time, it includes: reading the current policy state, which includes: Write-Back, Write-Guard, and Write-Through. If the current policy state is Write-Back and the predicted remaining protection time is less than or equal to 72 hours, the decision is to switch to Write-Guard mode; if the current policy state is Write-Guard and the predicted remaining protection time is greater than or equal to 80 hours, the decision is to switch to Write-Back mode; if the current policy state is Write-Guard and the predicted remaining protection time is less than or equal to 24 hours, the decision is to switch to Write-Through mode; if the current policy state is Write-Through and the predicted remaining protection time is greater than or equal to 30 hours, the decision is to switch to Write-Guard mode; if a minimum residency time check is triggered, the decision is to maintain the current state; otherwise, the decision is to maintain the current state.

[0134] a. Basis and calculation formula for setting strategy thresholds

[0135] The switching thresholds for the three-tier strategy (72h and 24h) are derived based on engineering parameters of a typical enterprise-level storage system. The core principle for threshold setting is that the RPT threshold should ensure that, even in the worst-case scenario (when the actual BBU protection time is lower than the predicted value due to aging or sudden failure), the storage system still has sufficient protection time to complete data flushing. Introducing a safety factor γ (SafetyFactor, ranging from 1.5 to 2.0), the strategy degradation trigger condition is RPT_predicted ≤ γ × t_required.

[0136] The derivation of the 72h write-back to write-guard switching threshold: Assuming a typical medium-sized storage system configuration (16GB DRAM cache, 8GB write cache, average dirty data rate of 50% (4GB dirty data), SSD write speed of 1GB / s, total power-down flush time of approximately 64s, total discharge power of 150W, and energy consumption per power-down of approximately 2.67Wh), if the BBU's nominal energy is 72Wh, it can support approximately 27 power-downs at full capacity. When the RPT drops to 72h, the equivalent number of power-downs it can support is 72 × 3600 / 64 ≈ 4050. The practical engineering significance of this threshold is that the remaining effective energy of the BBU is approximately 30%~40% of its nominal value, showing obvious aging but still capable of supporting multiple power-down protection cycles. The 72h threshold, as the first-level degradation threshold, is selected by comprehensively considering the response time of maintenance personnel (enterprise IT can usually intervene within 4 to 8 hours during working hours), the error boundary of RPT prediction error ≤ 10% (72h ± 7.2h), and the sufficient operational margin to reserve for subsequent Write-Guard mode.

[0137] The derivation of the 24-hour Write-Guard to Write-Through switching threshold: When the RPT drops to 24 hours, the BBU is in a severely aged state, with remaining effective energy possibly only 10%~20% of the nominal value. Continuing to operate in Write-Guard mode at this point is too risky, as a prediction error of 10% (±2.4 hours) is close to the order of magnitude of the time required for a single complete refresh. The engineering implication of the 24-hour threshold is that the emergency response window for the operations team has been shortened to less than one business day, and it is imperative to immediately switch to the most conservative Write-Through mode until the BBU is replaced.

[0138] General formula for threshold calculation:

[0139] Threshold=k×(D_cache_max×dirty_ratio_max) / W_ssd_min×P_total / E_batt_nominal× SOH_threshold,

[0140] Where k is the safety factor (1.5~2.0), D_cache_max is the maximum cache capacity (GB), dirty_ratio_max is the maximum dirty data rate (0.8), W_ssd_min is the worst-case SSD write rate (GB / s, 0.5GB / s), P_total is the total discharge power (W), E_batt_nominal is the nominal energy of the BBU (Wh), and SOH_threshold is the SOH level corresponding to the protection threshold.

[0141] Substituting typical parameters (k=1.8, D_cache_max=32GB, dirty_ratio_max=0.8, W_ssd_min=0.5GB / s, P_total=150W, E_batt_nominal=80Wh, SOH_threshold=0.35), the calculated Threshold_first is approximately 77h, which matches the 72h selected in the project.

[0142] The second-level threshold is calculated similarly (k=1.5, SOH_threshold=0.15), yielding Threshold_second≈25h, which matches 24h.

[0143] b. Hysteresis mechanism design

[0144] To avoid frequent policy switching (jitter) caused by RPT fluctuations near the policy threshold boundary, the hierarchical policy engine introduces a hysteresis control mechanism.

[0145] The trigger condition for downgrading from Write-Back to Write-Guard is: RPT ≤ 72h (no hysteresis; downgrading occurs immediately upon triggering to ensure safety). The trigger condition for upgrading from Write-Guard back to Write-Back is: RPT ≥ 80h (hysteresis width of 8h, meaning the upgrade can only proceed after the RPT rises above the 72h threshold for at least 8 hours). The 80h threshold is selected as follows: 110% of the 72h threshold plus the prediction error margin (72 × 10% = 7.2h), rounded down to 80h.

[0146] The trigger condition for downgrading from Write-Guard to Write-Through is: RPT ≤ 24h (no hysteresis, downgrading occurs immediately upon triggering). The trigger condition for upgrading from Write-Through back to Write-Guard is: RPT ≥ 30h (hysteresis width of 6h, i.e., 125% of 24h).

[0147] The state machine transition logic of the hysteresis mechanism: The policy engine maintains a state variable strategy_state∈{WB,WG, WT}. In each RPT update cycle (1s), the state is updated according to the following rules: If the current state=WB and RPT≤72h → switch to WG, and record the reason and time of the switch; if the current state=WG and RPT≥80h → switch to WB; if the current state=WG and RPT≤24h → switch to WT; if the current state=WT and RPT≥30h → switch to WG; otherwise, the current state remains unchanged.

[0148] In addition, the policy engine also implements a minimum dwell time mechanism: after any policy switch is completed, the new policy continues to be executed for at least 300 seconds (5 minutes). During this period, even if the RPT crosses the threshold again, a reverse switch will not be executed, further suppressing jitter.

[0149] c. Write-Back full performance mode (Level 1, RPT>72h)

[0150] When the predicted remaining protection time (RPT) exceeds 72 hours, the BBU is in a healthy state, and the storage system operates at full performance in Write-Back mode. In this mode, write I / O requests directly enter the DRAM write cache and immediately return an acknowledgment (ACK) to the host. The background process flushes dirty data to the back-end storage at an adaptive frequency. The write cache refresh cycle is the default configuration (typically a background flush every 5-30 seconds, dynamically adjusted by the storage system based on write load). The I / O current limiter is fully open, and no additional bandwidth restrictions are imposed. The supercapacitor bank is in online standby mode, and the power distribution controller is configured in "LFP main supply + supercapacitor instantaneous response" mode.

[0151] d. Write-Guard Enhanced Protection Mode (Level 2, 24h) <RPT≤72h)

[0152] Write-Guard mode is one of the core technical components of this solution. When the BBU shows obvious signs of aging but has not yet reached a dangerous level, it extends the power-loss protection time by actively limiting write performance, thus achieving a smooth transition of "trading performance for safety".

