Battery charge and state of health online estimation method based on multi-parameter fusion

Abstract

The application discloses a kind of based on multi-parameter fusion's battery charge and health state online estimation method, belong to battery state monitoring technical field.In the state that battery pack and load keep connection, group voltage, current, environmental temperature and each monomer voltage, internal resistance, temperature are collected online and double calibrated;According to error covariance and prediction confidence interval, state confidence is calculated respectively, and only when below threshold, selective bidirectional correction is triggered;When monomer dispersion exceeds limit, health state falls into maintenance interval or low confidence cumulative length reaches set value, online capacity is triggered based on DC-DC boost, to update model parameters with capacity measured by capacity, to BP network incremental learning and reset error covariance and prediction confidence interval.The application avoids SOC and SOH error cross contamination, inhibits long-term float estimation drift, fully releases capacity calibration value, and is suitable for communication power supply and other non-power-off scenarios.

Description

Technical Field

[0001] This invention belongs to the field of battery status monitoring technology, and relates to an online estimation method for battery charge and health status based on multi-parameter fusion. Background Technology

[0002] Battery banks are the last line of defense in ensuring uninterrupted power supply in communication power systems. In accordance with the mandatory requirement in the "Design Code for Installation Engineering of Communication Power Supply Equipment" (GB 51171-2016) that "backup power supplies must not be arbitrarily disconnected," communication batteries must be continuously connected to the DC bus in a floating charge standby state. To ascertain the true backup capacity of the batteries, maintenance units need to periodically conduct verification capacity tests (referred to as "capacity verification") and continuously monitor their state of charge and health.

[0003] Traditional capacity assessment methods involve offline discharge testing, requiring the battery pack to be disconnected from the DC bus and connected to an external dummy load for discharge. This not only poses a risk of communication interruption but also relies on manual on-site operation, resulting in high labor costs and long operation cycles. To address this, existing technologies have developed online capacity assessment schemes based on DC / DC converters. These schemes maintain the battery pack connected to the load while adjusting the battery's charging and discharging current using a DC / DC converter. The actual discharged capacity is then calculated based on the integral of the discharge current over time, thus completing the capacity assessment without affecting the power supply to the communication load.

[0004] In terms of state estimation, early methods mainly used the ampere-hour integration method to estimate the state of charge (SOC). This method is simple to implement but suffers from cumulative errors and cannot reflect capacity degradation caused by battery aging. Subsequently, existing technologies have developed SOC estimation methods based on the Thevenin equivalent circuit model combined with extended Kalman filtering (EKF), as well as methods for estimating state of equilibrium (SOH) based on impedance parameters, cycle counts, and other features combined with neural networks. Furthermore, existing technologies have also disclosed a collaborative estimation scheme that combines SOC and SOH estimation: using ampere-hour integration as a bridge, employing dual filters to iteratively calculate SOC and SOH separately, or using impedance parameter identification results in conjunction with neural networks to simultaneously estimate SOC and SOH, and using the estimation results for graded early warning.

[0005] However, the applicant has found through research that the aforementioned prior art has at least the following problems:

[0006] First, existing joint estimation schemes for SOC and SOH employ indiscriminate bidirectional feedback between the two state variables. In reality, SOC is a variable that changes rapidly on a second-by-second scale, while SOH is a variable that changes slowly on an hourly or even daily scale. Indiscriminately substituting SOH back to correct SOC and then substituting SOC deviation back to correct SOH within each estimation cycle will lead to SOC estimation being contaminated by SOH errors that have not yet converged, or SOC transient errors contaminating SOH estimation, forming a vicious cycle of error cross-propagation. After long-term operation, the estimation results will diverge.

[0007] Second, existing online capacity assessment schemes treat capacity assessment as an independent calibration action performed at fixed intervals, which is disconnected from the front-end monitoring and early warning processes: when to assess capacity is determined solely by the plan, rather than by the actual state of the battery. Especially for communication battery packs that are frequently float-charged and shallow-charged / discharged, deep-charged / discharge opportunities are extremely rare. The estimation model lacks a true capacity correction benchmark between two capacity assessments, and the estimation errors of SOC and SOH gradually accumulate and drift over time. Existing timed capacity assessment mechanisms cannot respond to this drift in a timely manner.

[0008] Third, existing methods, when using measured capacity for calibration, only use it to update the numerical values ​​of model parameters without addressing the confidence level of the estimation results. In other words, the system's "confidence in its own estimation results" remains unchanged before and after the capacity is adjusted, and quantities characterizing estimation uncertainty, such as error covariance and prediction confidence interval, are not reset, resulting in the full release of the calibration value of the capacity.

[0009] Fourth, the existing solution assumes that the front-end acquisition stage is accurate, and sensor calibration is only used to correct the sampled values ​​themselves; however, in actual engineering, the reference source will age and the temperature drift compensation curve will shift. The acquisition uncertainty represented by the calibration residual is not transmitted to the noise model of the back-end estimation model, and the acquisition accuracy and estimation accuracy are disconnected from each other in the data link.

[0010] Fifth, existing solutions mostly take the battery pack as the estimation object. When an individual cell experiences a micro-short circuit, sudden degradation, or is replaced, its abnormal parameters will be mixed into the health characteristics of the entire pack, contaminating the estimation results of the entire pack's SOC and SOH. Existing technologies lack a mechanism for identifying and isolating mutated cells.

[0011] Therefore, how to provide an online SOC and SOH estimation method that can avoid cross-contamination of SOC and SOH errors, trigger core capacity on demand and fully release the value of core capacity calibration, and has adaptive capability to acquisition drift and single-cell mutation in communication battery application scenarios with long-term float charging and non-disconnection of power is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0012] The purpose of this invention is to provide an online estimation method for battery charge and health status based on multi-parameter fusion, thereby solving the aforementioned technical problems.

