A battery online monitoring and fault prediction system

Abstract

This invention relates to the field of power system technology, specifically disclosing a battery online monitoring and fault prediction system. It adopts an edge-cloud collaborative architecture, including the following modules and an end-to-end closed-loop collaborative process. Each module achieves real-time data and command linkage through a cross-layer feature interaction protocol. Specifically, it includes: a sensing and acquisition module, a hierarchical fusion diagnostic module, an adaptive prediction and decision-making module, and a closed-loop optimization and digital twin module. This invention, through its edge-cloud collaborative architecture and multi-dimensional sensing and acquisition, ensures comprehensive and real-time monitoring, reduces transmission load, and resolves the contradictions of traditional single-parameter monitoring and processing modes. Relying on the hierarchical fusion diagnostic module, it improves the accuracy of fault prediction and manages risks in advance. Combined with adaptive decision-making, closed-loop optimization, and digital twin technologies, it achieves personalized dynamic optimization of maintenance strategies and a full-link closed loop.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to an online monitoring and fault prediction system for batteries. Background Technology

[0002] As a core energy storage component in power systems, energy storage devices, communication base stations, and transportation vehicles, the operational stability and service life of batteries directly determine the reliable operation and maintenance costs of the overall system. With increasing reliance on energy storage devices across various sectors, battery banks often operate continuously under complex conditions, making them susceptible to performance degradation and localized anomalies due to multiple factors such as charge-discharge cycles, operating environment, and material aging. Failure to monitor and intervene in a timely manner can lead to escalating faults, system shutdowns, and even safety hazards.

[0003] Currently, existing battery monitoring technologies are mostly based on offline or semi-online monitoring, which has significant limitations. Firstly, monitoring methods often focus on collecting single electrical parameters, lacking comprehensive capture of multi-dimensional information such as the operating environment and battery status indicators, making it difficult to fully reflect the battery's health status. Secondly, data processing models have shortcomings; either relying on lightweight edge computing leads to insufficient diagnostic accuracy, or relying entirely on cloud processing causes data transmission delays, failing to achieve real-time monitoring and rapid response. Thirdly, fault prediction capabilities are weak, often relying on post-event analysis or simple trend judgments based on historical fault data, lacking in-depth analysis of fault evolution patterns and consideration of multiple risk coupling effects, making it difficult to provide early warnings of potential faults. Fourthly, existing systems are mostly open-loop architectures, lacking effective linkage between monitoring, diagnosis, and maintenance strategies; model parameters cannot be dynamically optimized based on actual operational feedback, resulting in insufficient flexibility to adapt to different application scenarios and failing to meet the needs of refined battery operation and maintenance under complex operating conditions.

[0004] Furthermore, traditional monitoring systems lack the ability to trace and manage the entire lifecycle status of batteries. Maintenance strategies are mostly standardized and cannot be customized to take into account the differences between individual batteries and the needs of different scenarios, resulting in low operation and maintenance efficiency and wasted resources. Therefore, developing an online battery monitoring and fault prediction system with multi-dimensional perception, real-time collaborative processing, accurate fault prediction, and adaptive optimization capabilities has become an urgent need in the field of energy storage equipment operation and maintenance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an online battery monitoring and fault prediction system, which solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online monitoring and fault prediction system for batteries, which adopts an edge-cloud collaborative architecture, including the following modules and an end-to-end closed-loop collaborative process, wherein each module achieves real-time linkage of data and instructions through a cross-layer feature interaction protocol; The sensing and acquisition module, deployed at the edge, is used to acquire multi-dimensional operating parameters of individual cells and battery clusters in the battery pack in real time, and form a time-series observation sequence after anti-interference preprocessing. The hierarchical fusion diagnostic module adopts a lightweight edge computing and cloud-based deep analysis mode. Its input end is connected to the output end of the sensing and acquisition module. Based on the time-series observation sequence, it performs health status assessment, abnormal status identification and risk status prediction in parallel, and outputs a fused health status vector, risk status vector and comprehensive fault risk level. The adaptive prediction and decision-making module is deployed in the cloud. Its input end is connected to the output end of the hierarchical fusion diagnosis module. It combines health status vector, risk status vector, historical fault data and scenario constraints to dynamically generate maintenance strategy suggestions and early fault warnings at multiple time scales. It supports hierarchical push of strategies and manual intervention interface. The closed-loop optimization and digital twin module is deployed across the edge and cloud, and communicates with the perception and acquisition module, the hierarchical fusion diagnosis module, and the adaptive prediction and decision-making module respectively. Based on actual operation feedback data and twin simulation results, it realizes online incremental updates of the diagnostic model and the prediction model. Through the constructed full-physical field digital twin, it completes strategy simulation verification and effect prediction, forming a full-link adaptive closed loop of acquisition, diagnosis, decision-making and optimization.

[0007] Preferably, the sensing and acquisition module adopts a three-level architecture of distributed acquisition, centralized synchronization, and edge preprocessing, and has the ability to adaptively adjust the sampling frequency. The acquired multi-dimensional operating parameters include: Electrical parameters: including at least the individual cell voltage, total battery cluster voltage, charging and discharging current, dynamic internal resistance, and individual cell and battery cluster temperature; Operating environment parameters: at least include ambient temperature, relative humidity, vibration intensity, and ambient dust concentration; Status identification parameters: must include at least the cumulative number of cycles, current state of charge (SOC), and battery aging stage label; After the collected data is synchronized with timestamps and preprocessed at the edges, it forms a data structure with a length of [length missing]. Time series observation sequence ,in For single cell batteries At time step The single-step multi-dimensional fusion data is used, and only feature data is uploaded at the edge to reduce the transmission bandwidth usage.

[0008] Preferably, the hierarchical fusion diagnostic module adopts a three-layer sub-network architecture of parameter sharing, task separation, and feature distillation. Each layer achieves information linkage through feature interaction channels, and a cross-sub-network knowledge transfer mechanism is introduced to improve diagnostic accuracy, specifically including: Health State Quantization Subnetwork: Based on time-series electrical parameters and state identification parameters, a hybrid model employing extended Kalman filtering, attention mechanism, and particle filter correction is used to calculate the health state vector of each individual cell. , Includes capacity decay Internal resistance growth rate and energy efficiency The three-dimensional integrated vector supports the adaptation of model parameters for batteries with different chemical systems; Anomaly pattern matching subnetwork: Based on real-time electrical and environmental parameters, it performs fast matching through a dynamically updated anomaly pattern library. It uses a cosine similarity and Mahalanobis distance fusion algorithm to calculate the matching degree, accurately identify the anomaly types and confidence levels related to overvoltage, undervoltage, overtemperature, internal resistance mutation, and insulation degradation, and outputs the temporal characteristics and spatial location of the anomaly occurrence. Risk State Deduction Subnetwork: Based on the current Based on anomaly identification results, historical fault sequences, and environmental trend data, an improved gated cyclic unit time-series prediction model is used to extrapolate the risk state vector within a preset time window. , This is used to quantify the potential risk intensity and coupling effects of four core faults: lithium plating, thermal runaway, capacity drop, and tab corrosion.

[0009] Preferably, the health status vector The calculation process incorporates an adaptive mechanism for the battery aging stage, as detailed below: The first step is to use the extended Kalman filter algorithm to initially estimate the real-time capacity based on the voltage and current sequence within the sliding time window. and dynamic internal resistance Then, a secondary correction is performed using a particle filter algorithm to reduce nonlinear errors; The second step involves calculating using a stage-adaptive coupling model. The core component, the formula is: ; ; in, This refers to the battery's rated capacity, which is the factory-calibrated value. This is the initial internal resistance of the battery, which is the factory calibration value. This represents the average temperature rise within the sliding window, which is the difference between the highest and lowest temperatures within the window. For reference temperature, the default value is 25℃; The cumulative number of battery cycles is obtained from the statistics of charge and discharge cycles. The capacity decay-temperature coupling coefficient is related to the battery chemistry system and was calibrated through offline experiments. The temperature effect index characterizes the nonlinear amplification effect of temperature rise on capacity decay. The internal resistance growth-cycle count coupling coefficient is related to the battery material properties and was calibrated through offline experiments. As a correction factor for the aging stage, according to Values ​​within the specified interval: Initial Next, take 0.8-0.9, mid-term. Next, take 1.0, final stage Next, a value of 1.1-1.2 is used to compensate for model biases at different aging stages; The third step is to calculate the energy efficiency of the most recent complete charge-discharge cycle. , This represents the total energy of the discharge. To determine the total charging energy, a charge / discharge rate correction term is introduced. , This refers to the charging rate. To optimize the calculation accuracy for discharge rate, the final result is... .

[0010] Preferably, the risk state vector This is achieved by combining a time-series risk propagation model with the fault coupling effect. The model definition and parameter relationships are as follows: ; in: For the current time step The risk state vector, Meaning of each component: For lithium plating risk, For the risk of thermal runaway, To mitigate the risk of a sudden drop in capacity, To represent the risk of tab corrosion, the values ​​are all in the range of [0,1], with larger values ​​indicating higher risk. For changes in health status, It is obtained by joint prediction based on the current operating conditions; This is the abnormal mode vector, where each element represents the confidence level of overvoltage, undervoltage, overtemperature, sudden change in internal resistance, and decrease in insulation, with values ​​ranging from [0,1]. This is a risk decay matrix. The diagonal elements range from 0.85 to 0.95, representing the natural decay trend of risk over time. The off-diagonal elements range from 0.01 to 0.05, representing the weak coupling effect between different risks. This is a health-risk coupling matrix, where non-zero elements range from 0.1 to 0.3, representing the contribution of deteriorating health status to different risks. The impact coefficient of corresponding capacity decay on lithium plating risk; This is an anomaly-risk trigger matrix, with elements ranging from 0 to 0.6, representing the trigger strength of different anomalies for specific risks; This is a risk-anomaly interaction matrix, with elements ranging from 0.05 to 0.2, representing the synergistic amplification effect between anomalies and existing risks. The Hadamard product is used to perform element-wise multiplication of corresponding vector elements; the parameters of the above matrix are obtained through training with historical fault data and are updated periodically by the closed-loop optimization module.