[0153] The Write-Guard pattern includes the following three key technical mechanisms:

[0154] I / O Rate Limiting Mechanism: A token bucket algorithm is used to dynamically limit the front-end write I / O bandwidth. The core parameters of the token bucket are configured as follows: The token generation rate R_token is dynamically calculated based on the RPT value, R_token = R_max × min(1.0, RPT / 72h), where R_max is the peak write bandwidth in Write-Back mode (typically 500MB / s). When RPT = 72h, R_token = R_max (unlimited); when RPT = 36h, R_token = 0.5 × R_max (50% rate limiting); when RPT → 24h, R_token → 0.33 × R_max. The token bucket capacity B_bucket is set to 4 × R_token (i.e., allowing a maximum of 4 seconds of burst writes) to balance bandwidth smoothness and burst performance. Token bucket overflow handling: When a host write request arrives, if there are enough tokens in the bucket, it is processed normally and the corresponding token amount is deducted; if there are insufficient tokens, the I / O request enters a waiting queue with a maximum depth of 1024 requests. After a timeout (500ms), a host retry signal (EAGAIN) is returned. Rate limiting granularity: Token bucket rate limiting is implemented at the LUN (Logical Unit Number) level of the storage controller. Each LUN maintains an independent token bucket instance to ensure that a burst of writes to a single LUN does not affect the bandwidth allocation of other LUNs.

[0155] Disk flushing control logic: In Write-Guard mode, the background flushing frequency of the write cache is actively increased to maintain a low amount of dirty data. The disk flushing cycle is shortened to:

[0156] T_flush_guard=max(T_flush_min, T_flush_default×(RPT / 72h)), where T_flush_min is the minimum flush cycle (taken as 2s).

[0157] When RPT=72h, T_flush_guard=T_flush_default; when RPT=36h, T_flush_guard≈0.5×T_flush_default (refresh frequency doubles); when RPT→24h, T_flush_guard→T_flush_min (maximum refresh frequency).

[0158] Hard limit on dirty data volume: In Write-Guard mode, an additional dirty data volume threshold D_guard = DRAM_cache_size × 0.15 is set (i.e., the dirty data volume does not exceed 15% of the total cache capacity). When the dirty data volume reaches this limit, a forced flush is triggered immediately, regardless of whether the flush cycle has been reached, and backpressure is applied to new write requests at the same time. Gradual flush strategy: After switching from Write-Back to Write-Guard, the flush cycle is linearly reduced from T_flush_default to the target value T_flush_guard within 10 minutes, rather than abruptly changing.

[0159] Dynamic calculation method for power outage protection time: In Write-Guard mode, the policy engine continuously calls the protection time calculation interface of the digital twin engine to update the RPT value in real time. As the RPT decreases from 72h to 24h, the I / O current limiting rate and disk flushing frequency are automatically and smoothly adjusted without manual intervention. The policy engine recalculates the RPT every 1 second and refreshes the token bucket rate and disk flushing cycle based on the current RPT value to ensure a real-time balance between protection performance and business performance.

[0160] Supercapacitor bank coordination mechanism: In Write-Guard mode, the supercapacitor bank remains online in standby mode, and the power distribution controller configuration remains unchanged. However, due to the reduced write load in Write-Guard mode, the amount of dirty data in DRAM is actively controlled, and the actual amount of data flushed during power loss is reduced. This means that under the same BBU energy state, the actual protection margin of Write-Guard mode is higher than that of Write-Back mode—this is precisely the engineering significance of the hierarchical protection strategy.

[0161] e. Write-Through security degradation mode (Level 3, RPT≤24h)

[0162] When the predicted remaining protection time (RPT) is ≤24h, the storage controller switches the cache write policy to Write-Through mode. In this mode, all write I / O requests are simultaneously written to the DRAM cache and the backend non-volatile storage (SSD), and an ACK is returned to the host only after both have been written. Write-Through mode eliminates the risk of data loss due to power failure—because the data is persisted to the SSD while being written to DRAM, even if the BBU completely fails, there will be no data loss. The cost of this mode is a significant increase in write latency (an additional SSD write latency, typically 0.5~2ms), and write throughput is limited by the write capacity of the SSD (typically decreasing by 30%~50%).

[0163] Upon entering Write-Through mode, the policy engine performs the following additional operations: immediately sends a Critical-level alarm to the operations platform via SNMP Trap and Redfish Event, with the alarm content including the current RPT value, BBU SOH value, and recommended replacement time window; the BBU local MCU switches to the highest frequency telemetry mode (all sensor sampling rates are increased to the maximum value, and the data reporting cycle is shortened to 200ms); the digital twin engine switches to high-frequency prediction mode (RPT calculation cycle is shortened from 1s to 200ms) to ensure that any subtle state changes are captured in a timely manner.

[0164] In summary, in this step, the digital twin engine pushes the RPT value to the hierarchical policy engine of the storage controller via the IPMB / IPC channel. The hysteresis decision module of the policy engine executes the following logic: read the current policy state strategy_state∈{WB, WG, WT}; if state=WB and RPT≤72h → decision result=switch to WG; if state=WG and RPT≥80h → decision result=switch to WB; if state=WG and RPT≤24h → decision result=switch to WT; if state=WT and RPT≥30h → decision result=switch to WG; if the minimum dwell time check is triggered (within 300s) → decision result=maintain the current state; otherwise → decision result=maintain the current state. The policy decision module simultaneously calculates the token bucket rate R_token and the flush cycle T_flush_guard target value in Write-Guard mode. This step takes <5ms.

[0165] The policy decision result is distributed to the cache management layer and I / O scheduling layer via the storage controller's internal interface. If the decision result is a state switch, then: the cache management layer executes the switch operation according to the target policy—when switching to Write-Guard, a token bucket instance is created and a progressive disk flushing frequency adjustment timer is started; when switching to Write-Through, the write-through path is immediately enabled and the Write-Back path is paused; when switching to Write-Back, the token bucket instance is destroyed and the default flushing cycle is restored. The I / O scheduling layer enables / disables / updates the parameters of the I / O rate limiter. The policy engine records the policy switch event (timestamp, state before and after the switch, RPT value, switch reason) to the local log and pushes it to the operation and maintenance platform via BMC. This step takes: 20ms for the policy to be distributed to the controller + 10ms for the policy to take effect.

[0166] This embodiment provides a method for predicting the BBU protection capability of a storage system based on digital twins, which further includes:

[0167] The system sends the predicted remaining protection time, health status degradation trend, and health status parameters of the battery energy storage unit and capacitor energy storage unit to the operation and maintenance platform. The operation and maintenance platform generates and displays the health status trend of the battery backup unit, the prediction of the remaining service life, and the replacement suggestion window. When the predicted remaining protection time is lower than the preset warning threshold, the system pushes an active warning message to the operation and maintenance platform.

[0168] Specifically, the operations and maintenance platform is deployed on the storage management platform side, and receives predictive maintenance data output by the digital twin engine through standardized interfaces to provide decision support for operations and maintenance personnel.

[0169] The operation and maintenance platform implements the following functions:

[0170] a. Data reception and visualization

[0171] The operations and maintenance platform receives data from the BMC via SNMP v3 Trap, Redfish standard Event Schema, or HTTPS RESTful API. The data includes: real-time BBU SOC / SOH / RPT values ​​(push cycle 1 second); SOH trend data (30-day rolling curve, daily SOH sampling points); RUL prediction report (365-day long-term trend, predicted SOH curve and the date of intersection with the 60% replacement threshold); predicted remaining protection time trend chart (72-hour short-term prediction curve); policy switching event logs (recording the time, reason, and RPT value of each Write-Back / Write-Guard / Write-Through mode switch); and abnormal event alarms (abnormal ESR increase, sudden drop in Q_max, abnormal temperature, etc.).