[0013] The technical solution adopted in this invention is as follows:

[0014] A method for online estimation of battery charge and state of health based on multi-parameter fusion includes: with the battery pack connected to the load, online acquisition of the pack voltage, charging / discharging current, ambient temperature, and individual cell voltage, internal resistance, and temperature, followed by dual calibration to obtain a calibration parameter set; on a fast time scale, applying improved ampere-hour integral and extended Kalman filtering to the calibration parameter set based on the Thevenin equivalent circuit model, and weighted fusing the two estimation results to obtain a state of charge value; on a slow time scale, constructing a health feature vector from the individual cell internal resistance, available capacity, and cycle count in the calibration parameter set, inputting it into a pre-trained BP neural network to obtain a health state value; calculating the state of charge confidence score based on the error covariance of the extended Kalman filter, and calculating the health state confidence score based on the prediction confidence interval of the BP neural network; when the state of charge... When the confidence level is lower than a preset state of charge confidence threshold, the available capacity benchmark is corrected using the health state value and the state of charge value is substituted back for correction; when the health state confidence level is lower than a preset health state confidence threshold, the health feature vector is corrected in reverse using the estimation deviation of the state of charge value to obtain the corrected state of charge value and health state value; when the dispersion of the individual cell voltage or individual cell internal resistance in the calibration parameter set is greater than a preset dispersion threshold, the health state value falls into a preset maintenance range, or the cumulative duration of the confidence level being lower than the threshold reaches a set duration, online capacity verification based on DC-DC boost is triggered to obtain the measured capacity; the parameters of the Thevenin equivalent circuit model are updated with the measured capacity and the BP neural network is incrementally learned, while the error covariance matrix and the prediction confidence interval are reset to the initial state based on the measured capacity.

[0015] Optionally, the dual calibration includes hardware calibration based on a reference source and software compensation for zero-point and temperature drift; and the gain calibration error, zero-point offset, and temperature drift compensation residual obtained from the calibration are used as quantitative values ​​of acquisition uncertainty to adjust the process noise covariance and observation noise covariance of the extended Kalman filter online.

[0016] Optionally, the calculation of the state of charge confidence includes: taking the variance of the component corresponding to the state of charge in the error covariance matrix as the basic confidence metric; when the battery is operating in the plateau region of the open-circuit voltage-state of charge curve, the basic confidence metric is weighted and corrected according to the voltage observation sensitivity of the plateau region.

[0017] Optionally, the health status confidence is determined based on the variance of the BP neural network's predictions of multiple perturbations of the same health feature vector. The larger the prediction variance, the lower the health status confidence. Both the charge state confidence threshold and the health status confidence threshold have hysteresis intervals. The charge state confidence or the health status confidence must cross the corresponding hysteresis interval before the corresponding correction is triggered or released.

[0018] Optionally, the update cycle of the fast timescale is on the order of seconds, and the update cycle of the slow timescale is on the order of hours or days; the health status value is updated only when online capacity is completed, a deep charge / discharge event occurs, or the reverse correction is triggered.

[0019] Optionally, for battery packs operating under long-term float charging, the set duration is determined by weighting the cumulative duration of confidence levels below a threshold with the float charging duration, in order to suppress estimation drift caused by the lack of a core capacity anchor point under long-term shallow charge and discharge conditions.

[0020] Optionally, updating the Thevenin equivalent circuit model parameters with the measured capacity, incrementally learning the BP neural network, and resetting the error covariance matrix and the prediction confidence interval to an initial state based on the measured capacity specifically includes: calculating the capacity deviation between the measured capacity and the model predicted capacity; when the capacity deviation exceeds a set deviation threshold, identifying the Thevenin equivalent circuit model parameters online, and incrementally learning the BP neural network using data collected during the capacity verification process as new samples; resetting the error covariance matrix to the product of the initial matrix and the deviation correction coefficient, and resetting the prediction confidence interval to a benchmark interval determined by the measured capacity.

[0021] Optionally, it also includes single-cell mutation isolation: when a sudden change in the single-cell voltage or single-cell internal resistance of an individual cell is identified, the parameters of that single cell are isolated from the health feature vector used for the overall estimation, and an estimation model is established for that single cell separately; when the replacement of that single cell is detected, the parameters of its single-cell estimation model are reset.

[0022] Optionally, it also includes online maintenance: when the health status value falls into the preset maintenance range, the online capacity verification is first triggered. After verification, the DC-DC boost module is controlled to perform equalization charging or activation charging and discharging on the battery pack. After maintenance is completed, the state of charge value estimation on the fast time scale and the health status value estimation on the slow time scale, as well as the selective bidirectional coupling correction triggered by the confidence threshold, are re-executed.

[0023] Optionally, it also includes tiered early warning: outputting the final state of charge value and its state of charge confidence, as well as the health state value and its health state confidence;

[0024] When the state of charge value exceeds the preset normal range of the state of charge and / or the health value exceeds the preset normal range of the health, and the corresponding confidence level is not lower than the corresponding preset confidence threshold, the preset warning level corresponding to the over-limit item and the degree of over-limit is directly output.

[0025] When the state of charge value and / or the health status value exceed the preset normal range, but the corresponding confidence level is lower than the corresponding preset confidence threshold, the online capacity verification is first triggered to obtain the measured capacity. The model parameters are then updated with the measured capacity and the confidence level is reset. Finally, the warning level is determined based on the re-estimated state of charge value and / or health status value.

[0026] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0027] I. This invention allocates the estimation of SOC and SOH to two time scales, fast and slow, respectively, and selectively controls the bidirectional coupling correction through a confidence threshold: the healthy state value is used for back-substitution correction only when the confidence of the state of charge is below its threshold, and the healthy feature vector is corrected in reverse using the state of charge estimation deviation only when the confidence of the healthy state is below its threshold. Compared with the indiscriminate bidirectional feedback in the prior art, this invention retains the advantages of collaborative estimation of SOC and SOH while cutting off the error cross-propagation path of "erroneous SOH contaminating SOC and transient SOC error contaminating SOH", which significantly improves the stability and convergence of the estimation results under long-term operating conditions.

[0028] Second, this invention improves online capacity verification from timed execution to event-driven execution: online capacity verification based on DC-DC boost is only triggered when the dispersion of single-cell voltage or single-cell internal resistance exceeds the limit, the health status value falls into the preset maintenance range, or the cumulative time when the confidence level is below the threshold reaches a set time, forming a closed loop of "monitoring-judgment-capacity verification-update". On the one hand, it avoids unnecessary capacity verification operations that damage battery life, and on the other hand, it ensures that capacity verification resources are accurately allocated to battery packs that truly need calibration; especially for communication battery packs that are float-charged and shallow-charged and shallow-discharged for a long time, the low-confidence cumulative time triggering mechanism makes up for the scarcity of capacity verification anchors and effectively suppresses estimation drift.

