Dynamic adjustment method for safety threshold of reliable battery management system

Abstract

The invention discloses a dynamic adjustment method for a safety threshold of a reliable battery management system, and relates to the technical field of battery management. Acquiring real-time operation data of the battery and the aging degree of the battery cell; calculating a battery voltage normal operation threshold interval through a preset Gaussian process regression model to obtain a voltage threshold; taking the threshold value as a safety boundary, and respectively calculating a battery unified threshold value and a battery cell single body differential threshold value by combining the aging degree of the battery cell and the real-time operation data except the voltage to obtain an initial threshold value set; after the initial threshold value set is calibrated, the temperature and the voltage are predicted through an LSTM model; performing risk assessment according to the calibration threshold set and the predicted value to obtain a safety risk value; and adjusting the calibration threshold set according to the security risk value to obtain a security threshold set. According to the method, real-time, aging and prediction data are fused to dynamically adjust the threshold value, the full-period safety of the battery is guaranteed, the performance utilization rate is improved by adapting to different working conditions, unified management of the battery pack and differential protection of the single battery cells are considered, and thus the safety judgment accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of battery management technology, specifically relating to a method for dynamically adjusting the safety threshold of a reliable battery management system. Background Technology

[0002] In the field of new energy, batteries are the core energy storage unit, and their safety performance and service life directly determine the operational reliability of terminal products such as energy storage systems and electric vehicles. The battery management system is the core component that ensures the safe and stable operation of batteries. The safety threshold is the key judgment basis for the battery management system to realize overcharge, over-discharge, overcurrent and over-temperature protection. The safety threshold of traditional battery management systems mostly adopts fixed values, which are usually determined based on the nominal parameters of the battery at the time of manufacture or test data under standard laboratory conditions, and remain constant throughout the battery's entire life cycle.

[0003] Patent application CN120327265A discloses a battery thermal runaway and safety threshold dynamic adjustment system, comprising: a data acquisition module for acquiring battery state parameters, environmental parameters, and vehicle operating conditions; a data preprocessing module for preprocessing the acquired data; an initial safety threshold setting module for calculating the initial risk threshold of battery thermal runaway using statistical rules based on a large amount of historical sample data, serving as a judgment standard for the initial deployment stage; a dynamic threshold adjustment module for dynamically and individually adjusting the initial risk threshold according to a Bayesian optimization algorithm; and a thermal runaway prediction model module for predicting the battery risk level for the next 24 hours and the battery temperature for the next hour by combining traditional machine learning and deep learning models.

[0004] However, this method first sets the initial threshold based on statistical rules without deeply integrating the core physicochemical properties of the battery with the thermal runaway failure mechanism, resulting in the initial benchmark lacking physical rational support. Furthermore, the subsequent dynamic adjustment process fails to specifically correct the inherent defects of the initial benchmark, causing the threshold adjustment to always revolve around the initial value that lacks physical basis. Ultimately, it is difficult to accurately match the actual safety boundary of the battery, resulting in deviations in the dynamically adjusted threshold, thus affecting the accuracy of safety judgment. Summary of the Invention

[0005] The purpose of this invention is to solve the problem that the initial threshold set by statistical rules lacks support from the battery's physicochemical properties and failure mechanisms, and that dynamic adjustment does not specifically correct the defects of the initial benchmark, resulting in the threshold being difficult to match the actual safety boundary of the battery. Therefore, this invention proposes a dynamic adjustment method for the safety threshold of a reliable battery management system.

[0006] This invention proposes a method for dynamically adjusting the safety threshold of a reliable battery management system, the method comprising:

[0007] Acquire real-time operating data and cell aging status of the battery; the real-time operating data includes voltage, current, state of charge, temperature, and power.

[0008] The voltage confidence threshold is obtained by calculating the normal operating threshold range of the battery voltage using a preset Gaussian process regression model.

[0009] Using the voltage confidence threshold as the safety boundary, an initial threshold set is obtained by calculating the battery's unified safety threshold and the cell-specific threshold based on the temperature value, the current value, the charge state, the power, and the cell aging degree.

[0010] The initial threshold set is calibrated to obtain a calibration threshold set;

[0011] The predicted temperature and voltage values ​​are obtained by using an LSTM model.

[0012] A safety risk value is obtained by performing a predictive risk assessment on the battery based on the calibration threshold set, the predicted temperature value, and the predicted voltage value.

[0013] The safety threshold set is obtained by adjusting the calibration threshold set according to the safety risk value.