[0172] The operations and maintenance platform provides a Web GUI dashboard to visually display the above data in chart form. The dashboard layout includes: a BBU health status dashboard (green / yellow / red corresponding to RPT>72h / 24~72h / <24h respectively); a SOH trend line chart (overlaying 30-day historical data and 30-day forecast); a RUL countdown timer (displaying the remaining time before the recommended replacement in the form of "XX days remaining"); an RPT trend chart (a 72-hour rolling curve, marked with 72h and 24h threshold lines); and an alarm event list (displaying all policy switching and abnormal events in reverse chronological order).

[0173] b. Change the window suggestion algorithm

[0174] The operations and maintenance platform provides a function to suggest optimal replacement windows based on RUL and operational constraints. The calculation of the replacement window takes into account the following factors: RUL prediction value (remaining lifetime days output by the digital twin engine); low-peak business periods (identifying the lowest write load periods weekly / daily based on historical I / O load statistics); maintenance window constraints (allowed downtime maintenance periods stipulated by the enterprise's IT operations and maintenance policies); BBU inventory and logistics cycle (estimated time from order placement to delivery); safety margin (a 14-day safety margin reserved on top of the RUL).

[0175] The recommended output format for the replacement window is as follows: Recommended start date for replacement = RUL_end - 14 days (safety margin); Recommended execution period = the top 3 off-peak business periods (e.g., every Sunday from 2:00 AM to 6:00 AM); Replacement priority = calculated based on the current RPT value (RPT < 48h is urgent, 48~72h is important, > 72h is planned); Estimated replacement operation time (based on BBU hot-swap design, typically 15~30 minutes).

[0176] c. Proactive early warning and push mechanism

[0177] The digital twin engine proactively pushes alerts to the operations and maintenance platform through the following channels: SNMP Trap (using the standard SNMPv3 format, supporting confirmation delivery in inform / request mode); Redfish Event (following the DMTF Redfish standard Event Schema, supporting long-connection push of SSE / Server-Sent Events); and HTTPS RESTful Callback (the operations and maintenance platform registers a Webhook URL, and the digital twin engine initiates an HTTPS POST push when it detects a critical event).

[0178] The warning levels are as follows: Info level - Policy switching event (normal switch from Write-Back to Write-Guard); Warning level - RPT enters the 24~72h range, SOH drops below 70%, and the supercapacitor ESR increases to 150% of its initial value; Critical level - RPT drops below 24h, SOH drops below 60%, BBU communication is interrupted for more than 5 minutes, and a hard fault of overvoltage / undervoltage / overtemperature in a single battery cell is detected.

[0179] It should be noted that the digital twin engine pushes information such as RPT, SOC, SOH, and policy status for this cycle to the operations platform via SNMP Trap / Redfish Event. The operations platform then updates the dashboard data and alarm status. This step is an asynchronous operation and does not consume policy switching latency budget.

[0180] After the process is completed, wait for the next 1-second timer cycle to trigger, and repeat S1~S6. Total latency budget: S1 acquisition ~5ms + S2 fusion ~8ms + S3 estimation ~8ms + S4 prediction ~15ms + S5 decision ~5ms + S6 execution ~30ms = worst case total latency ~71ms << 100ms threshold.

[0181] This embodiment provides a closed-loop method for predicting the protection capability of a storage system's BBU based on digital twins, encompassing the entire process from sensor data acquisition, state estimation, trend prediction, strategy decision-making to strategy execution. This process executes cyclically with a 1-second cycle, with a total latency controlled within 100ms.

[0182] In this embodiment, a method for predicting the protection capability of a storage system BBU based on digital twins is provided. The update strategy of the digital twin engine adopts a multi-trigger mode, rather than a simple "update after each charge / discharge event," to ensure that the model is neither updated too frequently (consuming BMC computing resources) nor delayed in abnormal states. The multi-trigger mode includes: timed trigger update, event-triggered update, and abnormal trigger update.

[0183] Scheduled update trigger: A complete model refresh is performed every 24 hours, regardless of whether a charge / discharge event occurs during this period. This trigger ensures that the model remains active even during long-term inactivity, preventing parameter drift accumulation. Scheduled updates include operations such as full Dual EKF covariance matrix reset, LSTM input sequence integrity verification, and ESR baseline recalibration.

[0184] Event-triggered updates: These updates are triggered immediately after each BBU charge / discharge event (a complete charge → discharge or discharge → charge cycle). This trigger ensures the model promptly captures state changes resulting from the charge / discharge process. The update cycle is linked to the actual usage frequency of the BBU (typical enterprise-level storage systems undergo 1-4 discharge tests per month). Event-triggered updates focus on rapid convergence of SOC / SOH parameters.

[0185] Anomaly-triggered updates: Dual EKF immediately triggers an additional update when it detects the following anomalous patterns: Q_max decreases by more than 5% within two consecutive slow-scale cycles (indicating possible burst capacity decay); R0 increases by more than 20% within 24 hours (indicating possible internal micro-short circuits or loose connections); supercapacitor ESR increases by more than 10% per week; SOC estimation residuals continuously increase beyond a threshold (>100mV for 10 consecutive sampling points). During anomaly-triggered updates, Dual EKF automatically increases the gain of the process noise covariance Q_matrix to accelerate parameter convergence and track abrupt changes.

[0186] The specific process for updating a digital twin model is as follows:

[0187] The digital twin model update process drives the continuous evolution and calibration of the core models (Dual EKF and LSTM) of the digital twin engine, ensuring that model parameters are synchronized with the actual physical state of the BBU. This process is executed asynchronously using a multi-trigger mode, running in parallel with the main strategy switching process, and includes the following steps:

[0188] Step U100: Data Reporting and Reception. The BBU local MCU reports telemetry data frames to the BMC via PMBus / SMBus according to a cycle of 1 second in running mode and 60 seconds in idle mode. The communication thread of the digital twin engine listens for I2C bus interrupts and wakes up the parsing thread when data arrives. After the data frame is checked by CRC (Cyclic Redundancy Check), it is written to the original data receive buffer. If three consecutive CRC checks fail or no data is received for more than 5 seconds, the communication thread marks the BBU communication as abnormal and triggers a Critical alarm.

[0189] Step U200: Multi-trigger condition detection. The multi-trigger update scheduler continuously monitors the following three types of trigger conditions: timed trigger (checking if the 24-hour timed point has been reached); event trigger (checking if the BBU has completed a full charge / discharge event); and anomaly trigger (checking if the Dual EKF output meets the following anomaly conditions: Q_max decreases by more than 5% in adjacent cycles, R0 increases by more than 20% over 24 hours, ESR increases by more than 10% weekly, and SOC residuals exceed 100mV for 10 consecutive periods). If multiple trigger conditions are met simultaneously, the scheduler executes them in order of priority (anomaly > event > timed).

[0190] Step U300: Data Fusion and Feature Extraction. Once the triggering conditions are met, the scheduler initiates the data fusion module to process the most recent historical data in batches. The processing includes: time-series resampling (standardized to 1-second intervals), missing value interpolation (linear interpolation), outlier removal (3σ criterion), and feature extraction (extracting features such as charge / discharge capacity, average discharge rate, temperature integral, and voltage decay slope). The fusion results are used to construct the training samples or calibration inputs required for model updates.