[0029] Third, this invention uses the measured capacity of the core capacitor simultaneously as both a "parameter anchor" and a "confidence anchor": not only does it update the parameters of the equivalent circuit model and perform incremental learning on the BP neural network, but it also resets the error covariance matrix and the prediction confidence interval to their initial state based on the measured capacity of the core capacitor. Therefore, the core capacitor not only corrects the "estimated value" but also the "degree of confidence in the estimated value," fully releasing the calibration value of each core capacitor operation, and continuously improving the estimation accuracy with running time and the number of core capacitor operations.

[0030] Fourth, this invention uses the gain calibration error, zero-point offset, and temperature drift compensation residual obtained from dual calibration as quantitative values ​​of acquisition uncertainty, and adjusts the process noise covariance and observation noise covariance of the extended Kalman filter online, thus opening up the data link between "acquisition accuracy and estimation accuracy". This enables the estimation process to adapt to the degradation of the front-end acquisition links, such as reference source aging and temperature drift compensation curve shift, and overcomes the defect of the existing technology that assumes absolute accuracy of front-end acquisition.

[0031] Fifth, this invention identifies and isolates individual cells whose voltage or internal resistance changes abruptly, removes their parameters from the overall health feature vector and models them separately, and resets the corresponding cell estimation model after detecting cell replacement. This avoids individual abnormal cells from contaminating the overall SOC and SOH estimation results, and improves the adaptability of the estimation method to the discrete evolution of individual cells in the battery pack and maintenance replacement scenarios.

[0032] VI. The graded early warning system of this invention outputs the status value and its confidence level together. In the case of the value exceeding the limit but the confidence level is insufficient, online capacity verification is triggered first and then the early warning level is determined, which reduces the false alarm rate. Online maintenance is also first confirmed by capacity verification before equalization charging or activation charge and discharge is performed, which avoids damage to the battery caused by incorrect maintenance.

[0033] VII. This invention operates online while the battery pack remains connected to the load throughout the entire process, without needing to disconnect the battery from the communication load. This fundamentally eliminates the risk of communication interruption during the capacity assessment process and complies with the mandatory requirements of GB 51171-2016. At the same time, the entire process of estimation, capacity assessment, maintenance, and early warning is automated, promoting the transformation of communication battery operation and maintenance from "manual passive repair" to "intelligent proactive early warning," significantly reducing labor costs and shortening operation time. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein:

[0035] Figure 1 This is a schematic diagram illustrating the principle and process of an online estimation method for battery charge and health status based on multi-parameter fusion.

[0036] Figure 2 This is a schematic diagram of the parallel access structure described in an embodiment of the present invention;

[0037] Figure 3 This is a flowchart of the dual-channel weighted fusion estimation process described in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0039] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0040] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0041] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides an online estimation method for battery charge and health status based on multi-parameter fusion, applicable to communication battery packs connected to the DC bus of the communication power supply and in a long-term floating charge standby state. Figure 2 As shown, the DC-DC boost module used for online capacity verification is connected in parallel with the battery pack and the communication load. During capacity verification, the energy transfer of the battery pack is achieved by adjusting the charging and discharging current of the DC-DC boost module. The connection between the battery pack and the communication load does not need to be disconnected throughout the process, and the power supply to the communication load is uninterrupted.

[0044] The estimation method in this embodiment is interconnected in the data link: While the battery pack remains connected to the load, the group voltage, charging / discharging current, and ambient temperature of the battery pack, as well as the individual cell voltage, internal resistance, and temperature of each cell, are collected online. The collected parameters are then double-calibrated to obtain a calibration parameter set. This calibration parameter set drives state of charge (SOC) estimation on a fast timescale to obtain SOC values, and simultaneously drives state of health (SQ) estimation on a slow timescale to obtain SQ values. Simultaneously, the SOC confidence level and SQ confidence level are calculated for both types of SQ values. The confidence level determines whether to perform bidirectional coupling correction, and the final SOC and SQ values ​​are output after correction. When the dispersion exceeds the limit, the SQ value falls into a preset maintenance range, or the accumulated low confidence time reaches a set duration, online capacity verification based on DC-DC boost is triggered to obtain the measured capacity. The measured capacity is then fed back to the estimation model to update the model parameters and reset the confidence level. The output of each link in the link serves as the input to the next link, forming a closed-loop self-calibrating estimation system.

[0045] I. Multi-parameter acquisition and dual calibration:

[0046] Specifically, the group voltage is acquired using a voltage divider circuit in conjunction with a high-precision analog-to-digital converter; the charging and discharging current is acquired using a Hall sensor; the ambient temperature and individual cell temperature are acquired using a digital temperature sensor; and the individual cell internal resistance is acquired using a DC injection method, which involves injecting a disturbance current of a set amplitude into the individual cell, using a lock-in amplifier to extract the component in the individual cell voltage response that has the same frequency as the disturbance current, and calculating the individual cell internal resistance r accordingly. These six types of parameters, along with their acquisition times, constitute the original parameter set.

[0047] Dual calibration includes hardware calibration based on a reference source and software compensation for zero-point and temperature drift. Hardware calibration involves calibrating the gain of each voltage acquisition channel using a high-precision reference voltage source. Software compensation includes acquiring the zero-point offset of the current sensor during periods when the charging and discharging current is zero and subtracting it from subsequent sampled values, as well as performing temperature drift compensation on each acquisition parameter based on a pre-stored temperature-drift compensation curve. The parameters obtained after dual calibration constitute a calibration parameter set, which serves as the data input for all subsequent estimation stages.

[0048] Furthermore, this embodiment does not assume absolute accuracy of front-end acquisition upon completion of calibration. Instead, it retains the gain calibration error, zero-point offset, and temperature drift compensation residual generated during the calibration process as quantified values ​​of acquisition uncertainty. These values ​​are used to adjust the observation noise covariance of the extended Kalman filter online. The adjustment relationship is as follows:

[0049] R k = R b + ρ·( a g 2 + a o 2 + a T 2 );

[0050] In the formula, R k R represents the online adjusted observation noise covariance, which is the output of this relationship and directly determines the degree of confidence the extended Kalman filter has in the current measurement value. b The basic observation noise covariance is determined by the nominal accuracy of the acquisition channel; a g This refers to the gain calibration error, which is the residual gain deviation between the reference source and the acquisition channel after hardware calibration; a o This refers to the zero-point offset, which is the residual zero-point deviation of the current sensor after software compensation; a T The temperature drift compensation residual is the remaining drift amount that the temperature-drift compensation curve fails to cover; ρ is the residual weighting coefficient, used to unify the dimensions of the three types of calibration residuals to the dimensions of the observation noise covariance.