[0011] Preferably, the adaptive prediction and decision-making module employs a risk grading, scenario adaptation, and multi-objective optimization mechanism, maintains a multi-scale policy library, dynamically adjusts thresholds, and supports personalized policy customization. The strategy library includes three categories: long-term predictive maintenance plans, mid-term early warning and online balancing strategies, and emergency intervention strategies. Each strategy is associated with a cost-benefit evaluation model. The threshold dynamic adjustment rule is based on the failure rate over the past 30 days. Predicted remaining battery life The threshold is updated using the following formula: : ; ; in, This is a low-to-medium risk threshold used to trigger medium-term early warning strategies; This is a medium-to-high risk threshold used to trigger emergency intervention strategies; The failure rate over the past 30 days is calculated as the ratio of the number of failures to the total operating time. The remaining battery life is a predicted value, derived from the health state vector. It was deduced that; Strategy selection logic: When the overall fault risk level is low, output a long-term maintenance plan; when the level is medium or low, output a long-term maintenance plan. Any element in At the same time, it outputs mid-term early warning and online balancing strategies, prioritizing power supply to core loads; the level is high or Any element in In case of an emergency, an emergency intervention strategy is output, and the backup power switching mechanism is activated simultaneously to reduce the scope of the fault's impact.

[0012] Preferably, the online balancing strategy incorporates a risk contribution weighting and dynamic balancing cycle mechanism, ensuring dual adaptation of balancing weight calculation and operating mode to battery aging stage. ; in, For single cell batteries The equilibrium weights are determined by the weights; the higher the weight, the higher the equilibrium priority. For single cell batteries Real-time voltage; The average voltage of the battery cluster; For single cell batteries The real-time internal resistance; The average internal resistance of the battery cluster; For single cell batteries The maximum risk value in the risk state vector; For single cell batteries Capacity decay rate; For the weighting coefficients, satisfying Dual adaptation rules: Still state: ; Charging status: ; Discharge state: ; Late aging stage: additional Increase by 0.1 to enhance the balancing priority of degraded batteries; The balancing period is dynamically adjusted based on the maximum voltage difference within the cluster; the larger the voltage difference, the shorter the period.

[0013] Preferably, the closed-loop optimization and digital twin module integrates the entire process functions of data feedback, model update, simulation verification, and strategy iteration, and has the capability of virtual-real linkage calibration, specifically including: Feedback Data Pool: Stores actual fault events, maintenance operation records, twin simulation data, and corresponding preliminary diagnostic data, with a capacity of 1×10. 4 -1×10 6 The algorithm employs a priority experience replay mechanism to assign higher sampling weights to high-risk prediction hit data and fault missed / misjudged data, while also removing abnormal interference data through data cleaning algorithms. Model Incremental Learning Unit: An incremental learning strategy of freezing the bottom layer and fine-tuning the top layer is adopted. Every 500-2000 new feedback data are accumulated, the network parameters of the hierarchical fusion diagnostic module are fine-tuned, the coupling coefficient and risk propagation model matrix are updated, the AdamW optimizer is used, a regularization term is introduced to prevent overfitting, and the model update log is recorded for traceability. Digital twin simulation engine: Constructs a virtual battery pack based on an electrochemical-thermal-mechanical multiphysics coupling model, through... It dynamically corrects simulation parameters based on real-time running data, supports parallel simulation of different risk scenarios and maintenance strategies, and predicts the effect, potential chain reactions and economic costs of major maintenance strategies before implementation. The simulation error is ≤5%, and the simulation results serve as an important basis for strategy output.

[0014] Preferably, the digital twin simulation engine adopts a virtual-real synchronization-parameter self-calibration process to achieve dynamic optimization of simulation accuracy. The specific parameter correction process is as follows: based on Capacity decay in Internal resistance growth rate and energy efficiency The core parameters of the electrochemical model are corrected using the following formula: ; ; ; in, The capacity used in the simulation represents the actual usable capacity of the virtual battery. The internal resistance is used for simulation to characterize the conduction loss characteristics of the virtual battery. To simulate the amount of charge generated by the electrode reaction, and to characterize the reactivity of the electrode active materials in the virtual battery; The rated reaction capacity of the electrode is [value], which represents the battery design parameters; thermal model parameter correction: through [method / method]... Adjusting the heat dissipation coefficient during battery aging: ; in, For heat dissipation coefficient used in simulation; The baseline heat dissipation coefficient is determined by the battery casing material and structure. Mechanical stress model parameter correction: through internal resistance growth rate Corrected electrode expansion coefficient , The baseline expansion coefficient is used to ensure consistency between the simulated and physical battery pack's electro-thermal-mechanical coupling behavior; Simultaneous virtual-real calibration: Every hour, the deviation between the real-time data from the sensing and acquisition module and the twin simulation data is calculated. When the deviation of key parameters related to voltage and temperature exceeds 3%, a second parameter correction is triggered to ensure simulation accuracy.

[0015] Preferably, the battery online monitoring and fault prediction system is adapted and its functions are expanded for different application scenarios, supporting multi-scenario compatible deployment: In the energy storage power station scenario: the sensing and acquisition module supplements the insulation resistance of the battery cluster and the sealing performance of the cabinet; the risk state inference sub-network strengthens the coupled risk assessment of thermal runaway and insulation failure; the digital twin engine adds the functions of charge-discharge cycle life simulation and cluster collaborative control simulation; and the maintenance strategy prioritizes the needs of grid dispatch. Electric vehicle battery swapping station scenario: The sampling frequency is increased to 100Hz, a unique battery identification and full life cycle traceability function are added, the emergency intervention strategy adds battery swapping priority recommendation, the model incremental learning unit shortens the update cycle to accumulate 500 feedback data, and supports the balance between battery swapping efficiency and battery safety; For communication base station backup power scenarios: optimize the low-power acquisition mode in the idle state, add battery activation suggestions and backup power switching linkage strategies to the long-term maintenance plan, supplement the abnormal mode library with low power self-discharge and float charge voltage drift abnormal templates, and prioritize the policy output to ensure the power supply continuity of the base station core equipment.

[0016] This invention provides an online monitoring and fault prediction system for batteries, which has the following advantages: 1. By using the edge sensing and acquisition module, multi-dimensional operating parameters are synchronously collected and preprocessed. Combined with the "edge-cloud collaboration" architecture, the edge is responsible for real-time data collection and lightweight processing, while the cloud undertakes in-depth analysis tasks. This not only ensures the real-time nature of data collection but also resolves the contradiction between accuracy and latency in a single processing mode. At the same time, only feature data is uploaded, effectively reducing the transmission load and adapting to the monitoring needs of continuous operation scenarios.

[0017] 2. The hierarchical fusion diagnostic module adopts a multi-sub-network architecture to conduct health status assessment, anomaly identification and risk inference in parallel. It introduces a cross-sub-network knowledge transfer mechanism and fault coupling effect analysis. Compared with traditional single-dimensional prediction methods, it can more accurately quantify battery health status and potential fault risks, predict fault evolution trends in advance, reserve sufficient time for operation and maintenance intervention, and significantly reduce the fault incidence rate and expand risks.

[0018] 3. The adaptive prediction and decision-making module generates maintenance strategies at multiple time scales based on multi-dimensional evaluation results, combined with scenario constraints and historical data. It adapts to different operating conditions through dynamic threshold adjustments, and supports tiered strategy push and manual intervention to avoid resource waste caused by unified operation and maintenance. The closed-loop optimization mechanism, combined with digital twin technology, can dynamically update model parameters based on actual operation feedback, verify the effectiveness of maintenance strategies, form a full-link adaptive closed loop, and continuously improve the adaptability and effectiveness of strategies.

[0019] 4. Through scenario-based adaptation design, the system can optimize the data acquisition mode, risk assessment focus and maintenance strategy priority for different scenarios such as energy storage power stations, electric vehicle swapping stations and backup power for communication base stations. At the same time, the digital twin engine supports parallel simulation of multiple scenarios, and can achieve cross-scenario deployment without major adjustments to the hardware structure, adapting to the differentiated needs of different industries for battery operation and maintenance.

[0020] 5. By constructing a full physical field digital twin and linking data across the entire chain, the full life cycle traceability and visual management of battery operation status can be realized, accurately controlling the aging pattern and performance degradation trend of batteries. This can extend battery life through early maintenance and reduce ineffective operation and maintenance costs through optimized strategies, thus balancing safety, reliability and economy.

[0021] In summary, this invention, through an "edge-cloud collaborative" architecture and multi-dimensional perception and acquisition, ensures comprehensive and real-time monitoring while reducing transmission load, resolving the contradictions of traditional monitoring methods that rely on single parameters and processing modes. It enhances fault prediction accuracy and proactively manages risks by leveraging a layered fusion diagnostic module. Combining adaptive decision-making, closed-loop optimization, and digital twin technology, it achieves personalized dynamic optimization of maintenance strategies and a closed-loop end-to-end system. With scenario-based adaptation, it possesses broad applicability and can be deployed across scenarios without significant hardware modifications. Ultimately, through full lifecycle management, it extends battery life, improves operational efficiency, reduces overall costs, and balances system security, reliability, and economy. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the principle of an online battery monitoring and fault prediction system according to the present invention; Figure 2 This is a block diagram illustrating the principle of the closed-loop optimization and digital twin module described in this invention. Figure 3 This is a block diagram illustrating the principle of the hierarchical fusion diagnostic module described in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figures 1-3 As shown, the present invention provides a technical solution: an online monitoring and fault prediction system for batteries, which adopts an edge-cloud collaborative architecture, including the following modules and an end-to-end closed-loop collaborative process. Each module realizes real-time linkage of data and instructions through a cross-layer feature interaction protocol. Specifically, it includes: a sensing and acquisition module, a hierarchical fusion diagnostic module, an adaptive prediction and decision-making module, and a closed-loop optimization and digital twin module. The sensing and acquisition module is deployed at the edge to acquire multi-dimensional operating parameters of individual cells and battery clusters in the battery pack in real time, forming a time-series observation sequence after anti-interference preprocessing. The hierarchical fusion diagnostic module adopts a lightweight edge computing and cloud-based deep analysis mode. Its input is connected to the output of the sensing and acquisition module, and it performs health status assessment, abnormal state identification, and risk state prediction in parallel based on the time-series observation sequence, outputting a fused health status vector, a risk status vector, and a comprehensive fault risk level. The adaptive prediction and decision-making module is deployed in the cloud, and its input is... The output of the layered fusion diagnostic module is connected to dynamically generate maintenance strategy suggestions and early warnings for multiple time scales by combining health status vectors, risk status vectors, historical fault data, and scenario constraints. It supports hierarchical push of strategies and manual intervention interfaces. The closed-loop optimization and digital twin module is deployed across the edge and cloud, and communicates with the perception and acquisition module, the layered fusion diagnostic module, and the adaptive prediction and decision-making module. Based on actual operation feedback data and twin simulation results, it realizes online incremental updates of diagnostic and prediction models. Through the constructed full-physical field digital twin, it completes strategy simulation verification and effect prediction, forming a full-link adaptive closed loop of acquisition, diagnosis, decision-making, and optimization.