[0191] Step U400: Dual EKF parameter calibration. The SOC / SOH joint estimation module performs Dual EKF parameter calibration based on the fused data: For timed updates, a full EKF reset is performed—the covariance matrix P is restored to its initial value, and the full convergence process of the SOC and SOH parameters is re-executed; for event-triggered events, a complete Q_max and R0 identification is performed using the complete voltage-current-time series of this charge-discharge cycle; for abnormal triggers, the process noise covariance Q_matrix is ​​increased (temporarily amplified by 2~5 times) to accelerate parameter convergence and track possible sudden state changes. After the EKF parameter calibration is completed, the updated Q_max and R0 are immediately used to update the model parameters for SOC estimation and written to persistent storage (eMMC) to prevent loss during BMC restart.

[0192] Step U500: LSTM Model Update and Prediction Calibration. At the daily settlement time (0:00), the trend prediction module feeds the daily SOH feature value sequence of the past 30 days into the LSTM model for forward propagation prediction. Simultaneously, model calibration is performed: the residual is calculated by comparing the LSTM's predicted value for day 1 with the actual SOH value identified by EKF-SOH in real-time today. If the absolute value of the residual is >10%, drift correction of the LSTM input sequence is triggered. If the absolute value of the residual is >10% for 7 consecutive days, LSTM model fine-tuning is triggered—using the SOH data from the most recent 90 days as the fine-tuning training set, performing 10-20 epochs of lightweight gradient descent training (learning rate 0.001, freezing the underlying LSTM weights, training only the output layer and the last two layers). The fine-tuning process is executed in the BMC background as a low-priority thread, without affecting the main inference flow, and takes approximately 5-10 minutes. After the LSTM prediction is completed, the SOH prediction curve and RUL value (the intersection date of the curve and the 60% replacement threshold) for the next 365 days are generated.

[0193] Step U600: Result Persistence and Feedback. After the model update is complete, the model management module performs the following persistence operations: Writes the current state vector, covariance matrix, and model parameters of the Dual EKF to the BBU state file of the eMMC (atomic write, first writes to a temporary file and then renames it to ensure file integrity); writes the LSTM model weights (INT8 quantization format) and prediction result sequence to the model persistent storage area; records the model update log (trigger type, update time, key parameter changes, comparison before and after calibration). The model management module maintains the model version history (retaining the most recent 30 versions) and supports one-click rollback to the previous stable version when a model anomaly is detected. The updated RPT and SOH prediction results are fed back to the hierarchical policy engine via the IPC channel and simultaneously pushed to the operation and maintenance platform via SNMP / Redfish.

[0194] The following is a summary table of key technical parameters in this embodiment.

[0195] Table 1 Key Estimation Accuracy Parameters

[0196]

[0197] Table 2 System Response and Delay Parameters

[0198]

[0199] Table 3 Threshold parameters for hierarchical strategy

[0200]

[0201] Table 4 Specifications of Lightweight Digital Twin Model

[0202]

[0203] Table 5 Model Update Strategy Parameters

[0204]

[0205] Table 6 Hybrid BBU Hardware Specifications

[0206]

[0207] Table 7 Write-Guard Mode Control Parameters

[0208]

[0209] The following section provides a specific example to further illustrate the BBU protection capability prediction method for a storage system based on digital twins provided in this embodiment.

[0210] Example: A centralized storage controller equipped with a hybrid BBU of LFP battery packs and supercapacitors.

[0211] 1. Hardware Configuration

[0212] The storage system adopts a centralized storage architecture and is equipped with dual controllers (primary / backup mode). BBU modules are installed within each controller, employing a hybrid topology of LFP battery packs and supercapacitors. The LFP battery packs provide the primary backup power, with nominal parameters of 12.5V / 20Ah and a rated energy of approximately 250Wh. Under typical storage loads (write cache power consumption of approximately 2.6W), they can provide approximately 96 hours of theoretical protection time. The supercapacitors buffer instantaneous power fluctuations, providing peak power support during write bursts and seamlessly taking over power during power outages, preventing accelerated aging of the LFP batteries due to high-current discharge.

[0213] The BMC module uses an Aspeed 2500 processor, is equipped with 512MB of DDR4 RAM, and runs an embedded Linux operating system. The digital twin engine is deployed on the BMC side, with the LSTM model parameters controlled to within 500K and the model file size less than 2MB, meeting the storage constraints of the embedded environment.

[0214] 2. Digital Twin Engine Operation

[0215] The digital twin engine acquires voltage, current, and temperature data output from the BMS (Battery Management System) at a frequency of 1Hz, and executes the Dual EKF algorithm to identify SOC and SOH parameters online. The Dual EKF consists of two coupled filters: the first EKF uses SOC as the primary state variable and describes the battery's dynamic characteristics using the Thevenin equivalent circuit model; the second EKF uses SOH (characterized by capacitance decay rate and ESR growth rate) as the state variable and describes parameter drift using an aging empirical model. The two filters execute alternately and mutually correct each other to achieve joint online estimation of SOC and SOH.

[0216] The LSTM prediction module receives the SOC / SOH time series and outputs prediction results at two time scales: short-term prediction is the protection time curve within the next 72 hours, with a time resolution of 1 hour and a prediction error of no more than 10%; long-term prediction is the RUL estimate within the next 12 months, with a prediction error of no more than 15%. In this embodiment, the calculation of the protection time comprehensively considers the available capacity of the LFP battery pack (constrained by both SOC and SOH), the auxiliary support capability of the supercapacitor, and the power consumption requirements of the current write cache of the storage system, and finally outputs a unified index in "hours": the predicted remaining protection time.

[0217] End-to-end inference latency meets real-time requirements: Dual EKF single execution time is 10ms~30ms, LSTM single inference time is 5ms~20ms, data acquisition and preprocessing time is about 10ms, and the total latency is less than 100ms.

[0218] 3. BBU aging process and protection time decay curve

[0219] In this embodiment, the aging process of the LFP battery pack conforms to the typical degradation pattern of lithium batteries: slow degradation in the early stage, accelerated degradation in the middle stage, and rapid degradation in the later stage. Correspondingly, the predicted remaining protection time exhibits a non-linear degradation curve with the service life. Typical parameters for each stage are given below:

[0220] Year 0 (Initial Deployment): SOH is 100%, ESR is the nominal value (approximately 20mΩ), and the predicted remaining protection time is 96 hours. The digital twin engine is in the initial learning phase, using the running data from the first 30 days to complete model parameter calibration. At this time, the system enables Write-Back full-performance mode, the storage controller allows dirty data to reside in the write cache for the longest possible time, and the background flushing frequency maintains the baseline value (1x).

[0221] Years 1-2 (Slow Decline Period): SOH slowly decreases to approximately 95%, ESR increases to approximately 24 mΩ (a 20% increase), and the predicted remaining protection time is approximately 84 hours. During this stage, SOH decreases by an average of approximately 2.5% per year, and the protection time decreases by an average of approximately 6 hours per year. The predicted remaining protection time remains greater than 72 hours, and the hierarchical strategy engine maintains the Write-Back mode. The digital twin engine enters a continuous learning phase, continuously optimizing the prediction accuracy of the LSTM model using accumulated runtime data.

[0222] Years 3-4 (linear decay period): SOH decreases from approximately 95% to approximately 85%, ESR increases from approximately 24mΩ to approximately 34mΩ (cumulative increase of 70%), and predicted remaining protection time decreases from approximately 84h to approximately 72h. During this stage, SOH decays by an average of approximately 5% per year, and protection time decreases by an average of approximately 6h per year. The decay rate doubles compared to the previous period but remains within the linear range. When the predicted remaining protection time first drops to 72h, the tiered strategy engine triggers a downgrade switch from Write-Back to Write-Guard. Since the hysteresis mechanism sets the upgrade threshold to 80h, and the BBU aging process is irreversible, the Write-Back mode is not restored in this embodiment.