[0051] The logic behind this relationship is as follows: the larger the three types of calibration residuals, the higher the probability that the current acquired value deviates from the true value, and the greater the observation noise covariance. Consequently, the extended Kalman filter automatically reduces its trust in the measurement value and relies more on model prediction. Conversely, the smaller the calibration residuals, the stronger the filter's trust in the measurement value. Thus, changes in the accuracy of the front-end acquisition stage are transmitted to the back-end estimation stage in real time, connecting the acquisition accuracy and the estimation accuracy in the data link.

[0052] II. Fast timescale estimation of state of charge:

[0053] On a fast timescale, state of charge (SOC) estimation is performed on a second-level cycle. Since the actual discharge capacity of an aging battery is lower than its rated capacity, using the rated capacity as the capacity benchmark for ampere-hour integration would introduce a systematic bias that worsens with aging. Therefore, the usable capacity benchmark is first corrected using the health status value output on a slow timescale:

[0054] Q N = Q R •SOH;

[0055] In the formula, Q N The usable battery capacity, corrected for its current health condition, is the output of this equation, serving as the capacity reference for ampere-hour integration; Q R The rated capacity of the battery is taken as the factory-specified capacity; the State of Health (SOH) value is the output of the slow time-scale estimation stage and used as the input of this stage. The lower the SOH value, the smaller the usable capacity benchmark, and the greater the change in state of charge produced by the ampere-hour integral for the same charge and discharge capacity, thus ensuring that the capacity benchmark estimated by the state of charge is consistent with the battery's current actual storage capacity.

[0056] Based on the revised capacity reference, the first state-of-charge value is calculated using the improved ampere-hour integral method that incorporates a charge-discharge efficiency coefficient:

[0057] ;

[0058] In the formula, SOC A (t) is the first state of charge value, which is the output of the relationship and represents the remaining charge level of the battery at time t; SOC(t0) is the estimated initial state of charge at the starting time t0, which is obtained by looking up the open circuit voltage-state of charge characteristic curve after resting, or by taking the state of charge value stored at the last shutdown. The accuracy of its value directly affects the accuracy of the integration starting point; η is the charge and discharge efficiency coefficient, which is a binary function of ambient temperature T and charge and discharge current I. It is obtained by looking up a pre-calibrated two-dimensional table and is used to characterize the difference between the actual convertible capacity and the input capacity under low temperature and high current conditions, and to overcome the cumulative error caused by the fixed efficiency assumption; I(u) is the charge and discharge current corresponding to the integration time variable u, with the discharge direction defined as positive and the charging direction as negative; Q N The usable capacity has the same meaning as the aforementioned relationship. The larger the charge / discharge current and the further the ambient temperature deviates from normal temperature, the greater the deviation of the charge / discharge efficiency coefficient from one, and the more significant the correction effect on the first state of charge value.

[0059] As an optional implementation method, the charge-discharge efficiency coefficient η is pre-calibrated through standard charge-discharge tests: at multiple ambient temperature points (e.g., -20℃, 0℃, 25℃, 60℃) and multiple charge-discharge current rate points (e.g., 0.05C, 0.1C, 0.2C, 0.5C, 1C), the sample battery is subjected to a complete standard charge-discharge cycle. The ratio of the actual amount of electricity discharged or charged to the theoretical amount of electricity at each operating point is used as the charge-discharge efficiency coefficient for that operating point, thereby constructing a two-dimensional calibration table. During battery operation, bilinear interpolation is performed on the two-dimensional calibration table according to the current ambient temperature and charge-discharge current to obtain the charge-discharge efficiency coefficient at the current moment.

[0060] Therefore, the improvements of the improved ampere-hour integration method in this embodiment compared to the conventional ampere-hour integration method are reflected in three aspects: First, the capacity reference is improved from a fixed rated capacity to a usable capacity that is dynamically corrected according to the health status value; second, the charge and discharge efficiency is improved from a fixed coefficient to an efficiency coefficient that varies with ambient temperature and charge and discharge current and is calibrated through standard tests; third, the initial integration value is improved from an arbitrarily set value to a value calibrated from the open-circuit voltage-state-of-charge characteristic curve. These three improvements simultaneously suppress accumulated errors from the perspectives of reference, coefficient, and initial value, and they work together and are indispensable.

[0061] Meanwhile, a Thevenin equivalent circuit model of the battery is established, using ohmic internal resistance, polarization resistance, and polarization capacitance to characterize the battery's dynamic characteristics. The observation equations of the model are:

[0062] U t = U oc (SOC, T) − I·R0 − U p ;

[0063] In the formula, U t U is the battery terminal voltage, which is the output of this equation and serves as an observation for the extended Kalman filter; oc U is the open-circuit voltage, a bivariate function of the state of charge and ambient temperature, given by the open-circuit voltage-state-of-charge characteristic curve with a temperature compensation factor; R0 is the ohmic internal resistance, characterizing the instantaneous response of the terminal voltage to a sudden change in current; U p The polarization voltage characterizes the gradual change in terminal voltage caused by electrochemical polarization and concentration polarization; I represents the charging and discharging current, whose meaning is completely consistent with the aforementioned relationship. The greater the deviation between the terminal voltage and the open-circuit voltage and current, the more significant the polarization effect, and the filter corrects the estimation of the polarization voltage accordingly.

[0064] The state variables of the model are the state of charge and the polarization voltage, and its discrete state transition relationship is as follows:

[0065] ;

[0066] ;

[0067] In the formula, k is the discrete time index; Δt is the discrete sampling period, the value of which determines the speed of state update; R p For polarization resistance, C pThe polarization capacitor and its product constitute the polarization time constant, which determines the response speed of the polarization voltage to a current step; I(k) is the charging and discharging current at the k-th discrete moment; exp(·) is an exponential function; the meanings of the remaining parameters are completely consistent with the aforementioned relationships. The ohmic internal resistance, polarization resistance, and polarization capacitor are obtained online based on the open-circuit voltage-state-of-charge characteristic curve with a temperature compensation factor. Based on the above state transition relationship and observation equation, an extended Kalman filter is performed to obtain the second state-of-charge value (SOC). K .