[0025] More specifically, the sensing and acquisition module adopts a three-level architecture of distributed acquisition, centralized synchronization, and edge preprocessing. It has the ability to adaptively adjust the sampling frequency and acquire multi-dimensional operating parameters, including: electrical parameters, operating environment parameters, and status identification parameters. The electrical parameters include at least the individual cell voltage (sampling accuracy ±0.001V), total battery cluster voltage, charging and discharging current (sampling frequency 10-100Hz, which can be dynamically adjusted according to operating conditions), dynamic internal resistance (measured using high-frequency impedance spectroscopy), and individual cell and battery cluster temperature (sampling resolution 0.1℃). The operating environment parameters include at least the ambient temperature, relative humidity, vibration intensity (range 0-50g, supporting multi-directional acquisition), and ambient dust concentration. The status identification parameters include at least the cumulative number of cycles, the current state of charge (SOC) (estimation error ≤3%), and the battery aging stage label. After the collected data is synchronized with timestamps (synchronization accuracy ≤1ms) and preprocessed for edges (wavelet threshold denoising + outlier removal), a length of [length missing] is formed. ( Time series observation sequences (with adaptively adjustable window length) ,in For single cell batteries At time step The single-step multi-dimensional fusion data is used, and only feature data is uploaded at the edge to reduce the transmission bandwidth usage.

[0026] The sensing and acquisition module adopts a three-tier architecture of distributed acquisition, centralized synchronization, and edge preprocessing. Combined with adaptive sampling frequency adjustment and multi-dimensional parameter acquisition design, it fundamentally solves the core pain points of traditional monitoring systems: single parameters, asynchronous data, heavy transmission load, and excessive noise interference. The distributed acquisition mode achieves comprehensive coverage of the operating status of individual batteries and battery clusters, avoiding misjudgments caused by missing local parameters. Centralized synchronization (synchronization accuracy ≤1ms) ensures the temporal consistency of multi-dimensional data, providing an accurate data foundation for subsequent health status quantification and risk projection. Pre-processing at the edge, including wavelet threshold denoising and outlier removal, effectively filters environmental interference and acquisition errors. This enhances data purity; the adaptive sampling frequency (10-100Hz) can be dynamically adjusted according to charging and discharging conditions, increasing sampling density to ensure data integrity under high dynamic conditions and reducing frequency to save energy under static conditions; multi-dimensional parameters (electrical, environmental, and status indicators) comprehensively capture the battery's own performance and external operating conditions, breaking the limitation that a single electrical parameter cannot reflect the full picture of battery health; the design of uploading only feature data at the edge significantly reduces the bandwidth usage of cloud transmission, while the time-series observation sequence with a length ≥10 and adaptive adjustment provides sufficient historical data support for the hierarchical fusion diagnostic module, ensuring the accuracy and real-time performance of health status assessment, anomaly identification, and risk prediction.

[0027] In this embodiment, for example in a backup power supply scenario for a communication base station, the three-level architecture of the sensing and acquisition module is specifically implemented as follows: a distributed deployment of 16 individual cell voltage acquisition channels, 2 battery cluster total voltage acquisition channels, 1 charge / discharge current acquisition channel, 8 temperature acquisition channels (4 for individual cells and 4 for battery clusters), 1 dynamic internal resistance acquisition module (using 1kHz high-frequency impedance spectroscopy), and independent acquisition units for ambient temperature, relative humidity, vibration intensity (range 0-50g, X / Y / Z direction acquisition), and ambient dust concentration. Simultaneously, a charge / discharge cycle statistics module records the cumulative number of cycles, and an extended Kalman filter algorithm is used to estimate the current load. The State of Charge (SOC) is calculated (ensuring an estimation error ≤3%), and aging stage labels are assigned based on the cumulative number of cycles (≤500 cycles for initial stage, 501-1500 cycles for intermediate stage, and >1500 cycles for final stage). The sampling frequency is adaptively adjusted according to the base station backup power supply's operating status, using a 50Hz sampling frequency during charging and discharging phases, and automatically dropping to 1Hz during the idle phase to achieve low-power operation (power consumption ≤5W). All collected data is time-aligned to ≤1ms using a timestamp synchronization module, and then preprocessed by wavelet threshold denoising (3 decomposition layers, threshold coefficient 0.8) and 3σ criterion outlier removal using an edge-embedded processor, ultimately forming the length... Time series observation sequence ( It includes 12-dimensional fused data such as unit voltage, dynamic internal resistance, temperature, ambient humidity, and SOC. At the edge, only 8 types of feature parameters such as mean, variance, and peak value of the sequence are extracted and uploaded to the cloud. Compared with the original data transmission, the bandwidth usage is reduced by more than 75%, which not only ensures the data transmission efficiency, but also provides high-value input data for the subsequent hierarchical fusion diagnostic module.

[0028] More specifically, the hierarchical fusion diagnostic module adopts a three-layer sub-network architecture of parameter sharing, task separation, and feature distillation. Each layer achieves information linkage through feature interaction channels and introduces a cross-sub-network knowledge transfer mechanism to improve diagnostic accuracy. Specifically, it includes: a health status quantification sub-network, an abnormal pattern matching sub-network, and a risk status inference sub-network. The health state quantification subnetwork, based on time-series electrical parameters and state identification parameters, employs a hybrid model combining extended Kalman filtering, attention mechanism, and particle filter correction to calculate the health state vector for each individual cell. , Includes capacity decay Internal resistance growth rate and energy efficiency The three-dimensional comprehensive vector supports model parameter adaptation for batteries with different chemical systems; the anomaly pattern matching subnetwork is based on real-time electrical and environmental parameters, and performs fast matching through a dynamically updated anomaly pattern library (combined with incremental learning to expand templates in real time). It uses a cosine similarity and Mahalanobis distance fusion algorithm to calculate the matching degree, accurately identify the anomaly types and confidence levels (value range [0,1]) related to overvoltage, undervoltage, overtemperature, internal resistance mutation, and insulation degradation, and outputs the temporal characteristics and spatial location of the anomaly occurrence; the risk state inference subnetwork is based on the current Based on anomaly identification results, historical fault sequences, and environmental trend data, and using an improved gated cyclic unit time-series prediction model, the future preset time window is extrapolated. Risk state vector (supports custom configuration) , This is used to quantify the potential risk intensity and coupling effects of four core faults: lithium plating, thermal runaway, capacity drop, and tab corrosion.

[0029] The hierarchical fusion diagnostic module adopts a three-layer sub-network architecture of parameter sharing, task separation, and feature distillation. Combined with cross-sub-network knowledge transfer mechanism and feature interaction channel, it breaks through the limitations of traditional single diagnostic models from three aspects: diagnostic dimension, accuracy, and foresight, and builds a full-chain accurate diagnostic system: the parameter sharing design enables the reuse of basic features between sub-networks, reduces redundant calculations and ensures feature consistency, while the feature distillation mechanism strengthens the retention of key information and improves model lightweighting and diagnostic efficiency.

[0030] The task separation mode allows the three sub-networks to focus on the core tasks of health quantification, anomaly identification, and risk extrapolation, avoiding diagnostic bias caused by task coupling. At the same time, the cross-sub-network knowledge transfer mechanism realizes information linkage through interactive channels, feeds back the results of health status quantification to anomaly pattern matching to improve identification accuracy, and integrates anomaly identification features into risk extrapolation to strengthen trend prediction capabilities.

[0031] The hybrid model of the health state quantification subnetwork can be adapted to batteries with different chemical systems, achieving... The precise quantization of three-dimensional vectors solves the problem of the one-sidedness of traditional single-parameter assessment of health status; the anomaly pattern matching subnetwork relies on incremental learning to dynamically update the pattern library, combined with a fusion similarity algorithm, to quickly identify multiple types of anomalies and locate spatiotemporal features, overcoming the pain point that traditional fixed templates cannot adapt to new anomalies; the risk state inference subnetwork, through an improved GRU model and a custom time window design, quantifies the potential risks and coupling effects of four types of core faults, predicts the fault evolution trend 1-72 hours in advance, and provides sufficient buffer time for subsequent maintenance decisions. Overall, it realizes a closed-loop diagnosis of accurate battery status assessment, rapid anomaly identification, and early prediction, greatly improving the scientificity and foresight of battery operation and maintenance.