[0223] The specific parameters of the Write-Guard enhanced mode are as follows: I / O rate limiting adopts the token bucket algorithm, and the token generation rate is set to 70% of the baseline write bandwidth, that is, the peak write rate is limited to 70% of the baseline value (30% lower than Write-Back mode); at the same time, the background disk flushing frequency is increased from 1x to 2x, and the maximum residence time of dirty data in the cache is shortened by 50%. With the assistance of supercapacitors, the increase in I / O latency in Write-Guard mode is controlled within 15%~25%, which is far lower than the more than 50% latency increase in Write-Through mode.

[0224] Years 5-6 (Accelerated Decay Period): SOH decreases from approximately 85% to approximately 75%, ESR increases from approximately 34mΩ to approximately 50mΩ (cumulative increase of 150%), and the predicted remaining protection time decreases from approximately 72 hours to approximately 48 hours. During this stage, SOH decreases by an average of approximately 5% per year, but the protection time decreases by an average of approximately 12 hours per year, doubling the decay rate compared to years 3-4, entering the nonlinear accelerated decay range. At this point, the predicted remaining protection time is between 48 hours and 72 hours, falling within the middle working zone of Write-Guard mode. The hierarchical strategy engine dynamically adjusts the current limiting parameters based on the current protection time value: the closer the protection time is to 48 hours, the stricter the token bucket generation rate (down to a minimum of 50% of the baseline value); the closer the protection time is to 72 hours, the more lenient the current limiting (up to a maximum of 70% of the baseline value).

[0225] During this period, the operations and maintenance (O&M) decision support layer continuously monitors the SOH (Storage Overload) decay rate. When the ESR (Expiratory Storage Rate) growth rate exceeds a preset threshold (e.g., ESR growth > 50% within 6 months), the BBU (Storage Unit) is determined to have entered an accelerated aging phase, triggering a preventative replacement warning. The O&M decision support layer combines historical load data from the storage controller to predict low-load periods within the next 6 months, generating candidate maintenance windows (e.g., the continuous low I / O period from 2:00 AM to 4:00 AM), and pushes replacement recommendations and candidate windows to the O&M management platform. O&M personnel can then plan spare parts procurement and replacement schedules 6 months in advance based on these recommendations.

[0226] Year 7 (Rapid Decay Phase): SOH rapidly decreases from approximately 75% to approximately 65%, ESR increases from approximately 50mΩ to approximately 70mΩ (cumulative increase of 250%), and the predicted remaining protection time rapidly decreases from approximately 48 hours to approximately 20 hours. During this phase, SOH decreases by an average of approximately 10% annually, and the protection time decreases by an average of approximately 28 hours annually. The decay rate doubles again compared to years 5-6, entering the rapid decay phase. When the predicted remaining protection time drops below 24 hours, the tiered policy engine triggers an emergency degradation switch from Write-Guard to Write-Through, and simultaneously sends a highest-priority replacement alarm to the operations and maintenance management platform, requiring operations and maintenance personnel to complete the BBU replacement within 48 hours. In Write-Through mode, all write requests are directly written to disk, the write caching function is completely disabled, there is zero risk of data loss, but write performance is reduced to the lowest level.

[0227] 4. Explanation of the function of the hysteresis mechanism

[0228] The hysteresis mechanism in this embodiment is designed to prevent frequent oscillations during policy switching. The degradation thresholds are set to 72h (Write-Back → Write-Guard) and 24h (Write-Guard → Write-Through), while the upgrade thresholds are set to 80h (Write-Guard → Write-Back) and 36h (Write-Through → Write-Guard). Under normal BBU aging conditions, the protection time exhibits a monotonically decreasing trend. The hysteresis window primarily addresses the following two boundary scenarios:

[0229] Scenario 1: Boundary fluctuations caused by prediction errors. When the actual protection time is around 72 hours, short-term prediction errors (±10%) may cause the predicted value to fluctuate between 65 hours and 79 hours. Without a hysteresis mechanism, the strategy may frequently switch between Write-Back and Write-Guard. An 8-hour hysteresis window between the 72-hour downgrade threshold and the 80-hour upgrade threshold ensures strategy stability within the prediction error range.

[0230] Scenario 2: Temporary changes in protection time due to sudden load changes. When the storage system experiences a sudden increase in write cache power consumption due to a business surge, the predicted remaining protection time may momentarily drop beyond the 72-hour threshold. The hysteresis mechanism ensures that policy rollback is only allowed when the protection time substantially recovers to over 80 hours (corresponding to a significant reduction in load or recovery of SOH, the latter of which does not occur in actual aging).

[0231] 5. Collaborative operation of hybrid BBUs

[0232] In this embodiment, the collaborative working mechanism between the supercapacitor and the LFP battery pack is as follows: During normal write operations, the supercapacitor buffers the instantaneous power peaks caused by write bursts, preventing the LFP battery from frequently experiencing high current surges and delaying battery aging; when a power failure occurs, the supercapacitor seamlessly takes over the power supply within a millisecond timescale, ensuring that the controller has enough time to perform the cache flushing operation; the LFP battery continues to provide sustained power support after the supercapacitor's energy is depleted, and its protection time is predicted in real time by the digital twin engine.

[0233] 6. Technical Effect Analysis

[0234] The technical effects can be quantitatively analyzed from the following three dimensions:

[0235] First, BBU utilization is improved. Under a fixed 2-year replacement cycle, the storage system requires 5 BBU replacements within its 10-year design lifespan (at the end of years 2, 4, 6, 8, and 10), with an average annual replacement cost of 0.5 BBUs. With the predictive maintenance method proposed in this application, the BBU can remain in service until the State of Health (SOH) drops to approximately 65% ​​(around year 7), requiring only about 2 replacements within 10 years (after year 7 and year 10), reducing the average annual replacement cost to approximately 0.2 BBUs. The single service life of a BBU is extended from 2 years to approximately 5-6 years, and the total spare parts usage is reduced by approximately 60%.

[0236] Second, the risk of unplanned downtime is reduced. In traditional solutions, if the BBU fails unexpectedly before replacement (e.g., due to accelerated aging caused by high temperature leading to a sudden drop in SOH), it will cause unplanned downtime. The predictive maintenance method in this application incorporates BBU failure events into planned maintenance through continuous monitoring and replacement warnings up to 6 months in advance, reducing the risk of unplanned downtime by more than 90%.

[0237] Third, the smoothness of performance under tiered protection. In the 7-year service life of this embodiment, Write-Back mode covers approximately 4 years (years 0-4), Write-Guard mode covers approximately 3 years (years 4-7), and Write-Through mode is triggered only briefly (approximately several days to weeks) after the protection time drops below 24 hours. Compared to traditional binary switching schemes (direct switching between Write-Back and Write-Through), this application's three-tiered strategy transforms a precipitous performance drop into a smooth, gradual performance change. Write-Guard mode maintains average write performance at approximately 75% or more of the baseline value during the 3-year service life, significantly improving the user experience of the storage system during periods of BBU sub-optimal performance.

[0238] It should be noted that the specific algorithms, threshold parameters and hardware configurations involved in this embodiment are only preferred implementations of this embodiment, and the protection scope of this embodiment is not limited thereto.