[0068] Subsequently, as Figure 3 As shown, the two estimation results are dynamically weighted and fused:

[0069] SOC F = w1·SOC A + w2·SOC K w1 + w2 = 1;

[0070] In the formula, SOC F The weighted and fused state of charge (SOC) is the output of this equation. A The first state of charge (SOC) value. K The second state of charge value is represented by the improved ampere-hour integral method and the extended Kalman filter method, respectively. w1 and w2 are the fusion weights of the two estimation results. The weight adjustment rules are as follows: when the rate of change of charging and discharging current is less than the first threshold and the ambient temperature is in the normal temperature range, w1 is increased because the ampere-hour integral drifts slowly and has higher accuracy under steady-state low current conditions; when the rate of change of charging and discharging current is greater than the second threshold, or the ambient temperature is in the high temperature range or low temperature range, w2 is increased because the equivalent circuit model filtering under dynamic conditions and extreme temperatures is more robust to transient disturbances and model mismatch; the second threshold is greater than the first threshold, and a transition band is formed between the two thresholds to avoid frequent weight jumps near the boundary.

[0071] As an optional implementation, the first threshold is 0.02C / s, the second threshold is 0.05C / s, and the normal temperature range is 15℃ to 35℃; under steady-state conditions, w1 is 0.7 and w2 is 0.3, and under dynamic or extreme temperature conditions, w1 is 0.3 and w2 is 0.7. When in the transition zone, the weights are linearly interpolated according to the rate of change of current and the degree of temperature deviation, so that the weights change continuously with the operating conditions.

[0072] III. Slow-timescale estimation of health status:

[0073] On slow timescales, health status estimation is performed on hourly or daily cycles and updated only when online core capacity is completed, deep charge / discharge events occur, or reverse corrections are triggered, avoiding slow variables being driven by high-frequency noise. The multi-dimensional health feature vector required for health status estimation consists of the growth rate of internal resistance relative to the initial value, capacity retention rate, normalized cycle count, and ambient temperature. The growth rate of internal resistance relative to the initial value is:

[0074] Δr = (r − r0) / r0;

[0075] In the formula, Δr is the growth rate of the internal resistance relative to the initial value, which is the output of the equation and characterizes the degree of degradation of the battery's conductive path; r is the current internal resistance of the single cell measured by the DC injection method; r0 is the initial internal resistance of the single cell, calibrated when the battery is put into operation. By using the relative growth rate instead of the absolute value of the internal resistance, the difference in initial internal resistance between different batches and specifications of batteries can be eliminated, making this characteristic comparable between different battery packs.

[0076] The growth rate of internal resistance relative to the initial value, capacity retention rate, normalized cycle count, and ambient temperature are combined to form a health feature vector, which is then input into a BP neural network pre-trained with historical nuclear capacity data to obtain the health status value.

[0077] SOH = f BP (Δr, κ, n, T);

[0078] In the formula, SOH is the health status value, which is the output of the relationship; f BP This represents the nonlinear mapping fitted by the BP neural network; the meaning of Δr is completely consistent with the aforementioned relationship; κ is the capacity retention rate, taken as the ratio of the most recent measured capacity to the rated capacity, reflecting the degree of loss of active material; n is the normalized cycle number, taken as the ratio of the cumulative equivalent charge-discharge cycle number to the designed cycle life, reflecting the battery's usage intensity; T is the ambient temperature, reflecting the environmental stress level the battery has been exposed to for a long time. These four dimensions respectively characterize conductivity degradation, capacity decay, usage intensity, and environmental stress, complementing each other; as any dimension increases, the output of the health status value decreases accordingly, thus ensuring that the health status estimation remains sensitive to different types of aging modes.

[0079] As an optional implementation, the BP neural network adopts a three-layer structure: an input layer, hidden layers, and an output layer. The input layer includes four nodes, which receive the growth rate of internal resistance relative to the initial value, the capacity retention rate, the normalized number of iterations, and the ambient temperature, respectively. The hidden layer includes ten nodes, and the activation function is the sigmoid function. The output layer includes one node, which outputs the health status value. Training samples are taken from historical nuclear capacity data: the health feature vectors collected before and after each nuclear capacity measurement are used as input, and the actual health status determined by the measured nuclear capacity is used as the label for offline training.

[0080] IV. Confidence Assessment:

[0081] While obtaining the two types of state values, calculate their confidence scores for each. The confidence score for the charged state is:

[0082] c S = 1 / (1 + λ· p s ), λ = g ref / g;

[0083] In the formula, c S p represents the confidence level of the state of charge, which is the output of this relation. A larger value indicates a more reliable state of charge value. s λ is the variance of the component corresponding to the state of charge in the error covariance matrix, directly characterizing the uncertainty of the extended Kalman filter's estimation of the state of charge; the larger the variance, the lower the confidence level. λ is the weighting coefficient in the plateau region; g is the slope of the open-circuit voltage-state of charge characteristic curve at the current state of charge, i.e., the voltage observation sensitivity. ref The slope of the linear region of the curve is taken as the reference slope. The logic behind this relationship is that when the battery is operating in the plateau region of the curve, the terminal voltage is not sensitive to changes in the state of charge. Decreasing g increases λ, and even if the variance remains unchanged, the confidence level is still reduced in weight, thus accurately reflecting the physical fact that the voltage observation in the plateau region has a weakened ability to constrain the state of charge.

[0084] The confidence level for health status is:

[0085] c H = 1 / (1 + σ B );

[0086] In the formula, c H The confidence level for health status is the output of this relation; σ B This represents the variance of the prediction results obtained by the BP neural network after applying multiple small perturbations to the same health feature vector. A larger perturbation variance indicates that the network input falls into an area insufficiently covered by the training samples, the degree of extrapolation is high, and the reliability of the health status value is lower; conversely, a smaller perturbation variance indicates higher reliability. Furthermore, both confidence thresholds have hysteresis intervals; the corresponding correction is only triggered or deactivated when the confidence level crosses the hysteresis interval, preventing frequent start and stop of correction actions near the threshold.

[0087] As an optional implementation, the multiple small perturbations are applied as follows: a random perturbation of 1% to 5% is applied to each dimension of the health feature vector, and the forward propagation is repeated 50 times. The variance of the 50 prediction results is taken as σ. B The hysteresis interval of the confidence threshold can be set, for example, as follows: the threshold for triggering correction is 0.6, and the threshold for canceling correction is 0.7.