[0014] Optionally, the training process of the preset Gaussian process regression model includes:

[0015] Obtain historical operating datasets of multiple independent battery samples in a healthy state, and construct a training data point set based on the dataset; each training data point in the training data point set includes voltage value, current value, charge state and temperature value;

[0016] Assign a trainable memory unit parameter vector to each battery sample;

[0017] For each data point in the training data point set, the current value, charge state and temperature value contained therein are concatenated with the memory cell parameter vector of the battery sample corresponding to that data point to obtain the memory enhancement input feature set;

[0018] Using the memory-enhanced feature set as input and voltage value as output, a joint optimization objective function is constructed; the joint optimization objective function is a fusion function of the negative logarithmic marginal likelihood term of all training data points and the distribution consistency regularization term of all memory unit parameter vectors;

[0019] The optimal hyperparameters and the optimal memory unit parameter vector are obtained by minimizing the joint optimization objective function using a preset gradient coupling optimization algorithm.

[0020] The optimal hyperparameters and the optimal memory cell parameter vector are input into the Gaussian process regression model to obtain the preset Gaussian process regression model.

[0021] Optionally, constructing a joint optimization objective function using the memory-enhanced feature set as input and voltage values ​​as output includes:

[0022] The joint optimization objective function is L = 0.5y T ×K -1 ×y+0.5ln|K|+(n / 2)×ln(2π)+β×∑||m i -μ|| 2 ;

[0023] Where L is the joint optimization objective function value, y is the actual voltage value vector of all training data points, and K is the kernel matrix constructed based on the memory-enhanced input feature set. -1 Let |K| be the inverse of the kernel matrix, |K| be the determinant of the kernel matrix, n be the number of training data points, β be the regularization coefficient, and m be the inverse of the kernel matrix. i Let be the memory cell parameter vector of the i-th battery sample, and μ be the mean of all memory cell parameter vectors.

[0024] Optionally, the working principle of the preset gradient coupling optimization algorithm includes:

[0025] Calculate the first gradient vector of the hyperparameters and the second gradient vector of all memory cell parameter vectors according to the joint optimization objective function;

[0026] The first gradient vector and the second gradient vector are concatenated to obtain the joint gradient vector;

[0027] A linear transformation is performed on the joint gradient vector to obtain the attention weight vector;

[0028] The first gradient vector and the second gradient vector are weighted and combined according to the attention weight vector to obtain the second coupling vector and the first coupling vector, respectively.

[0029] The first updated gradient is obtained by weighting the first gradient vector and the first coupling vector.

[0030] The second updated gradient is obtained by weighted summation of the second gradient vector and the second coupling vector.

[0031] Based on the minimization objective of the joint optimization objective function, the optimal hyperparameters and the optimal memory cell parameter vector are obtained by performing gradient descent operations on the hyperparameters and the memory cell parameter vector according to the first update gradient and the second update gradient, respectively.

[0032] Optionally, an initial threshold set is obtained by calculating a uniform safety threshold and a cell-specific threshold based on the temperature value, the current value, the state of charge, the power, and the cell aging degree, respectively. This includes:

[0033] A two-layer time-series graph network model is constructed with the battery as the global node and each individual cell as the local node.

[0034] Based on the temperature value, the current value, the charge state, and the power, the dynamic coupling weights between the global node and each local node, and between each local node and each other, are calculated to obtain a dynamic coupling weight set;

[0035] In the time-series network model, the voltage confidence threshold is input into the first risk quantization function to calculate the initial risk value of the global node, and the cell aging degree of each cell is input into the second risk quantization function to calculate the initial risk value of each local node.

[0036] Based on the dynamic coupling weight set, the initial risk values ​​of all nodes in the time sequence graph network model are calculated in multiple rounds of coupled iterations until the absolute value of the change in the risk value of all nodes in two adjacent iterations is less than the preset convergence threshold.

[0037] Risk values ​​are extracted and transformed from the global nodes and local nodes in the time-series graph network model to obtain an initial threshold set.

[0038] Optionally, the first risk quantification function and the second risk quantification function include:

[0039] The first risk quantification function is Where R1 is the initial risk potential value of the global node, V1 is the rated battery voltage value, V2 is the voltage confidence threshold, V3 is the real-time voltage value, and k1 is the first normalization coefficient.

[0040] The second risk quantification function is Where S represents the health status of a single battery cell and its value ranges from (0,1). S represents the decay rate of a single cell, α represents the aging rate sensitivity coefficient, and k2 represents the first normalization coefficient.