[0239] Regarding state estimation algorithms, Dual EKF can be replaced by other nonlinear state estimation methods with equivalent functionality, including but not limited to UKF (Unscented Kalman Filter), PF (Particle Filter), or deep learning-based end-to-end state estimation networks. The parameterization methods of SOC and SOH can also be adjusted accordingly.

[0240] In terms of trend prediction algorithms, LSTM can be replaced by other deep learning architectures with equivalent temporal modeling capabilities, including but not limited to GRU (Gated Recurrent Unit), Transformer temporal prediction model, TCN (Temporal Convolutional Network), or N-BEATS and other temporal decomposition models.

[0241] Regarding the thresholds for the tiered strategy, the Write-Back / Write-Guard switching threshold of 72h and the Write-Guard / Write-Through switching threshold of 24h, as well as the hysteresis window upgrade thresholds of 80h and 36h, can all be adjusted according to the actual configuration of the storage system. The adjustment criteria include, but are not limited to: the total capacity of the storage system's write cache, the system's peak write bandwidth, and the service's tolerance for data loss. For example, for a high-end storage array equipped with a 1TB write cache, the 72h threshold can be increased to 96h or higher.

[0242] Regarding the BBU type, the LFP battery pack + supercapacitor hybrid topology in the embodiment can be replaced with other types of energy storage units, including but not limited to NCM (Nickel Cobalt Manganese) batteries, LTO (Lithium Titanate) batteries, solid-state batteries, or pure supercapacitor solutions.

[0243] In terms of hardware platform, the processor architecture of the BMC module is not limited to Aspeed 2500, and can be replaced with other embedded processors with equivalent computing power, such as ARM Cortex-M series, RISC-V architecture processors, or x86 low-power processors.

[0244] Regarding storage architecture, the technical solution of this application is not only applicable to centralized storage controllers, but also to node-level BBU protection in distributed storage systems, host-level BBU protection in software-defined storage (SDS) environments, and endpoint protection scenarios in NVMe-oF (NVMe over Fabrics) storage architectures.

[0245] The BBU protection capability prediction method for a storage system based on digital twins provided in this embodiment has the following advantages:

[0246] 1. Proactive prediction replaces passive alarms, improving protection reliability.

[0247] Current technologies generally employ a passive alarm mechanism triggered by a State of Health (SOH) threshold. This means that a replacement alarm is triggered when the SOH of the Battery Backup Unit (BBU) falls below a fixed threshold (typically 70%–80%). This approach has two inherent drawbacks: first, the BBU is already in a sub-healthy state when the alarm is triggered, leaving insufficient response time for maintenance personnel; second, the fixed threshold cannot reflect the dynamic differences in the BBU's actual protection capability as the load changes.

[0248] This embodiment deploys a lightweight digital twin engine on the BMC (Baseboard Management Controller) side, employs a Dual EKF (Extended Kalman Filter) to identify SOC (State of Charge) and SOH parameters online, and combines this with LSTM (Long Short-Term Memory) to predict the decay trend of the remaining protection time of the BBU, achieving a shift from "replacing after failure" to "advance planning." Specifically, the prediction method in this embodiment can issue an early warning 72 to 96 hours before the BBU protection time drops to the policy switching threshold, which is about 4 to 6 times more early warning time compared to traditional SOH threshold alarms (which are usually triggered when the protection capability is severely insufficient). At the same time, the short-term prediction error is controlled within 10% (within the 72-hour prediction window), and the long-term RUL (Remaining Useful Life) prediction error is controlled within 15%, providing a quantitative basis for operation and maintenance decisions.

[0249] 2. The protection strategy is dynamically adaptive, achieving a balance between performance and security.

[0250] Existing storage systems typically employ a binary logic for switching write strategies: when the BBU (Browser Protector) is functioning normally, Write-Back mode is enabled for optimal performance; when the BBU malfunctions, it switches to Write-Through mode to ensure data security. This either-or switching approach leaves little room for compromise between performance loss and security, creating a risk of sudden data loss during the transition period when the BBU's protection capabilities decline.

[0251] The tiered protection strategy proposed in this embodiment dynamically adjusts the write strategy based on the predicted remaining protection time output by the digital twin engine. When protection capability is sufficient (predicted remaining protection time greater than 72 hours), it maintains the Write-Back full-performance mode. When protection capability decreases (predicted remaining protection time between 24 hours and 72 hours), it activates the Write-Guard intermediate state. When protection capability is severely insufficient (predicted remaining protection time less than 24 hours), it switches to Write-Through mode. This smooth degradation mechanism avoids the abrupt switch from full performance to full security in traditional solutions, ensuring that the I / O performance degradation of the storage system is controllable. Data from the embodiment shows that using Write-Guard mode can smoothly limit the peak write rate to 70% of the baseline value (i.e., a 30% reduction), while increasing the background disk flushing frequency to twice the original value. Even within the range of 25% to 50% protection capability decay, it can still maintain approximately 70% to 85% of write performance, significantly outperforming the performance of traditional binary switching solutions during BBU sub-health periods.

[0252] 3. Three-level graded degradation system for refined power failure protection.

[0253] Existing Write-Back / Write-Through binary switching strategies only have two extreme states, failing to achieve a fine balance between security and performance. This embodiment introduces a Write-Guard intermediate state, forming a three-level hierarchical system, coupled with hysteresis-based strategy switching boundary management (degradation threshold 72h, upgrade threshold 80h), effectively avoiding frequent oscillations of the strategy near the boundary.

[0254] The Write-Guard mode controls the peak write rate through a token bucket-based I / O rate limiting mechanism and increases the disk flushing frequency to shorten the residence time of dirty data in the cache. When the protection time is between 24h and 72h, the rate limiting parameters are dynamically adjusted according to the current protection time value: the closer the protection time is to 24h, the stricter the rate limiting and the more frequent the disk flushing; the closer the protection time is to 72h, the more lenient the rate limiting. This continuously adjustable protection strength curve allows system administrators to achieve the optimal balance between security and performance without manual intervention. The hysteresis mechanism ensures the stability of policy switching: Write-Guard degradation is triggered when the protection time drops to 72h, and Write-Back mode is only allowed to be restored when the protection time rises back to 80h or more. Considering the irreversibility of the BBU aging process, this hysteresis window mainly prevents policy oscillations caused by prediction errors or load fluctuations in actual operation, ensuring the determinism of graded switching.

[0255] 4. Intelligent operation and maintenance decision-making reduces total lifecycle costs.

[0256] In this embodiment, the operation and maintenance decision support layer proactively pushes replacement suggestions and optimal maintenance windows based on the SOH trend and RUL prediction results output by the digital twin engine before the BBU actually fails. Compared to traditional fixed-cycle replacement schemes (usually mandatory replacement every 2-3 years), the predictive maintenance method in this embodiment can dynamically adjust the replacement timing according to the actual degradation status of the BBU.