[0088] V. Selective bidirectional coupling correction triggered by confidence threshold:

[0089] Unlike existing technologies that perform bidirectional feedback in every estimation cycle, the bidirectional coupling correction in this embodiment is selectively triggered by a confidence threshold. Only when the state of charge confidence level c... S When the value falls below its threshold, the available capacity baseline is updated according to the aforementioned available capacity correction relationship based on the current health status value, and the first and second state of charge values ​​are recalculated back to correct the merged state of charge value; only when the health status confidence level c H When the value is below a threshold, the health eigenvector is corrected by reverse calculation of the deviation based on the state of charge, where the correction relationship for capacity retention is as follows:

[0090] κ ← κ + μ·δ ;

[0091] In the formula, κ represents the capacity retention rate, which has the same meaning as the aforementioned relationship. This relationship outputs the corrected capacity retention rate. δ represents the state of charge (SOC) estimation deviation, which is the difference between the weighted fused SOC value and the coulomb measurement reference value at the same moment. The coulomb measurement reference value refers to the SOC trajectory obtained by accumulating the subsequent charging and discharging currents in ampere-hours, based on the known SOC time. The known SOC time is, for example, the time when the capacity is fully utilized and the battery pack has been sufficiently rested, and the SOC is determined by looking up the open-circuit voltage-SOC characteristic curve from the open-circuit voltage. μ is the reverse correction coefficient, with a value between zero and one, which determines the step size of a single correction. The logic behind this relationship is that if the fused SOC continuously deviates from the coulomb measurement reference, it indicates that the capacity retention rate in the health feature vector is overestimated or underestimated. Correction is performed according to the proportion of the deviation and with a step size of less than one, which can both bring the health feature closer to the actual state of the battery and avoid oscillations caused by a single overcorrection. The correction is performed iteratively. When the change in state of charge between two consecutive iterations is less than the first convergence threshold and the change in health state is less than the second convergence threshold, or when the number of iterations reaches the set maximum number of iterations, the iteration stops and the final state of charge and health state values ​​are output. Since the coupled correction is only initiated unidirectionally in the direction of insufficient confidence, the cross-propagation path of erroneous health state contaminating the state of charge and transient state of charge error contaminating the health state is cut off.

[0092] As an optional implementation, both the first convergence threshold and the second convergence threshold are set to 0.5%, and the maximum number of iterations is set to 10.

[0093] VI. Event-driven online capacity triggering and execution logic:

[0094] In this embodiment, online capacity verification is not performed on a fixed cycle, but is triggered by events. The triggering conditions include: the dispersion d of the cell voltage or cell internal resistance exceeds the limit, i.e., d is greater than the dispersion threshold d0, indicating that there is consistent degradation within the battery pack; the health status value falls into a preset maintenance range, indicating that the battery has degraded to the point where intervention is required; and the cumulative time τ for the confidence level to be below the threshold reaches a set time τ0, indicating that the estimation result has been unreliable for a long time. Meeting any one of these three conditions triggers online capacity verification based on DC-DC boost.

[0095] As an optional implementation, the dispersion d is the ratio of the difference between the maximum and minimum values ​​of the individual cell voltage or the individual cell internal resistance to the mean value; the dispersion threshold d0 is 5% for individual cell voltage and 20% for individual cell internal resistance; the preset maintenance range is 60% to 80% of the healthy state value, and the scrap threshold is 60%; the set duration τ0 is 720 hours.

[0096] For battery packs that operate under long-term float charging, deep charging and deep discharging opportunities are extremely rare, and capacity anchor points are scarce. In this embodiment, the determination of whether the set duration has been reached is based on the rule that the cumulative duration is weighted by the float charging duration: the longer the float charging duration, the greater the contribution of the unit time to the cumulative duration, thereby enabling the capacity trigger to be brought forward with the increase of float charging duration, suppressing the estimation drift caused by the lack of a real capacity correction benchmark under long-term shallow charging and shallow discharging conditions.

[0097] Once the triggering conditions are met, the capacity verification is executed according to the following logic: The remote monitoring terminal or local control unit issues a capacity verification command; the current equalization charge, float charge, or discharge state of the battery pack is monitored; the DC-DC boost module is started to discharge using ramp control, with the charge / discharge current change rate limited to within 0.1C / s to avoid current surges causing voltage fluctuations on the communication power bus; voltage, current, and temperature are monitored in real time during discharge; discharge stops when the set cutoff threshold is reached; charging is then started using ramp control and stops when the total battery pack voltage recovers to the DC bus voltage; the measured capacity Q is obtained by integrating the charge / discharge current over time during discharge. M The system generates a capacity assessment report and uploads it to the monitoring terminal. If any abnormal voltage, excessive temperature, or abnormal communication power output occurs during the capacity assessment process, the capacity assessment will automatically stop and an alarm will be triggered.

[0098] VII. Circuit Topology of DC-DC Boost Module:

[0099] like Figure 2As shown, as an optional implementation, the DC-DC boost module adopts an interleaved parallel synchronous rectification Buck-Boost topology. Its main circuit includes switching transistors, synchronous rectifier diodes, inductors, and capacitors, and is equipped with a drive circuit and overvoltage, overcurrent, and overtemperature protection circuits. To adapt to the wide voltage range of the communication battery pack, the module adopts segmented topology control: when the battery pack voltage is in the 42V to 48V range, the two circuits operate in interleaved parallel to distribute current stress and reduce inductor current ripple; when the battery pack voltage is in the 48V to 56V range, it switches to single-circuit operation to reduce switching losses. Simultaneously, the module employs adaptive inductor current control, switching the inductor with different inductance values ​​via relays according to the battery pack voltage, limiting the inductor current ripple to within 5%, and achieving an overall efficiency of no less than 92%. Soft-switching technology is also used to reduce switching losses, and the drive timing of the synchronous rectifier diodes is optimized to reduce the conduction time of the body diode.

[0100] Furthermore, a supercapacitor is connected in parallel to the output of the DC-DC boost module to absorb voltage ripple during the charging and discharging process. The module also coordinates with the communication power supply via a CAN bus, feeding back the charging and discharging current to the communication power supply in real time. The communication power supply dynamically adjusts its output power accordingly, ensuring that the voltage fluctuation of the communication power supply bus during the charging and discharging process does not exceed ±1%. As an optional implementation, the module's output voltage range is 42V to 72V, the discharge current is adjustable from 0.1C to 1C, the voltage regulation accuracy is ±0.5V, and the current regulation accuracy is ±0.2A. One module can switch between managing two battery packs.