[0041] Optionally, calibrating the initial threshold set to obtain a calibration threshold set includes:

[0042] Obtain the actual safety boundary offset under historical operating conditions;

[0043] Calculate the uniform calibration factor and the differential calibration factor based on the actual safety boundary offset;

[0044] The unified security threshold in the initial threshold set is calibrated according to the unified calibration factor to obtain the calibrated unified security threshold;

[0045] The calibration differential threshold is obtained by calibrating the differential threshold of each individual cell in the initial threshold set according to the differential calibration factor;

[0046] The calibration threshold set is obtained by statistically analyzing the uniform safety threshold and all differential calibration thresholds.

[0047] Optionally, the safety risk value obtained by performing a predictive risk assessment of the battery based on the calibration threshold set, the predicted temperature value, and the predicted voltage value includes:

[0048] If both the predicted voltage value and the predicted temperature value are greater than the unified safety threshold of the calibration threshold set, then the degree of deviation is calculated for the voltage and temperature respectively to obtain the voltage deviation risk value and the temperature deviation risk value.

[0049] The individual cell consistency risk coefficient is obtained by calculating the parameter dispersion between cells based on the numerical distribution characteristics of the differential thresholds of each cell in the calibration threshold set.

[0050] The safety risk value is obtained by weighting and fusing the voltage deviation risk value, the temperature deviation risk value, and the individual unit consistency risk coefficient.

[0051] The beneficial effects of this invention are as follows: This invention proposes a dynamic adjustment method for the safety threshold of a reliable battery management system. First, real-time operating data such as battery voltage, current, state of charge, temperature, and power, as well as cell aging, are collected. A voltage confidence threshold is generated using a preset Gaussian process regression model. Then, using this threshold as a safety boundary, a unified battery safety threshold and a cell-specific differentiated threshold are calculated based on the aforementioned data and aging, forming an initial threshold set and completing calibration. Simultaneously, an LSTM model is used to predict temperature and voltage values. Based on the calibrated threshold set and the predicted data, a predictive risk assessment of the battery is conducted to obtain a safety risk value. Finally, the calibrated threshold set is adjusted based on the safety risk value to determine the final safety threshold set. This method integrates real-time operating data, cell aging status, and predicted data to achieve dynamic adjustment of the safety threshold, ensuring the safe operation of the battery throughout its entire life cycle, adapting to different operating conditions to improve performance utilization, balancing unified battery pack management and cell-specific protection, and improving the accuracy of safety judgments. Attached Figure Description

[0052] The present invention will now be further described with reference to the accompanying drawings.

[0053] Figure 1 This is a flowchart illustrating a method for dynamically adjusting the safety threshold of a reliable battery management system, as provided in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] 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.

[0056] This invention provides a method for dynamically adjusting the safety threshold of a reliable battery management system. See also... Figure 1 , Figure 1 A flowchart illustrating a method for dynamically adjusting the safety threshold of a reliable battery management system, provided in an embodiment of the present invention. The method includes the following steps:

[0057] S101, acquires real-time operating data of the battery and cell aging status;

[0058] S102, the voltage confidence threshold is obtained by calculating the normal operating threshold range of the battery voltage through a preset Gaussian process regression model.

[0059] S103, using the voltage confidence threshold as the safety boundary, calculates the unified safety threshold of the battery and the individual cell-specific threshold based on the temperature value, current value, charge state, power and cell aging degree to obtain the initial threshold set;

[0060] S104, calibrate the initial threshold set to obtain the calibration threshold set;

[0061] S105, the predicted temperature and voltage values ​​are obtained by using the LSTM model;

[0062] S106, a predictive risk assessment of the battery is performed based on the calibration threshold set, predicted temperature value, and predicted voltage value to obtain a safety risk value;

[0063] S107, The safety threshold set is obtained by adjusting the calibration threshold set according to the safety risk value;

[0064] The real-time operating data includes voltage, current, charge status, temperature, and power.