[0032] In this embodiment, for example in a communication base station backup power supply scenario, the hierarchical fusion diagnostic module is deployed based on the feature data uploaded by the sensing and acquisition module. The three-layer sub-networks share basic feature vectors such as voltage, temperature, and SOC, and complete information linkage every 100ms through a dedicated feature interaction channel. The knowledge transfer mechanism weight is set to 0.3 (balancing the independence and linkage of sub-networks). The health status quantification sub-network adopts a hybrid process of extended Kalman filter initial estimation, attention mechanism to strengthen the internal resistance and capacity feature weights, and particle filter to correct nonlinear errors, adapting to the base station lead-acid battery system and outputting a three-dimensional health status vector. ,in Calculated by combining the ratio of real-time capacity to rated capacity with an aging correction factor. Based on the comparison between dynamic internal resistance monitoring data and initial internal resistance, the results are obtained. After correcting for the impact of charge / discharge rate, the error is controlled within ±2%. The abnormal pattern matching subnetwork constructs an initial abnormal template library (containing 200 templates for each of 5 core abnormalities). Incremental learning is used to update the template library every 50 actual abnormal data points. The matching degree is calculated by fusing cosine similarity (weight 0.4) and Mahalanobis distance (weight 0.6), with an identification accuracy ≥95%. When a single cell voltage exceeds 2.4V (overvoltage), the confidence score of 0.92, the timestamp of the occurrence, and the spatial location of the corresponding 8th cell are output. The risk state inference subnetwork adopts an improved GRU model (64 hidden layer neurons, dropout coefficient 0.2) with a preset time window. Enter the current Based on the anomaly identification results and the historical fault sequences of the past 72 hours, a risk state vector is derived and output. When a certain single cell battery When a slight over-temperature anomaly is detected, output This indicates a risk of thermal runaway within 24 hours. ) and the risk of sudden capacity drop ( There is a weak coupling effect, which provides targeted early warning for base station operation and maintenance personnel.

[0033] More specifically, health status vector The calculation process incorporates an adaptive mechanism for the battery aging stage, as detailed below: The first step involves using the extended Kalman filter algorithm to initially estimate the real-time capacity based on the voltage and current sequence within a sliding time window (window size 5-10 charge-discharge cycles, dynamically adjusted according to the number of cycles). and dynamic internal resistance Then, a secondary correction is performed using a particle filter algorithm to reduce nonlinear errors; The second step involves calculating using a stage-adaptive coupling model. The core component, the formula is: ; ; in, The rated capacity of the battery (unit: Ah) is the factory-calibrated value. This is the initial internal resistance of the battery (unit: mΩ), which is the factory rated value; The average temperature rise within the sliding window (unit: °C) is the difference between the highest and lowest temperatures within the window. This is a reference temperature (unit: °C), with a default value of 25 °C. The cumulative number of battery cycles (unit: times) is obtained from the statistics of charge and discharge cycles. The capacity decay-temperature coupling coefficient (range 0.01-0.05) is related to the battery chemistry system and is calibrated through offline experiments. The temperature effect index (range 1.2-1.8) characterizes the nonlinear amplification effect of temperature rise on capacity decay. The internal resistance growth-cycle number coupling coefficient (range 0.08-0.15) is related to the battery material characteristics and is calibrated through offline experiments. As a correction factor for the aging stage, according to Values ​​within the specified interval: Initial Next, take 0.8-0.9, mid-term. Next, take 1.0, final stage Next, a value of 1.1-1.2 is used to compensate for model biases at different aging stages; The third step is to calculate the energy efficiency of the most recent complete charge-discharge cycle. , Total discharge energy, in Wh; The total charging energy, in Wh, is introduced with a charge / discharge rate correction term. , This refers to the charging rate. To optimize the calculation accuracy for discharge rate, the final result is... .

[0034] The hierarchical fusion diagnostic module adopts a three-layer sub-network architecture of parameter sharing, task separation, and feature distillation. Combined with cross-sub-network knowledge transfer mechanisms and feature interaction channels, it breaks through the limitations of traditional single diagnostic models in terms of diagnostic dimensions, accuracy, and forward-looking capabilities, constructing a full-chain precision diagnostic system. Among these features, the health status vector... The adaptive mechanism for battery aging stages introduced by the calculation significantly improves the accuracy of quantifying the health status of batteries at different life cycles, and solves the pain point of evaluation bias caused by the traditional model ignoring the differences in aging stages.

[0035] Precise quantification is achieved through a three-step progressive process: a sliding time window dynamically adapts to the charge and discharge cycle, combined with preliminary estimation using extended Kalman filtering and secondary correction using particle filtering, effectively reducing nonlinear errors in the voltage and current sequences and ensuring... and Reliability of basic data; Aging stage correction factor incorporated into the stage adaptive coupling model. To address the differences in battery characteristics at the initial, intermediate, and final stages, a dynamic compensation model for biases is implemented, coupled with offline calibration coefficients adapted to the chemical system and material properties, thus achieving... , Precise matching with the actual aging state of the battery; charge / discharge rate correction item. Then it was optimized Calculation accuracy is improved to avoid interference from rate differences in energy efficiency assessment.

[0036] The overall process not only achieves the scientific quantification of the three-dimensional components of health status, but also adapts to changes throughout the battery's life cycle through an adaptive mechanism, providing highly accurate basic health data for subsequent anomaly identification and risk projection, and further strengthening the core diagnostic capabilities of the hierarchical fusion diagnostic module.

[0037] In this embodiment, for example in a communication base station backup power supply scenario, the hierarchical fusion diagnostic module is deployed based on the feature data uploaded by the sensing and acquisition module. The three-layer sub-networks share basic feature vectors such as voltage, temperature, and SOC, and complete information linkage every 100ms through a dedicated feature interaction channel. The knowledge transfer mechanism weight is set to 0.3 (balancing the independence and linkage of sub-networks), where the health status vector... The calculation is performed according to the following steps to adapt to the base station lead-acid battery system: First, set the sliding time window size (dynamically adjusted according to the cumulative number of cycles). For ≤500 cycles, initially use 5 charge-discharge cycles; for 500 < ≤1500 times, take 8 periods in the middle stage. (After 1500 cycles, 10 cycles are selected at the end of the period). The time series sequences of individual cell voltage and charge / discharge current within the window are extracted. The extended Kalman filter algorithm is used to preliminarily estimate the real-time capacity. With dynamic internal resistance Then, a second correction is performed using a particle filtering algorithm (particle count set to 200, resampling threshold 0.5) to control the nonlinear error within 3%; the second step is to call the stage adaptive coupling model to calculate the core components, given the lead-acid battery. , Reference temperature calibrated through offline experiments , , If a certain single cell (Mid-term) Average temperature rise within the sliding window Calculated , The third step is to analyze the data from the most recent complete charge-discharge cycle. , Charging rate Discharge rate First calculate the correction term. , then The final output is a health state vector. .

[0038] The abnormal pattern matching subnetwork is based on this The anomaly detection threshold is optimized, and the risk state deduction subnetwork will... As a core input, it improves the accuracy of risk prediction when a single entity is detected. In the event of a slight over-temperature anomaly, output a risk state vector. It indicates the risk of weak coupling between thermal runaway and sudden capacity drop within 24 hours, providing targeted early warning for operation and maintenance personnel.

[0039] More specifically, risk state vector This is achieved by combining a time-series risk propagation model with the fault coupling effect. The model definition and parameter relationships are as follows: ; in: For the current time step The risk state vector (dimension 4×1). Meaning of each component: For lithium plating risk, For the risk of thermal runaway, To mitigate the risk of a sudden drop in capacity, To represent the risk of tab corrosion, the values ​​are all in the range of [0,1], with larger values ​​indicating higher risk. Changes in health status (dimension 3×1). It is predicted by combining the current operating conditions (charge / discharge rate, ambient temperature, and humidity); This is an abnormal mode vector (dimension 5×1), where each element represents the confidence level of overvoltage, undervoltage, overtemperature, sudden change in internal resistance, and decrease in insulation, with values ​​ranging from [0,1]. The risk decay matrix (dimension 4×4) has diagonal elements with values ​​of 0.85-0.95, representing the natural decay trend of risk over time, and off-diagonal elements with values ​​of 0.01-0.05, representing the weak coupling effect between different risks. The health-risk coupling matrix (4×3 dimension) has non-zero elements ranging from 0.1 to 0.3, representing the contribution of deteriorating health status to different risks. The impact coefficient of corresponding capacity decay on lithium plating risk; This is an anomaly-risk trigger matrix (4×5 dimensions), with elements ranging from 0 to 0.6, representing the trigger strength of different anomalies for specific risks, such as... The trigger coefficient for the risk of thermal runaway corresponding to overtemperature anomalies; The risk-anomaly interaction matrix (4×4 dimension) has elements ranging from 0.05 to 0.2, representing the synergistic amplification effect between anomalies and existing risks. The Hadamard product is used to perform element-wise multiplication of corresponding vector elements; the parameters of the above matrix are obtained through training with historical fault data and are updated periodically by the closed-loop optimization module.

[0040] The hierarchical fusion diagnostic module adopts a three-layer sub-network architecture of parameter sharing, task separation, and feature distillation. Combined with cross-sub-network knowledge transfer mechanisms and feature interaction channels, it breaks through the limitations of traditional single diagnostic models in terms of diagnostic dimensions, accuracy, and forward-looking capabilities, constructing a full-chain precision diagnostic system. Among these features, the health status vector... The computationally introduced adaptive mechanism for battery aging stages significantly improves the accuracy of quantifying battery health status at different lifecycle stages, overcoming the assessment bias caused by traditional models ignoring differences in aging stages, and addressing the risk state vector... The time-series risk propagation model adopted further realizes the accurate extrapolation of failure risk and the quantification of coupling effects, solving the core pain point of traditional risk prediction that only assesses the failure probability and ignores the time-series propagation of risk and the synergistic effect of multiple failures.

[0041] This model integrates multi-dimensional inputs through matrix operations and Hadamard products to construct a full-link risk evolution logic encompassing current risk, health changes, anomaly triggers, and interactive amplification: a risk self-decay matrix. Precisely depicting the natural decay pattern of risk and its weak coupling characteristics across risks, the health-risk coupling matrix With the anomaly-risk trigger matrix The direct impacts of health deterioration and abnormal events on risk are quantified separately, forming a risk-anomaly interaction matrix. The Hadamard product highlights the synergistic amplification effect of anomalies and existing risks, enabling dynamic projection of four core fault risks. Simultaneously, the matrix parameters are trained based on historical fault data and periodically updated by a closed-loop optimization module, ensuring the model adapts to different operating conditions and battery aging states. This provides forward-looking and accurate risk data support for the subsequent adaptive decision-making module, extending the fault warning window.