[0101] VIII. Confidence Anchoring Update for Actual Measured Capacity:

[0102] The measured capacity is used not only to update the model parameters but also to reset the confidence level of the estimate. First, the deviation rate between the measured capacity and the model predicted capacity is calculated:

[0103] e = |Q M - Q P )| / Q P ;

[0104] In the formula, e is the capacity deviation rate, which is the output of the equation and quantifies the degree of deviation of the estimation model from the actual capacity; Q M This refers to the actual measured capacity of the nuclear reactor; Q P The model predicts the capacity, which is the capacity predicted by the current estimation model for this capacity reduction process. The larger the deviation rate, the more severe the model drift, and the stronger the subsequent updates and resets will be.

[0105] When the capacity deviation rate e exceeds the set deviation threshold ε, the ohmic internal resistance, polarization resistance, and polarization capacitance of the Thevenin equivalent circuit model are identified online. Data collected during the capacity grading process is used as new samples to incrementally learn the BP neural network. This means retraining with a smaller learning rate while retaining the original network weights, preventing new capacity grading samples from overwriting historical knowledge. Simultaneously, the error covariance matrix and prediction confidence interval are reset to their initial state based on the measured capacity of the capacity grading system. The reset relationship for the error covariance matrix is ​​as follows:

[0106] P ← β·P0, β = 1 + e;

[0107] In the formula, P is the error covariance matrix, and this relation outputs the reset error covariance matrix; P0 is the initial error covariance matrix; β is the deviation correction coefficient; and e is the capacity deviation rate, whose meaning is completely consistent with the aforementioned relation. The construction logic of this relation is as follows: the core capacity provides a true capacity benchmark. During long-term operation, the filter may fall into spurious convergence due to model mismatch, that is, the variance is very small but the estimated value is wrong. Resetting the error covariance matrix back to the initial level can eliminate this spurious convergence. The deviation correction coefficient increases with the capacity deviation rate, so that the battery pack with more severe drift maintains a larger openness after resetting, and the filter can reconverge to the true state more quickly. The predicted confidence interval is simultaneously reset to the benchmark interval determined by the measured capacity of the core capacity. Thus, the core capacity not only corrects the estimated value itself, but also corrects the system's degree of confidence in the estimated value, and the calibration value of the core capacity is fully released. As an optional implementation method, the deviation threshold ε is set to 5%.

[0108] IX. Monomorphic Mutation Isolation:

[0109] Furthermore, this embodiment implements abrupt change monitoring for individual cells: when the change in the cell voltage or internal resistance of an individual cell exceeds the abrupt change criterion within adjacent sampling windows, the cell is determined to have undergone abrupt change. Its parameters are isolated from the health feature vector used for the overall group estimation, and a separate estimation model is established for that cell, ensuring that the overall state of charge and health estimation is not contaminated by individual abnormal cells. When it is detected that the cell has been replaced, for example, when its cell voltage and internal resistance return to the factory characteristic range of a new cell, the parameters of the cell estimation model are reset, allowing it to re-participate in the overall group estimation. As an optional implementation, the sampling window is 10 minutes. When the change in the cell voltage of a certain cell exceeds 50mV between adjacent sampling windows, or the change in its internal resistance exceeds 30%, the cell is determined to have undergone abrupt change.

[0110] 10. Online maintenance and tiered early warning:

[0111] When the health status value falls within the preset maintenance range, the online capacity verification is first triggered to confirm the actual status of the battery. After confirmation, the DC-DC boost module is controlled to perform equalization charging or activation charging and discharging on the battery pack. After maintenance is completed, dual time scale estimation and selective bidirectional coupling correction are re-executed to update the estimation results. This avoids applying maintenance operations to the battery based on a single potentially unreliable estimation result, preventing damage to the battery from erroneous maintenance.

[0112] The tiered early warning system outputs the final state of charge and health values ​​along with their confidence levels: when the value exceeds the limit but the confidence level is not lower than the threshold, the corresponding level of early warning information is directly output; when the value exceeds the limit but the confidence level is lower than the threshold, online capacity verification is triggered first, and then the early warning level is determined based on the estimated results after capacity verification, thereby reducing the false alarm rate.

[0113] The solution described in this embodiment is designed for communication battery packs that require long-term float charging and cannot be powered off. It constructs a closed-loop online estimation system consisting of "data acquisition—dual time-scale estimation—confidence assessment—selective correction—event-driven capacity verification—dual update of model and confidence".

[0114] First, the group voltage, charging and discharging current, ambient temperature, and individual cell voltage, internal resistance, and temperature are acquired online. After dual calibration by hardware calibration and software compensation, a calibration parameter set is obtained. At the same time, the calibration residuals (gain error, zero-point offset, temperature drift residuals) are quantified and fed back in real time to the observation noise covariance of the extended Kalman filter, realizing adaptive linkage between acquisition accuracy and estimation accuracy.

[0115] At the data application level, the method adopts a dual timescale strategy of fast and slow separation: the fast timescale (second level) is based on the Thevenin equivalent circuit model, and the improved ampere-hour integral and extended Kalman filter are executed in parallel, and the state of charge (SOC) value is obtained by dynamic weighted fusion according to the operating conditions; the slow timescale (hour level / day level) constructs the health feature vector of the cell internal resistance growth rate, capacity retention rate, normalized cycle number and ambient temperature, and inputs it into the BP neural network to obtain the state of health (SOH) value.

[0116] At the state coupling level, the confidence levels of SOC and SOH are calculated using the EKF error covariance and the BP network prediction confidence interval, respectively. Selective bidirectional correction is triggered only when the corresponding confidence level is below the threshold. When the SOC confidence level is insufficient, the available capacity benchmark is corrected by SOH and then back-substituted to correct the SOC. When the SOH confidence level is insufficient, the healthy feature vector is corrected by the SOC estimation bias. This cuts off the error cross-propagation path and avoids transient errors from polluting slow variables or aging errors from polluting fast variables.