[0257] Taking the LFP (Lithium Iron Phosphate) + supercapacitor hybrid BBU in this embodiment as an example: Under the traditional fixed 2-year replacement cycle, the BBU is replaced when the State of Health (SOH) is still about 95% at the end of the second year, utilizing only about 5% of the capacity decay space during a single service life; with the predictive maintenance method of this application, the BBU can be used until the SOH drops to about 65% in the 6th or 7th year before replacement, utilizing about 35% of the capacity decay space during a single service life. Within the 10-year design lifecycle of the storage system, the fixed 2-year replacement plan requires 5 replacements, while the predictive maintenance plan only requires about 2 replacements, reducing BBU spare parts usage by about 60%. Simultaneously, since the replacement window can be planned 6 months in advance, maintenance personnel can choose to perform replacement operations during periods of low storage system load, avoiding unplanned downtime impacting business continuity. The long-term error of the RUL prediction is controlled within 15%, ensuring a confidence level of over 85% for the replacement time window, providing reliable data support for maintenance budget preparation and spare parts inventory management.

[0258] 5. Lightweight edge deployment, avoiding network dependence

[0259] In this embodiment, the digital twin engine is deployed on the BMC side rather than in the cloud or on a standalone server. It uses an ARM Cortex-A53 processor, requiring only 512MB to 1GB of RAM, and the LSTM model parameters are kept below 500K. This deployment scheme offers the following advantages: First, it avoids dependence on external network connections, ensuring normal execution of BBU health monitoring and tiered protection strategies even during network outages, meeting the high availability requirements of the storage system. Second, inference latency is controllable, with DualEKF execution time of 10ms to 30ms and LSTM inference time of 5ms to 20ms, resulting in a total end-to-end latency of less than 100ms, meeting the timeliness requirements for real-time policy switching. Third, the BMC, as a standard component of the storage system, requires no additional hardware investment, reducing overall system cost and deployment complexity. This lightweight architecture allows the technical solution of this application to be directly applied to resource-constrained environments such as edge computing nodes and embedded storage devices.

[0260] In summary, this embodiment uses digital twin technology as the core driving force. Through the accurate prediction of BBU health status and the dynamic linkage of storage system protection strategies, a closed-loop management system of "prediction-decision-execution-feedback" is formed. This system systematically solves the four key defects of existing technologies in the field of back-end storage data protection, maximizes the I / O performance of the storage system while ensuring data security, and significantly improves the refinement and predictability of BBU operation and maintenance management.

[0261] Example 2

[0262] Please see Figure 2 This embodiment provides a storage system BBU protection capability prediction system based on digital twins. The overall architecture of the system consists of four logical layers, including the local MCU of the battery backup unit (BMS monitoring layer), the digital twin engine (BMC side), the hierarchical policy engine (storage controller side), and the operation and maintenance platform (operation and maintenance decision support layer). The digital twin engine communicates with the local MCU of the battery backup unit, the hierarchical policy engine, and the operation and maintenance platform. The digital twin engine is deployed in the baseboard management controller of the storage system to implement the storage system BBU protection capability prediction method based on digital twins provided in Embodiment 1. The hierarchical policy engine is deployed in the I / O scheduling software stack of the storage controller to dynamically adjust the cache write policy.

[0263] Specifically, the BMS monitoring layer is deployed within the BBU's local MCU (typically an ARM Cortex-M4 series, operating at 80~168MHz, with 128~256KB of RAM and 256~512KB of Flash). This MCU acquires analog signals such as the voltage of each string in the LFP battery pack, the voltage of each cell in the supercapacitor pack, the total current (Hall sensor or precision shunt), and the temperature at multiple points (NTC thermistor array) through an ADC multiplexer. The LFP battery pack and the supercapacitor pack are connected to the backup power bus of the storage system via an intelligent power distribution circuit. The power distribution controller automatically switches the power supply based on the power failure detection signal. The local MCU runs a lightweight ECM (Equivalent Circuit Model) to achieve millisecond-level local SOC estimation and fault detection, and has independent hard protection logic (overvoltage, undervoltage, overtemperature, overcurrent) to ensure basic protection is still provided during communication interruptions.

[0264] The digital twin engine is deployed in the Baseboard Management Controller (BMC) of the storage system. The typical hardware specifications of the BMC are a dual-core or quad-core ARM Cortex-A53 processor with a clock speed of 1.0~1.5GHz, 512MB~1GB of RAM, and an embedded Linux operating system. The digital twin engine, as a multi-threaded daemon process in the BMC user space, communicates with the local MCU of the BBU via PMBus / SMBus (System Management Bus) or I2C bus to receive telemetry data. Internally, the digital twin engine integrates a data fusion module, a SOC / SOH joint estimation module, a trend prediction module, a protection time calculation module, a model management module, a multi-trigger update scheduler, an edge inference engine, and a calibration feedback module. These modules work collaboratively through a shared memory queue and an inter-thread message passing mechanism.

[0265] More specifically, the digital twin engine is deployed in the BMC user space and is a multi-threaded real-time computing process responsible for building and driving the lightweight digital twin model of the BBU. In this embodiment, the digital twin engine adopts a lightweight digital twin architecture to ensure real-time deployability on the embedded BMC platform. Typical BMC hardware specifications are: processor ARM Cortex-A53 dual-core or quad-core, clock speed 1.0~1.5GHz; RAM capacity 512MB~1GB (of which the memory budget allocated to the digital twin engine is ≤128MB); persistent storage eMMC (embedded MultiMediaCard) 4~8GB; operating system is a trimmed version of embedded Linux (such as OpenBMC or a Yocto custom image).

[0266] The memory budget allocation for the digital twin engine is as follows: Circular buffer for the data fusion module (8MB reserved, supporting storage of 7200 historical records, covering a static 60s×120h runtime cycle); Dual EKF state vector and covariance matrix (2MB reserved); LSTM lightweight model weight parameters (4MB reserved); Edge inference engine runtime memory pool (8MB reserved, including tensor cache and intermediate results); Protection time calculation and trend prediction workspace (4MB reserved); Persistent storage for the model management module (8MB reserved); Process's own code segment and stack (4MB reserved). The total memory budget is ≤38MB, far below the 128MB allocation limit, with a memory margin >70%.

[0267] The hierarchical policy engine is deployed in the storage controller's software stack and is tightly coupled with the cache management layer and I / O scheduling layer. The hierarchical policy engine obtains status information such as Predicted Remaining Time (RPT) and State of Health (SOH) output by the digital twin engine through the IPMI (Intelligent Platform Management Interface) or internal IPC (Inter-Process Communication) channels provided by the BMC, and dynamically switches cache write policies according to the built-in hierarchical policy rules.

[0268] The operation and maintenance decision support layer is deployed on an external storage management platform or cloud operation and maintenance system. It interacts with the digital twin engine on the BMC side through SNMP (Simple Network Management Protocol) Trap, Redfish standard API or HTTPS RESTful interface to receive BBU status reports, predictive maintenance suggestions and proactive early warning events.

[0269] In the BBU protection capability prediction system for a storage system based on digital twins provided in this embodiment, the data flow and control flow are as follows:

[0270] Data Flow ① (Uplink, BBU→BMC): The BMS monitoring layer reports telemetry data frames to the BMC via PMBus / SMBus at fixed intervals (1s in running mode, 60s in idle mode). The data frame includes: voltage of each LFP battery string (accuracy ±5mV), total current (accuracy ±50mA), multi-point temperature (accuracy ±1℃), locally estimated SOC, total voltage of the supercapacitor bank, estimated ESR, power distribution circuit status word, and fault flag. The communication protocol adopts the PMBus 1.3 standard, with a maximum data rate of 400Kbps, a single reported data volume of approximately 64~128 bytes, and a single transmission time of <5ms.