[0042] In this embodiment, for example in a communication base station backup power supply scenario, the hierarchical fusion diagnostic module is deployed based on the feature data uploaded by the sensing and acquisition module. The three-layer sub-networks share basic feature vectors such as voltage, temperature, and SOC, and complete information linkage every 100ms through a dedicated feature interaction channel. The knowledge transfer mechanism weight is set to 0.3 (balancing the independence and linkage of sub-networks), where the health status vector... Calculate the risk state vector according to the established three-step process (adapted to base station lead-acid battery system). Then, through the time-series risk propagation model, a preset time window is used for deduction. The specific implementation is as follows: Given the current time step risk vector The change in health status was predicted. (Capacity decay and internal resistance growth rate increased slightly, while energy efficiency decreased slightly), abnormal mode vector (Only a slight over-temperature anomaly was detected, confidence level 0.4); the matrix parameters were determined through training with historical fault data of base station lead-acid batteries. ; ; ; .

[0043] Substitute into the model for calculation: First calculate , , , , ; ultimately obtained It is evident that abnormal temperature and deteriorating health synergistically amplify the risk of thermal runaway. The value of the matrix is ​​increased to 0.548, and it has a coupled impact on the risk of sudden capacity drop. The model accurately outputs the risk evolution results, providing a basis for operation and maintenance decisions. The matrix parameters are updated once every 1,000 operational data points accumulated by the closed-loop optimization module.

[0044] More specifically, the adaptive prediction and decision-making module employs mechanisms of risk grading, scenario adaptation, and multi-objective optimization, maintains a multi-scale policy library, dynamically adjusts thresholds, and supports personalized policy customization. The strategy library includes three categories: long-term predictive maintenance plans (cycle 1-3 months, focusing on lifespan extension), mid-term early warning and online balancing strategies (cycle 1-7 days, focusing on performance optimization), and emergency intervention strategies (immediate response, focusing on security protection). Each strategy is associated with a cost-benefit evaluation model; the threshold dynamic adjustment rule: the system is based on the failure rate of the most recent 30 days. Predicted remaining battery life The threshold is updated using the following formula: : ; ; in, This is a low-to-medium risk threshold used to trigger medium-term early warning strategies; This is a medium-to-high risk threshold used to trigger emergency intervention strategies; The failure rate over the past 30 days (value range [0,1]) is calculated as the ratio of the number of failures to the total operating time. The remaining battery life is the predicted value (in days), derived from the health state vector. It was deduced that; Strategy selection logic: When the overall fault risk level is low, output a long-term maintenance plan (including battery equalization activation and capacity calibration recommendations); when the level is medium or low, output a long-term maintenance plan (including battery equalization activation and capacity calibration recommendations). Any element in At the same time, it outputs mid-term early warning and online balancing strategies, prioritizing power supply to core loads; the level is high or Any element in In case of an emergency, the system will output emergency intervention strategies (power reduction operation, forced cooling, disconnection suggestions) and simultaneously coordinate with the backup power switching mechanism to reduce the scope of the fault's impact.

[0045] The adaptive prediction and decision-making module relies on the core mechanisms of risk classification, scenario adaptation and multi-objective optimization to build a multi-scale strategy library and introduce dynamic threshold adjustment logic. This effectively solves the pain points of traditional maintenance decision-making strategies, such as single strategies, fixed thresholds and poor scenario adaptability, and achieves accurate matching between maintenance strategies and battery operating status and working condition requirements.

[0046] The multi-scale strategy library covers long, medium, and short-cycle maintenance needs. Combined with a cost-benefit evaluation model, it can dynamically balance the three major objectives of lifespan extension, performance optimization, and security protection, avoiding cost waste caused by excessive maintenance or security risks caused by insufficient maintenance. Dynamic thresholds are based on failure incidence rates. With remaining lifespan Real-time calibration makes risk grading more closely match actual operating conditions, solving the problem of insufficient adaptability of fixed thresholds under different failure probabilities and life stages; a clear strategy selection logic is centered on risk level and threshold triggering conditions, linking with the output of the hierarchical fusion diagnostic module mentioned above. and The data enables full-scenario coverage of low-risk predictive maintenance, medium-risk optimization and adjustment, and high-risk emergency protection. It also supports personalized strategy customization and backup power linkage, ensuring the continuity of power supply to core loads while minimizing the impact of faults and maintenance costs, forming a closed-loop collaboration for diagnosis and decision-making.

[0047] In this embodiment, in the backup power scenario of a communication base station, the adaptive prediction and decision-making module is based on the output of the hierarchical fusion diagnostic module. and health status vector The simulation data is implemented as follows: First, calculate the dynamic threshold and statistically analyze the base station backup power supply failure rate over the past 30 days. (Only two minor over-temperature anomalies occurred, with a total operating time of 720 hours), through middle , The remaining battery life was calculated. Substitute into the formula to calculate: ; .

[0048] Combination Analysis of each component , , All exceeded However, it did not exceed The risk level was determined to be medium, triggering a mid-term early warning and online balancing strategy.

[0049] The strategy execution plan is as follows: prioritize power supply to the core communication load of the base station, and activate the online power balancing algorithm (the balancing weight is configured according to the discharge state). For high-risk individual cells, the equalization cycle is adjusted to 1 hour. Cost-benefit analysis shows this strategy can reduce battery performance degradation rate by 15%, keeping maintenance costs within the monthly budget. Simultaneously, it integrates with a temperature monitoring module to enhance heat dissipation, providing feedback on the equalization effect every 2 hours. If the risk vector component rises to [a certain level], [further action is taken]. The above immediately triggers emergency intervention strategies (reduced power by 30%, forced air-cooled start-up) and switches to the backup battery pack to ensure uninterrupted base station communication.

[0050] More specifically, the online balancing strategy introduces a risk contribution weighting and dynamic balancing cycle mechanism, with dual adaptation to balancing weight calculation and operation mode, and battery aging stage: ; in, For single cell batteries The equilibrium weights (value range [0,1]) are determined by the weights; the higher the weight, the higher the equilibrium priority. For single cell batteries Real-time voltage (unit: V); The average voltage of the battery cluster (unit: V); For single cell batteries Real-time internal resistance (unit: mΩ); The average internal resistance of the battery cluster (unit: mΩ). For single cell batteries The maximum risk value in the risk state vector; For single cell batteries Capacity decay rate; For the weighting coefficients, satisfying Dual adaptation rules: Still state: ; Charging status: ; Discharge state: ; Late aging stage: additional Increase by 0.1 to enhance the balancing priority of degraded batteries; The equalization period is dynamically adjusted based on the maximum voltage difference within the cluster (0.5-2 hours), with a shorter period for larger voltage differences.

[0051] The adaptive prediction and decision-making module relies on the core mechanisms of risk classification, scenario adaptation, and multi-objective optimization to build a multi-scale strategy library and introduce dynamic threshold adjustment logic. This effectively solves the pain points of traditional maintenance decision-making strategies, such as single strategies, fixed thresholds, and poor scenario adaptability, and achieves accurate matching between maintenance strategies and battery operating status and working condition requirements.

[0052] The online balancing strategy specifically introduces a dual mechanism of risk contribution weighting and dynamic balancing cycle. At its core, it addresses issues such as priority confusion, energy redundancy, and insufficient adaptation to aging batteries caused by the one-size-fits-all approach of traditional balancing strategies through a balancing weight formula and dual adaptation rules. This significantly improves the accuracy and energy efficiency of battery cluster balancing optimization during the mid-term warning phase. The balancing weight formula systematically integrates four key indicators: voltage deviation, internal resistance deviation, maximum risk value, and capacity degradation, achieving multi-dimensional quantification of balancing requirements; weight coefficients... - Strictly adhering to the dual adaptation rules of operating mode and aging stage, focusing on voltage balancing during rest, strengthening risk guidance during charging, and emphasizing internal resistance optimization during discharging, while also providing additional enhancements for batteries in the late stages of aging. The coefficient ensures that resources are prioritized for monomers with severe performance degradation and high risk levels.

[0053] The dynamic balancing cycle is flexibly adjusted based on the maximum voltage difference within the cluster. When the voltage deviation is large, the cycle is shortened to ensure the timeliness of balancing, and when the deviation is small, the cycle is extended to reduce energy consumption. This forms a closed-loop balancing logic with precise demand quantification, dynamic priority adaptation, and flexible cycle adjustment. It not only makes up for the performance differences of individual cells and inhibits the uniform degradation of battery clusters, but also links risk states and aging characteristics, providing core support for extending the life of the battery cluster throughout its entire life cycle and enhancing the overall system's operation and maintenance optimization efficiency.

[0054] In this embodiment, for example in a communication base station backup power scenario, the adaptive prediction and decision module is based on the output of the hierarchical fusion diagnostic module. and health status vector The simulation data is implemented as follows: First, calculate the dynamic threshold and statistically analyze the base station backup power supply failure rate over the past 30 days. (Only two minor over-temperature anomalies occurred, with a total operating time of 720 hours), through middle , The remaining battery life was calculated. Substitute into the formula to calculate: ; .

[0055] Combination Analysis of each component , , All exceeded However, it did not exceed The risk level was determined to be medium, triggering a mid-term early warning and online balancing strategy.