[0117] At the closed-loop calibration level, when the individual cell voltage / internal resistance dispersion exceeds the limit, the SOH falls into the maintenance range, or the cumulative duration of low confidence reaches the set threshold, the event triggers online core capacity based on DC-DC boost. The obtained measured core capacity is not only used for online identification and updating of Thevenin model parameters and incremental learning of the BP network, but more importantly, it resets the EKF error covariance matrix and the SOH prediction confidence interval to the initial state based on the actual capacity, realizing the dual value release of "parameter anchoring" and "confidence anchoring", effectively suppressing estimation drift under long-term float charge and shallow charge and discharge conditions.

[0118] Furthermore, the system features individual mutation isolation, online maintenance, and tiered early warning functions: mutated monomers are modeled separately to prevent contamination during group estimation; both maintenance and early warning are based on joint confidence level assessment, and if the value exceeds the limit but the confidence level is insufficient, capacity verification is triggered first, significantly reducing false alarm rate and the risk of incorrect maintenance. The entire process does not require disconnecting the battery from the load, meeting the mandatory requirement that communication power supplies must not be interrupted.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online estimation of battery charge and state of health based on multi-parameter fusion, characterized in that, include: With the battery pack connected to the load, the group voltage, charging and discharging current, ambient temperature, and individual cell voltage, internal resistance, and temperature are collected online, and double calibration is performed to obtain a set of calibration parameters. On a fast time scale, the calibration parameter set is subjected to improved ampere-hour integration and extended Kalman filtering based on the Thevenin equivalent circuit model, and the two estimation results are weighted and fused to obtain the state of charge value; on a slow time scale, the cell internal resistance, available capacity and cycle number in the calibration parameter set are used to form a health feature vector, which is input into a pre-trained BP neural network to obtain the health state value. The confidence level of the state of charge is calculated based on the error covariance of the extended Kalman filter, and the confidence level of the health state is calculated based on the prediction confidence interval of the BP neural network. When the confidence level of the state of charge is lower than the preset confidence level of the state of charge threshold, the available capacity reference is corrected with the health status value and the state of charge value is corrected back. When the confidence level of the health status is lower than the preset confidence threshold of the health status, the health feature vector is corrected in reverse by the estimation deviation of the state of charge value to obtain the corrected state of charge value and health status value. When the dispersion of the individual cell voltage or individual cell internal resistance in the calibration parameter set is greater than the preset dispersion threshold, the health status value falls into the preset maintenance range, or the cumulative duration of the confidence level being lower than the threshold reaches the set duration, online capacity verification based on DC-DC boost is triggered to obtain the measured capacity of the capacity verification. The parameters of the Thevenin equivalent circuit model are updated with the measured capacity of the core, and the BP neural network is incrementally learned. At the same time, the error covariance matrix and the prediction confidence interval are reset to the initial state based on the measured capacity of the core.

2. The method for online estimation of battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, The dual calibration includes hardware calibration based on a reference source and software compensation for zero-point and temperature drift; and the gain calibration error, zero-point offset, and temperature drift compensation residual obtained from the calibration are used as quantitative values ​​of acquisition uncertainty to adjust the process noise covariance and observation noise covariance of the extended Kalman filter online.

3. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, The calculation of the state of charge confidence includes: taking the variance of the component corresponding to the state of charge in the error covariance matrix as the basic confidence metric; when the battery is operating in the plateau region of the open circuit voltage-state of charge curve, the basic confidence metric is reduced in weight according to the voltage observation sensitivity of the plateau region.

4. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, The confidence level of the health status is determined based on the variance of the BP neural network's predictions of multiple perturbations of the same health feature vector. The larger the prediction variance, the lower the confidence level of the health status. Both the confidence threshold of the charged state and the confidence threshold of the health status have hysteresis intervals. The confidence level of the charged state or the confidence level of the health status must cross the corresponding hysteresis interval before the corresponding correction is triggered or released.

5. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, The update cycle for the fast timescale is on the order of seconds, and the update cycle for the slow timescale is on the order of hours or days; the health status value is updated only when online capacity is completed, a deep charge / discharge event occurs, or the reverse correction is triggered.

6. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, For battery packs operating under long-term float charging, the set duration is determined by weighting the cumulative duration of confidence levels below the threshold with the float charging duration, in order to suppress estimation drift caused by the lack of core capacity anchor points under long-term shallow charge and discharge conditions.

7. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, The process of updating the Thevenin equivalent circuit model parameters with the measured capacity, incrementally learning the BP neural network, and resetting the error covariance matrix and the prediction confidence interval to an initial state based on the measured capacity specifically includes: calculating the capacity deviation between the measured capacity and the model predicted capacity; when the capacity deviation exceeds a set deviation threshold, identifying the Thevenin equivalent circuit model parameters online, and incrementally learning the BP neural network using data collected during the capacity grading process as new samples; resetting the error covariance matrix to the product of the initial matrix and the deviation correction coefficient, and resetting the prediction confidence interval to a benchmark interval determined by the measured capacity.

8. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, It also includes single-cell mutation isolation: when a sudden change in the single-cell voltage or single-cell internal resistance is identified, the parameters of that single-cell are isolated from the health feature vector used for the whole group estimation, and an estimation model is built for that single-cell separately; when the single-cell is detected to have been replaced, the parameters of its single-cell estimation model are reset.

9. The online estimation method for battery charge and state of health based on multi-parameter fusion according to claim 1, characterized in that, It also includes online maintenance: when the health status value falls into the preset maintenance range, the online capacity verification is first triggered. After verification, the DC-DC boost module is controlled to perform equalization charging or activation charging and discharging on the battery pack. After maintenance is completed, the state of charge value estimation on the fast time scale and the health status value estimation on the slow time scale, as well as the selective bidirectional coupling correction triggered by the confidence threshold, are re-executed.

10. A method for online estimation of battery charge and state of health based on multi-parameter fusion according to any one of claims 1-9, characterized in that, It also includes tiered early warning: the final state of charge value and its confidence level, as well as the health status value and its confidence level, are output together; When the state of charge value exceeds the preset normal range of the state of charge and / or the health value exceeds the preset normal range of the health, and the corresponding confidence level is not lower than the corresponding preset confidence threshold, the preset warning level corresponding to the over-limit item and the degree of over-limit is directly output. When the state of charge value and / or the health status value exceed the preset normal range, but the corresponding confidence level is lower than the corresponding preset confidence threshold, the online capacity verification is first triggered to obtain the measured capacity. The model parameters are then updated with the measured capacity and the confidence level is reset. Finally, the warning level is determined based on the re-estimated state of charge value and / or health status value.

Images