[0271] Data Flow ② (Internal, between modules of the digital twin engine): The data fusion module writes the calibrated state snapshot to a shared memory circular buffer; the SOC / SOH joint estimation module reads this buffer, executes the Dual EKF algorithm, and outputs parameters such as SOC, Q_max, and R0 to the trend prediction module; the LSTM model of the trend prediction module generates a future protection capability curve based on historical SOH sequences and outputs it to the protection time calculation module; the protection time calculation module integrates the discharge characteristics of the LFP battery pack and supercapacitor pack, the current estimated amount of dirty data in the cache, and a typical write load model to calculate the predicted remaining protection time (RPT). After verification by the calibration feedback module, the RPT result is pushed to the hierarchical strategy engine via a Unix Domain Socket.

[0272] Control Flow ③ (Downlink, BMC → Storage Controller): The digital twin engine encodes the RPT prediction results, SOH evaluation values, and policy recommendations into JSON-formatted policy instruction packets, which are then distributed via the IPMB (Intelligent Platform Management Bus) or a dedicated Ethernet channel between the BMC and the storage controller. The policy switching latency budget is: Digital twin calculation 10~50ms + policy decision 5ms + policy distribution to the controller 20ms + policy activation 10ms = Total latency <100ms.

[0273] Data Flow ④ (Uplink, BMC → Operations and Maintenance Platform): The digital twin engine simultaneously outputs the following to the operations and maintenance decision support layer via SNMP Trap / RedfishEvent / HTTPS active push or passive polling: BBU SOH trend chart data (30-day rolling), RUL prediction report (365-day long-term trend), suggested replacement time window (based on RUL and operations and maintenance window optimization algorithm), and hierarchical protection status change event log.

[0274] Configuration Stream ⑤ (Downlink, BMC→BBU MCU): The BMC can send configuration parameters to the local MCU of the BBU via PMBus, including: EKF observation noise covariance update value, protection threshold adjustment, firmware upgrade package, etc.

[0275] For details on the specific implementation of each module in a digital twin-based storage system BBU protection capability prediction system, please refer to the above description of the limitations of a digital twin-based storage system BBU protection capability prediction method, which will not be repeated here.

[0276] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

Claims

1. A method for predicting the protection capability of a storage system's BBU based on digital twins, characterized in that, The method is applied to a digital twin engine, which is deployed in the baseboard management controller of a storage system, and the method includes: Step 1: Receive telemetry data collected and reported locally by the battery backup unit; wherein, the battery backup unit is a hybrid topology structure in which battery energy storage unit and capacitor energy storage unit work together; the telemetry data includes: battery energy storage unit operation data, capacitor energy storage unit operation data and power distribution controller status word; Step 2: Perform data preprocessing on the telemetry data, and construct a real-time operating status snapshot of the battery backup unit based on the preprocessed telemetry data; the data preprocessing includes: time alignment, outlier removal, temperature compensation, and auxiliary data fusion; Step 3: Using a dual extended Kalman filter, combined with the real-time operating status snapshot, the health status parameters of the battery backup unit are identified online, and health status time series data is generated; Step 4: Based on the health status time series data, use a time series prediction model to predict the health status degradation trend of the battery backup unit; Step 5: Calculate and predict the remaining protection time based on the effective energy of the current battery energy storage unit and the capacitor energy storage unit, the current load information of the storage system, and the health status decay trend; Step 6: Send the predicted remaining protection time to the hierarchical policy engine; the hierarchical policy engine dynamically adjusts the cache write policy according to the predicted remaining protection time; wherein, the dynamic adjustment of the cache write policy includes: a three-level hierarchical protection policy of write-back, write protection and direct write; the hierarchical policy engine is deployed in the I / O scheduling software stack of the storage controller.

2. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, Also includes: The predicted remaining protection time, the health status decay trend, and the health status parameters of the battery energy storage unit and the capacitor energy storage unit are sent to the operation and maintenance platform so that the operation and maintenance platform can generate and display the health status trend, remaining service life prediction, and replacement suggestion window of the battery backup unit, and push proactive warning information to the operation and maintenance platform when the predicted remaining protection time is lower than the preset warning threshold.

3. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, In step 3, the dual extended Kalman filter includes: fast timescale state of charge estimation and slow timescale health state parameter identification. The fast timescale state of charge estimation and the slow timescale health state parameter identification are coupled and closed-loop corrected. The fast timescale state of charge estimation is updated once in each telemetry data reporting cycle, and the slow timescale health state parameter identification is updated once after accumulating multiple reporting cycles, and the updated health state parameters are fed back into the model parameters of the fast timescale state of charge estimation.

4. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, Step 4, which involves using a time-series prediction model to predict the health status degradation trend of the battery backup unit, specifically includes: Using the daily feature values ​​in the health status time series data as input, the time series prediction model performs forward propagation prediction and outputs the health status decay trend for the next preset number of days. When the deviation between the predicted value of the time series prediction model and the health status parameters identified in real time by the dual extended Kalman filter exceeds a preset threshold, a calibration process is triggered. The input sequence of the time series prediction model is drift corrected based on the real-time identification result of the dual extended Kalman filter, and the prediction is re-executed.

5. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, The time series prediction model adopts a long short-term memory network model; the long short-term memory network model is a single-layer structure with a hidden layer dimension of 32 or 64, and the input time step is a preset number of days.

6. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, In step 5, the formula for calculating the predicted remaining protection time is: RPT = E_total / P_total_discharge - t_total Wherein, E_total represents the total effective energy of the battery energy storage unit and the capacitor energy storage unit, P_total_discharge represents the total power consumption during the power outage discharge period, and t_total represents the total time from the occurrence of the power outage to the complete flushing of the cached data.

7. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, In step 6, the hierarchical strategy engine dynamically adjusts the cache write strategy based on the predicted remaining protection time, specifically including: Read the current policy status, which includes: write-back mode, write-protected mode, and write-through mode; If the current policy status is write-back mode and the predicted remaining protection time is less than or equal to 72 hours, the decision is to switch to write protection mode. If the current policy status is write protection mode and the predicted remaining protection time is greater than or equal to eighty hours, the decision is to switch to write-back mode. If the current policy status is write protection mode and the predicted remaining protection time is less than or equal to 24 hours, the decision is to switch to write-through mode. If the current policy status is write-through mode and the predicted remaining protection time is greater than or equal to 30 hours, the decision is to switch to write protection mode. If the minimum dwell time check is triggered, the decision is to maintain the current state; In all other cases, the decision is to maintain the current state.

8. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, The digital twin engine uses a multi-trigger mode for updates, which includes: timed updates, event-triggered updates, and exception-triggered updates.

9. The method for predicting the BBU protection capability of a storage system based on digital twins according to claim 1, characterized in that, The battery energy storage unit is a lithium iron phosphate battery pack, and the capacitor energy storage unit is a supercapacitor.

10. A storage system BBU protection capability prediction system based on digital twin, characterized in that, The system includes a local MCU for the battery backup unit, a digital twin engine, a hierarchical strategy engine, and an operation and maintenance platform. The digital twin engine is communicatively connected to the local MCU for the battery backup unit, the hierarchical strategy engine, and the operation and maintenance platform. The digital twin engine is deployed in the baseboard management controller of the storage system and is used to implement the storage system BBU protection capability prediction method based on digital twin as described in any one of claims 1-9. The hierarchical strategy engine is deployed in the I / O scheduling software stack of the storage controller and is used to dynamically adjust the cache write strategy.

Images