[0056] The online balancing strategy is executed precisely using a dual-adaptation mechanism: The current base station backup power supply is in a discharging state, and two types of typical individual cells within the cluster (mid-term and late-aging) are simultaneously selected to calculate balancing weights; one is the mid-term individual cell (… ), based on the discharge state reference configuration factor Given its real-time voltage Cluster average voltage Real-time internal resistance Cluster average internal resistance Maximum risk value Capacity decay Substituting into the formula, we get: ; The second is monomers at the end of aging ( The discharge baseline coefficient is further increased by μ4 to 0.2 (corresponding to a decrease of μ1 to 0.1, with the maintenance coefficient summing to 1), and its parameters are as follows: , , , The calculation yields: ; Prioritized significantly above mid-term cells, these cells were included in the first equalization queue. Simultaneously, a maximum voltage difference of 0.05V was detected within the cluster, and the equalization cycle was set to 1.5 hours according to dynamic cycle rules. Cost-benefit analysis showed that this strategy could improve cell cluster consistency by 20%, reduce performance degradation rate by 15%, and keep maintenance costs within the monthly budget. The temperature monitoring module was linked to enhance heat dissipation, and equalization results were reported every 2 hours. If the risk vector component rises to... The above immediately triggers emergency intervention strategies (reduced power by 30%, forced air-cooled start-up) and switches to the backup battery pack to ensure uninterrupted base station communication.

[0057] More specifically, the closed-loop optimization and digital twin module integrates the entire process of data feedback, model update, simulation verification, and strategy iteration, and has the capability for virtual-real linkage calibration, specifically including: Feedback Data Pool: Stores actual fault events, maintenance operation records, twin simulation data, and corresponding preliminary diagnostic data, with a capacity of 1×10. 4 -1×10 6 The algorithm employs a priority experience replay mechanism, assigning higher sampling weights (weight coefficient 1.5-2.0) to high-risk prediction hit data and fault missed / misjudged data, while simultaneously removing abnormal interference data through data cleaning algorithms. Incremental learning unit: Employing an incremental learning strategy of freezing the bottom layer and fine-tuning the top layer, the network parameters of the hierarchical fusion diagnostic module are fine-tuned and the coupling coefficients are updated every 500-2000 new feedback data points. , , ) and risk transmission model matrix ( , , , The AdamW optimizer (learning rate 1×10⁻⁶) was used. -4 -5×10 -4 Regularization terms are introduced to prevent overfitting, and model update logs are recorded for traceability. Digital twin simulation engine: Constructs a virtual battery pack based on an electrochemical-thermal-mechanical multiphysics coupling model, through... It dynamically corrects simulation parameters based on real-time running data, supports parallel simulation of different risk scenarios and maintenance strategies, and predicts the effect, potential chain reactions and economic costs of major maintenance strategies before implementation. The simulation error is ≤5%, and the simulation results serve as an important basis for strategy output.

[0058] The closed-loop optimization and digital twin module integrates the full-process functions of data feedback, model update, simulation verification and strategy iteration to build a self-optimizing system that links the virtual and real worlds. It solves the core pain points of traditional systems, such as fixed models, lack of prediction in strategy implementation and lack of continuous performance improvement, and provides closed-loop iterative support for the entire battery monitoring and fault prediction system.

[0059] The feedback data pool employs a priority experience replay mechanism, focusing on high-value data and eliminating interfering data to ensure the quality of data sources for model updates and provide accurate input for incremental learning. The model's incremental learning unit updates parameters using a strategy of freezing the bottom layer and fine-tuning the top layer, balancing model stability and adaptability. The AdamW optimizer and regularization terms balance training effectiveness and overfitting risk, achieving dynamic adaptation of the diagnostic model to actual working conditions. The digital twin simulation engine constructs a virtual battery pack based on a multi-physics coupling model, using health state vectors... Dynamic parameter correction keeps simulation errors within 5%, allowing for early prediction of maintenance strategy effectiveness and potential risks, thus avoiding cost waste or escalation of faults caused by blind maintenance.

[0060] The three elements work together to generate actual operational data that feeds back into the model for optimization. The optimized model supports simulation predictions, and the simulation results guide the closed-loop logic of strategy iteration, continuously improving the system's diagnostic accuracy, strategy adaptability, and scientific operation and maintenance, thereby extending the value of the battery's entire life cycle.

[0061] In this embodiment, for example in a communication base station backup power supply scenario, the closed-loop optimization and digital twin modules are deployed across the edge and cloud, linking the three modules mentioned above to achieve a closed-loop process. The specific implementation is as follows: the feedback data pool is configured with a capacity of 5×10 5 The database stores actual fault events of the base station over the past 6 months (including 2 instances of over-temperature anomalies and 3 instances of internal resistance surge warnings), maintenance operation records (online balancing, capacity calibration, etc.), twin simulation data, and corresponding diagnostic data. High-risk prediction hit data and 1 historical missed data are assigned a 1.8 times sampling weight. The criteria and outlier removal algorithm cleaned the data, removing 120 invalid data entries caused by environmental interference.

[0062] The incremental learning unit of the model is set to trigger an update every 1000 new feedback data points. For the hierarchical fusion diagnostic module, the parameters of the bottom feature extraction layer of the sub-network are frozen, and only the top parameters of the health status quantification and risk propagation model are fine-tuned. The capacity decay-temperature coupling coefficient ξ is updated to 0.032, and the internal resistance growth-cycle number coupling coefficient ζ is updated to 0.125. The risk self-decay matrix is ​​corrected simultaneously. The diagonal elements are [0.91, 0.89, 0.92, 0.93], and the AdamW optimizer (learning rate 3×10⁻⁶) is used. -4 The training was performed with L2 regularization (coefficient 0.001), and the update log fully recorded the parameter changes and training results.

[0063] The digital twin simulation engine constructs a virtual battery cluster based on the electrochemical-thermal-mechanical multiphysics model of lead-acid batteries, inputting the single cell from the previous section. And real-time operating parameters, correcting the simulation capacity. Internal resistance for simulation The simulation error was controlled within 4.2%. For the previously triggered mid-term online balancing strategy, the effects of two balancing cycles (1.5h and 1h) were verified in parallel using a simulation engine. The results showed that the 1.5h cycle improved battery cluster consistency by 20% and reduced energy consumption by 12%, which was superior to the 1h cycle, thus identifying it as the optimal strategy. Simultaneously, the simulation predicted the risk of thermal runaway within 24 hours after the strategy was implemented. It can be reduced to below 0.35 without any cascading risk. Based on this, the strategy execution instructions are output, and the strategy execution effect is subsequently fed back to the data pool to provide a basis for the next round of model updates and strategy iterations.

[0064] More specifically, the digital twin simulation engine adopts a virtual-real synchronization-parameter self-calibration process to achieve dynamic optimization of simulation accuracy. The specific parameter correction process is as follows: based on Capacity decay in Internal resistance growth rate and energy efficiency The core parameters of the electrochemical model are corrected using the following formula: ; ; ; in, The simulation capacity (unit: Ah) represents the actual usable capacity of the virtual battery. The internal resistance (unit: mΩ) is used for simulation to characterize the conduction loss characteristics of the virtual battery. The amount of electrode reaction charge (unit: Ah) used in the simulation is used to characterize the reactivity of the electrode active materials in the virtual battery. The rated reaction capacity of the electrode (unit: Ah) is the battery design parameter; thermal model parameter correction: through... Adjusting the heat dissipation coefficient during battery aging: ; in, Thermal efficiency coefficient for simulation (unit: W / (m²)) 2 ·K)); Reference thermal conductivity (unit: W / (m²)) 2 ·K), determined by the battery casing material and structure; Mechanical stress model parameter correction: through internal resistance growth rate Corrected electrode expansion coefficient , The baseline expansion coefficient is used to ensure consistency between the simulated and physical battery pack's electro-thermal-mechanical coupling behavior; Simultaneous virtual-real calibration: Every hour, the deviation between the real-time data from the sensing and acquisition module and the twin simulation data is calculated. When the deviation of key parameters related to voltage and temperature exceeds 3%, a second parameter correction is triggered to ensure simulation accuracy.

[0065] The closed-loop optimization and digital twin module integrates the full-process functions of data feedback, model update, simulation verification and strategy iteration to build a self-optimizing system that links the virtual and real worlds. It solves the core pain points of traditional systems, such as fixed models, lack of prediction in strategy implementation and lack of continuous performance improvement, and provides closed-loop iterative support for the entire battery monitoring and fault prediction system.

[0066] The digital twin simulation engine employs a virtual-real synchronization-parameter self-calibration process. By accurately correcting the parameters of the multi-physics model in different dimensions and verifying the virtual-real deviation in real time, it solves the core problems of fixed parameters, disconnection from physical battery behavior, and difficulty in maintaining accuracy in traditional simulation models. This achieves dynamic optimization of simulation accuracy and ensures consistency of electro-thermal-mechanical coupling behavior.

[0067] This process uses a health status vector Based on this core principle, the key parameters of the three major models—electrochemical, thermal, and mechanical—are specifically modified. Quantitative formulas deeply bind the battery aging state and temperature rise characteristics to the model parameters, ensuring that the virtual twin accurately replicates the actual operating characteristics of the physical battery. An hourly virtual-to-real synchronous calibration mechanism triggers secondary corrections with a 3% deviation threshold, forming a dynamic closed loop of initial parameter correction, virtual-to-real comparison, deviation verification, and precise correction, keeping simulation errors stably controlled within 5%. This approach provides highly reliable support for the pre-simulation verification of maintenance strategies and, by accurately replicating battery coupling behavior, helps predict potential cascading risks, further enhancing the optimization efficiency of the virtual-to-real linkage of the module.

[0068] In this embodiment, in the backup power supply scenario for communication base stations, the closed-loop optimization and digital twin modules are deployed across the edge and cloud, linking the three modules mentioned above to achieve a closed-loop process. The digital twin simulation engine is implemented according to a virtual-real synchronization-parameter self-calibration process, specifically as follows: the feedback data pool is configured with a capacity of 5×10... 5 The system stores nearly six months of actual fault events, maintenance operation records, twin simulation data, and corresponding diagnostic data for the base station. High-value data is assigned a 1.8x sampling weight and, after cleaning, provides input for model updates. The model incremental learning unit updates the diagnostic model parameters every 1000 data points. The digital twin simulation engine constructs a virtual battery cluster based on the electrochemical-thermal-mechanical multiphysics model of lead-acid batteries, inputting the individual cells mentioned earlier. and real-time operating parameters ( , ), Parameters are corrected according to the formula by dimension: known , , , , ; The calculated simulation capacity is: ; Internal resistance for simulation: ; Electrode reaction charge in simulation: ; Heat dissipation coefficient for simulation: ; Electrode expansion coefficient for simulation: .

[0069] A virtual-real synchronization calibration mechanism was initiated, extracting real-time voltage and temperature data from the sensing and acquisition module every hour and comparing them with the twin simulation data to calculate the deviation. In this scenario, the initial comparison showed a voltage deviation of 1.8% and a temperature deviation of 2.1%, both below the 3% threshold, requiring no secondary correction, and the simulation error was controlled at 4.2%. Regarding the previously mentioned mid-term online balancing strategy, parallel simulations were performed for 1.5-hour and 1-hour cycles, determining 1.5 hours as the optimal strategy and predicting the risk of thermal runaway within 24 hours after implementation. If the temperature drops below 0.35, there is no risk of cascading effects. After the strategy is implemented, the actual results will be fed back to the data pool, providing a basis for the next iteration. If a subsequent hourly calibration reveals a temperature deviation of 3.5%, a second correction will be immediately triggered, and adjustments will be made. Up to 14.8W / (m 2 •K), to control the deviation within 2.8%.

[0070] More specifically, the battery online monitoring and fault prediction system is adapted and expanded for different application scenarios, supporting multi-scenario compatible deployment: In the energy storage power station scenario: the sensing and acquisition module supplements the insulation resistance of the battery cluster and the sealing performance of the cabinet; the risk state inference sub-network strengthens the coupled risk assessment of thermal runaway and insulation failure; the digital twin engine adds the functions of charge-discharge cycle life simulation and cluster collaborative control simulation; and the maintenance strategy prioritizes the needs of grid dispatch. Electric vehicle battery swapping station scenario: The sampling frequency is increased to 100Hz, a new battery unique identification (RFID) and full life cycle traceability function are added, the emergency intervention strategy adds battery swapping priority recommendation (combining risk level and vehicle scheduling needs), the model incremental learning unit shortens the update cycle to accumulate 500 feedback data, and supports the balance between battery swapping efficiency and battery safety. Backup power scenario for communication base stations: Optimize the low-power acquisition mode in the static state (sampling frequency reduced to 1Hz, power consumption ≤5W), add battery activation suggestions and backup power switching linkage strategy to the long-term maintenance plan, supplement the abnormal mode library with low power self-discharge and float charge voltage drift abnormal templates, and prioritize the policy output to ensure the power supply continuity of the base station core equipment.

[0071] Addressing the core needs and operational characteristics of energy storage power stations, electric vehicle battery swapping stations, and backup power supplies for communication base stations, the system undergoes customized optimization across four dimensions: perception and data acquisition, risk assessment, digital twins, and maintenance strategies. For energy storage power station scenarios, it enhances cluster security and grid coordination, supplements key parameters such as insulation resistance, and deepens coupled risk assessment to meet the safety and scheduling requirements of large-scale cluster operation. For electric vehicle battery swapping stations, it focuses on balancing efficiency and safety in high-frequency battery swapping scenarios, increasing sampling frequency, adding full lifecycle traceability functions, and adapting to varying operating conditions by shortening model update cycles. For communication base station backup power supplies, it optimizes low-power operation and power supply continuity, supplementing scenario-specific anomaly templates and linkage strategies to meet the needs of long-term standby and core load priority for base stations. This adaptive design retains the system's full-link core capabilities of monitoring, diagnosis, decision-making, and optimization, while achieving deep integration with specific scenarios through differentiated functional expansion. It allows for cross-scenario deployment without significant hardware modifications, reducing scenario-based implementation costs, while simultaneously improving the accuracy, safety, and economy of operation and maintenance in various scenarios, thus broadening the system's application boundaries.

[0072] This system offers customized deployment options for three typical scenarios, with specific adaptation solutions as follows: In the energy storage power station scenario: The sensing and acquisition module, based on basic parameter acquisition, adds two channels for battery cluster insulation resistance acquisition (measurement range 0-500MΩ, measurement accuracy ±2%) and one channel for cabinet sealing monitoring (detection dew point temperature range -40℃~60℃, gas concentration resolution 0.01%VOL); the risk state deduction sub-network optimizes the risk propagation model matrix, increasing the coupling influence coefficient of thermal runaway and insulation failure from 0.03 to 0.3, strengthening the collaborative risk assessment of the two; the digital twin engine adds charge-discharge cycle life simulation (prediction accuracy ±3%) and cluster collaborative control simulation functions, supporting the simulation of charge-discharge power distribution for 10 parallel battery clusters; the maintenance strategy is linked with the grid dispatch system, pre-setting off-peak hours (00:00-08:00) for battery equalization activation and capacity calibration, and prioritizing power supply stability during peak hours (09:00-23:00) to avoid conflicts between maintenance operations and grid load peaks.

[0073] In the electric vehicle battery swapping station scenario: the sampling frequency of the sensing and acquisition module is fixedly increased to 100Hz, and an RFID battery identification unit is added (identification distance 0-5cm, identification rate 100%). A traceability platform covering the entire process of production, charging and discharging, battery swapping, and maintenance is built, linking core data such as battery number, cycle count, and health status. The emergency intervention strategy is embedded with a battery swapping priority recommendation algorithm, which generates a recommendation sequence based on risk level (weight 0.6) and vehicle scheduling needs (weight 0.4). High-risk batteries (any element of the risk vector ≥ 0.5) are given priority in the battery swapping queue. The model incremental learning unit shortens the update cycle to accumulate 500 feedback data points, quickly adapting to the high-frequency battery swapping conditions of more than 200 swapping operations per day at the battery swapping station, balancing battery swapping efficiency and battery safety.

[0074] For communication base station backup power scenarios: Optimize the low-power acquisition mode in the static state, reduce the sampling frequency to 1Hz, and control the power consumption of the core acquisition unit to 4.8W (≤5W); The long-term maintenance plan adds quarterly battery activation suggestions (adopting a stepped process of "0.1C constant current charging - static 1h - 0.05C constant current discharging") and a backup power switching linkage strategy. When the risk level of the main battery cluster rises to medium-high or the core parameter deviation exceeds 5%, the backup lithium battery pack is automatically triggered for seamless switching (switching delay ≤5ms); The abnormal mode library is supplemented with two types of scenario-specific abnormal templates: low power self-discharge (threshold ≤0.05Ah / 24h) and float charge voltage drift (deviation ±0.03V), expanding the feature templates by a total of 300; The strategy output always prioritizes the power supply of the base station's core communication equipment (main equipment, transmission equipment), and the power supply adjustment of non-core equipment (air conditioning auxiliary equipment) is delayed to ensure continuous and uninterrupted communication links.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A battery online monitoring and fault prediction system, characterized in that, It adopts an edge-cloud collaborative architecture, including the following modules and end-to-end closed-loop collaborative process. Each module realizes real-time linkage of data and instructions through cross-layer feature interaction protocol. The sensing and acquisition module, deployed at the edge, is used to acquire multi-dimensional operating parameters of individual cells and battery clusters in the battery pack in real time, and form a time-series observation sequence after anti-interference preprocessing. The hierarchical fusion diagnostic module adopts a lightweight edge computing and cloud-based deep analysis mode. Its input end is connected to the output end of the sensing and acquisition module. Based on the time-series observation sequence, it performs health status assessment, abnormal status identification and risk status prediction in parallel, and outputs a fused health status vector, risk status vector and comprehensive fault risk level. The adaptive prediction and decision-making module is deployed in the cloud. Its input end is connected to the output end of the hierarchical fusion diagnosis module. It combines health status vector, risk status vector, historical fault data and scenario constraints to dynamically generate maintenance strategy suggestions and early fault warnings at multiple time scales. It supports hierarchical push of strategies and manual intervention interface. The closed-loop optimization and digital twin module is deployed across the edge and cloud, and communicates with the perception and acquisition module, the hierarchical fusion diagnosis module and the adaptive prediction and decision module respectively. Based on the actual operation feedback data and the twin simulation results, it realizes online incremental updates of the diagnostic model and the prediction model. Through the constructed full physical field digital twin, it completes strategy simulation verification and effect prediction, forming a full-link adaptive closed loop of acquisition, diagnosis, decision and optimization. The health status vector The calculation process incorporates an adaptive mechanism for the battery aging stage, as detailed below: The first step is to use the extended Kalman filter algorithm to initially estimate the real-time capacity based on the voltage and current sequence within the sliding time window. and dynamic internal resistance Then, a secondary correction is performed using a particle filter algorithm to reduce nonlinear errors; The second step involves calculating using a stage-adaptive coupling model. The core component, the formula is: ； ； in, This refers to the battery's rated capacity, which is the factory-calibrated value. This is the initial internal resistance of the battery, which is the factory calibration value. This represents the average temperature rise within the sliding window, which is the difference between the highest and lowest temperatures within the window. For reference temperature, the default value is 25℃; The cumulative number of battery cycles is obtained from the statistics of charge and discharge cycles. The capacity decay-temperature coupling coefficient is related to the battery chemistry system and was calibrated through offline experiments. The temperature effect index characterizes the nonlinear amplification effect of temperature rise on capacity decay. The internal resistance growth-cycle count coupling coefficient is related to the battery material properties and was calibrated through offline experiments. As a correction factor for the aging stage, according to Values ​​within the specified interval: Initial Next, take 0.8-0.9, mid-term Next, take 1.0, final stage Next, a value of 1.1-1.2 is used to compensate for model biases at different aging stages; The third step is to calculate the energy efficiency of the most recent complete charge-discharge cycle. , This represents the total energy of the discharge. To determine the total charging energy, a charge / discharge rate correction term is introduced. , This refers to the charging rate. To optimize the calculation accuracy for discharge rate, the final result is... .

2. The battery online monitoring and fault prediction system according to claim 1, characterized in that, The sensing and acquisition module adopts a three-tier architecture of distributed acquisition, centralized synchronization, and edge preprocessing, and has the ability to adaptively adjust the sampling frequency. The acquired multi-dimensional operating parameters include: Electrical parameters: including at least the individual cell voltage, total battery cluster voltage, charging and discharging current, dynamic internal resistance, and individual cell and battery cluster temperature; Operating environment parameters: at least include ambient temperature, relative humidity, vibration intensity, and ambient dust concentration; Status identification parameters: must include at least the cumulative number of cycles, current state of charge (SOC), and battery aging stage label; After the collected data is synchronized with timestamps and preprocessed at the edges, it forms a data structure with a length of [length missing]. Time series observation sequence ,in For single cell batteries At time step The single-step multi-dimensional fusion data is used, and only feature data is uploaded at the edge to reduce the transmission bandwidth usage.

3. The battery online monitoring and fault prediction system according to claim 2, characterized in that, The hierarchical fusion diagnostic module adopts a three-layer sub-network architecture of parameter sharing, task separation, and feature distillation. Each layer achieves information linkage through feature interaction channels, and a cross-sub-network knowledge transfer mechanism is introduced to improve diagnostic accuracy, specifically including: Health State Quantization Subnetwork: Based on time-series electrical parameters and state identification parameters, a hybrid model employing extended Kalman filtering, attention mechanism, and particle filter correction is used to calculate the health state vector of each individual cell. , Includes capacity decay Internal resistance growth rate and energy efficiency The three-dimensional integrated vector supports the adaptation of model parameters for batteries with different chemical systems; Anomaly pattern matching subnetwork: Based on real-time electrical and environmental parameters, it performs fast matching through a dynamically updated anomaly pattern library. It uses a cosine similarity and Mahalanobis distance fusion algorithm to calculate the matching degree, accurately identify the anomaly types and confidence levels related to overvoltage, undervoltage, overtemperature, internal resistance mutation, and insulation degradation, and outputs the temporal characteristics and spatial location of the anomaly occurrence. Risk State Deduction Subnetwork: Based on the current Based on anomaly identification results, historical fault sequences, and environmental trend data, an improved gated cyclic unit time-series prediction model is used to extrapolate the risk state vector within a preset time window. , This is used to quantify the potential risk intensity and coupling effects of four core faults: lithium plating, thermal runaway, capacity drop, and tab corrosion.

4. The battery online monitoring and fault prediction system according to claim 3, characterized in that, The risk state vector This is achieved by combining a time-series risk propagation model with the fault coupling effect. The model definition and parameter relationships are as follows: ； in: For the current time step The risk state vector, Meaning of each component: For lithium plating risk, For the risk of thermal runaway, To mitigate the risk of a sudden drop in capacity, To represent the risk of tab corrosion, the values ​​are all in the range of [0,1], with larger values ​​indicating higher risk. For changes in health status, It is obtained by joint prediction based on the current operating conditions; This is the abnormal mode vector, where each element represents the confidence level of overvoltage, undervoltage, overtemperature, sudden change in internal resistance, and decrease in insulation, with values ​​ranging from [0,1]. This is a risk decay matrix. The diagonal elements range from 0.85 to 0.95, representing the natural decay trend of risk over time. The off-diagonal elements range from 0.01 to 0.05, representing the weak coupling effect between different risks. This is a health-risk coupling matrix, where non-zero elements range from 0.1 to 0.3, representing the contribution of deteriorating health status to different risks. The impact coefficient of corresponding capacity decay on lithium plating risk; This is an anomaly-risk trigger matrix, with elements ranging from 0 to 0.6, representing the trigger strength of different anomalies for specific risks; This is a risk-anomaly interaction matrix, with elements ranging from 0.05 to 0.2, representing the synergistic amplification effect between anomalies and existing risks. The Hadamard product is used to perform element-wise multiplication of corresponding vector elements; the elements of the matrix are obtained through training on historical fault data and are periodically updated by the closed-loop optimization module.

5. The battery online monitoring and fault prediction system according to claim 4, characterized in that, The adaptive prediction and decision-making module employs a risk grading, scenario adaptation, and multi-objective optimization mechanism, maintains a multi-scale policy library, dynamically adjusts thresholds, and supports personalized policy customization. The strategy library includes three categories: long-term predictive maintenance plans, mid-term early warning and online balancing strategies, and emergency intervention strategies. Each strategy is associated with a cost-benefit evaluation model. The threshold dynamic adjustment rule is based on the failure rate over the past 30 days. Predicted remaining battery life The threshold is updated using the following formula: : ； ； in, This is a low-to-medium risk threshold used to trigger medium-term early warning strategies; This is a medium-to-high risk threshold used to trigger emergency intervention strategies; The failure rate over the past 30 days is calculated as the ratio of the number of failures to the total operating time. The remaining battery life is a predicted value, derived from the health state vector. It was deduced that; Strategy selection logic: When the overall fault risk level is low, output a long-term maintenance plan; when the level is medium or low, output a long-term maintenance plan. Any element in At the same time, it outputs mid-term early warning and online balancing strategies, prioritizing power supply to core loads; the level is high or Any element in In case of an emergency, an emergency intervention strategy is output, and the backup power switching mechanism is activated simultaneously to reduce the scope of the fault's impact.

6. The battery online monitoring and fault prediction system according to claim 5, characterized in that, The online balancing strategy introduces a risk contribution weighting and dynamic balancing cycle mechanism, and is dually adapted to the balancing weight calculation and operation mode, as well as the battery aging stage. ； in, For single cell batteries The equilibrium weights are determined by the weights; the higher the weight, the higher the equilibrium priority. For single cell batteries Real-time voltage; The average voltage of the battery cluster; For single cell batteries The real-time internal resistance; The average internal resistance of the battery cluster; For single cell batteries The maximum risk value in the risk state vector; For single cell batteries Capacity decay rate; Let be the weighting coefficient, satisfying Dual adaptation rules: Still state: ; Charging status: ; Discharge state: ; Late aging stage: additional Increase by 0.1 to enhance the balancing priority of degraded batteries; The balancing period is dynamically adjusted based on the maximum voltage difference within the cluster; the larger the voltage difference, the shorter the period.

7. The battery online monitoring and fault prediction system according to claim 6, characterized in that, The closed-loop optimization and digital twin module integrates the entire process of data feedback, model update, simulation verification, and strategy iteration, and has the capability of virtual-real linkage calibration, specifically including: Feedback Data Pool: Stores actual fault events, maintenance operation records, twin simulation data, and corresponding preliminary diagnostic data, with a capacity of 1×10. 4 -1×10 6 The algorithm employs a priority experience replay mechanism to assign higher sampling weights to high-risk prediction hit data and fault missed / misjudged data, while also removing abnormal interference data through data cleaning algorithms. Model Incremental Learning Unit: An incremental learning strategy of freezing the bottom layer and fine-tuning the top layer is adopted. Every 500-2000 new feedback data are accumulated, the network parameters of the hierarchical fusion diagnostic module are fine-tuned, the coupling coefficient and risk propagation model matrix are updated, the AdamW optimizer is used, a regularization term is introduced to prevent overfitting, and the model update log is recorded for traceability. Digital twin simulation engine: Constructs a virtual battery pack based on an electrochemical-thermal-mechanical multiphysics coupling model, through... It dynamically corrects simulation parameters based on real-time running data, supports parallel simulation of different risk scenarios and maintenance strategies, and predicts the effect, potential chain reactions and economic costs of major maintenance strategies before implementation. The simulation error is ≤5%, and the simulation results serve as an important basis for strategy output.

8. The battery online monitoring and fault prediction system according to claim 7, characterized in that, The digital twin simulation engine adopts a virtual-real synchronization-parameter self-calibration process to achieve dynamic optimization of simulation accuracy. The specific parameter correction process is as follows: based on Capacity decay in Internal resistance growth rate and energy efficiency The core parameters of the electrochemical model are corrected using the following formula: ； ； ； in, The capacity used in the simulation represents the actual usable capacity of the virtual battery. The internal resistance is used for simulation to characterize the conduction loss characteristics of the virtual battery. To simulate the amount of charge generated by the electrode reaction, and to characterize the reactivity of the electrode active materials in the virtual battery; The rated reaction capacity of the electrode is [value], which represents the battery design parameters; thermal model parameter correction: through [method / method]... Adjusting the heat dissipation coefficient during battery aging: ； in, For heat dissipation coefficient used in simulation; The baseline heat dissipation coefficient is determined by the battery casing material and structure. Mechanical stress model parameter correction: through internal resistance growth rate Corrected electrode expansion coefficient , The baseline expansion coefficient is used to ensure consistency between the simulated and physical battery pack's electro-thermal-mechanical coupling behavior; Simultaneous virtual-real calibration: Every hour, the deviation between the real-time data from the sensing and acquisition module and the twin simulation data is calculated. When the deviation of key parameters related to voltage and temperature exceeds 3%, a second parameter correction is triggered to ensure simulation accuracy.

9. The battery online monitoring and fault prediction system according to claim 8, characterized in that, The battery online monitoring and fault prediction system is adapted and expanded for different application scenarios, supporting multi-scenario compatible deployment: In the energy storage power station scenario: the sensing and acquisition module supplements the insulation resistance of the battery cluster and the sealing performance of the cabinet; the risk state inference sub-network strengthens the coupled risk assessment of thermal runaway and insulation failure; the digital twin engine adds the functions of charge-discharge cycle life simulation and cluster collaborative control simulation; and the maintenance strategy prioritizes the needs of grid dispatch. Electric vehicle battery swapping station scenario: The sampling frequency is increased to 100Hz, a unique battery identification and full life cycle traceability function are added, the emergency intervention strategy adds battery swapping priority recommendation, the model incremental learning unit shortens the update cycle to accumulate 500 feedback data, and supports the balance between battery swapping efficiency and battery safety; For communication base station backup power scenarios: optimize the low-power acquisition mode in the idle state, add battery activation suggestions and backup power switching linkage strategies to the long-term maintenance plan, supplement the abnormal mode library with low power self-discharge and float charge voltage drift abnormal templates, and prioritize the policy output to ensure the power supply continuity of the base station core equipment.

Images