[0065] This invention provides a method for dynamically adjusting the safety threshold of a reliable battery management system. First, real-time operating data such as battery voltage, current, state of charge, temperature, and power, as well as cell aging, are collected. A voltage confidence threshold is calculated using a pre-set Gaussian process regression model. Then, using this voltage confidence threshold as a safety boundary, and combining the collected real-time operating data and cell aging, a unified safety threshold and a cell-specific differentiated threshold are calculated to construct an initial threshold set and complete the calibration operation. Simultaneously, an LSTM model is used to predict battery temperature and voltage values. Based on the calibrated threshold set and the predicted temperature and voltage data, a predictive risk assessment of the battery is performed, generating a corresponding safety risk value. Finally, the calibrated threshold set is dynamically adjusted based on this safety risk value to determine the final safety threshold set. This method integrates real-time battery operating data, cell aging status, and predicted data to achieve dynamic adjustment of the safety threshold. This ensures the operational safety of the battery throughout its entire lifecycle, adapts to different operating conditions to improve battery performance utilization, and balances unified control of the battery pack with differentiated protection for individual cells, effectively improving the accuracy of safety judgments.

[0066] In one implementation, the above steps anchor the safety benchmark with a voltage confidence threshold, which can accurately define the normal operating range of the battery voltage and effectively avoid the problem of insufficient adaptability to voltage fluctuations by traditional fixed thresholds. On this basis, a differentiated threshold design that integrates multi-dimensional operating parameters and cell aging characteristics ensures the overall operating safety of the battery pack while taking into account the performance differences of individual cells. Combined with the dynamic adjustment mechanism of predictive risk assessment, the battery operating risks can be predicted in advance and the thresholds can be optimized. Ultimately, a balanced improvement in safety and range throughout the battery's entire life cycle is achieved, significantly reducing the probability of safety hazards such as overcharging and over-discharging.

[0067] In one implementation, the logic for adjusting the calibration threshold set to obtain the safety threshold set is to establish a dynamic closed loop of risk assessment and threshold feedback to ensure that the threshold always adapts to the real-time safety status of the battery. The safety risk value serves as the core feedback signal, and its magnitude directly reflects the degree of matching between the current calibration threshold and the actual safety requirements of the battery: if the safety risk value is higher than the preset risk threshold, it indicates that the current threshold is too lenient and cannot effectively warn of potential risks, and the warning sensitivity needs to be improved by tightening the threshold; if the safety risk value is lower than the preset risk threshold, it indicates that the current threshold is too strict and may lead to unnecessary operational restrictions, and the threshold needs to be appropriately relaxed; if the safety risk value is within a reasonable range, the threshold remains unchanged.

[0068] In one embodiment, the training process of the pre-defined Gaussian process regression model includes:

[0069] Obtain historical operating datasets of multiple independent battery samples in a healthy state, and construct a training data point set based on the dataset; each training data point in the training data point set includes voltage value, current value, charge state and temperature value;

[0070] Assign a trainable memory unit parameter vector to each battery sample;

[0071] For each data point in the training data point set, the current value, charge state and temperature value contained therein are concatenated with the memory cell parameter vector of the battery sample corresponding to that data point to obtain the memory enhancement input feature set;

[0072] Using the memory-enhanced feature set as input and voltage value as output, a joint optimization objective function is constructed. The joint optimization objective function is a fusion function of the negative log-marginal likelihood term of all training data points and the distribution consistency regularization term of all memory unit parameter vectors.

[0073] The optimal hyperparameters and the optimal memory unit parameter vector are obtained by minimizing the joint optimization objective function using a pre-defined gradient coupling optimization algorithm.

[0074] The optimal hyperparameters and the optimal memory cell parameter vector are input into the Gaussian process regression model to obtain the preset Gaussian process regression model.

[0075] In one implementation, the logic for constructing the training data point set is to provide high-quality, highly correlated training samples for the Gaussian process regression model, ensuring that the model can learn the operating rules under the healthy state of the battery. First, multiple independent battery samples are selected, and their historical operating data under healthy state S∈(0.8,1], without aging damage and with stable performance are collected. The core is to extract the corresponding data sets of the four key parameters of voltage, current, charge state, and temperature in each sample, because these four parameters directly determine the voltage response characteristics of the battery. Each set of data is defined as a training data point, forming a training data point set. Since even batteries of the same model may have different operating characteristics due to slight differences in manufacturing process and initial state, by assigning an independent memory unit parameter vector to each sample, the model can specifically learn the unique operating rules of each sample during training. This vector is essentially a sample-specific feature encoding.

[0076] In one implementation, the core of the training scheme for optimizing the traditional Gaussian process regression model is to add a trainable memory unit parameter vector to each battery sample. By jointly optimizing the model hyperparameters and memory unit parameters, the individual characteristics of a single battery are preserved, overfitting between samples is avoided, and the model parameters and sample features are adapted to each other. Ultimately, compared with existing models, it achieves the dual advantages of adapting to individual differences and improving prediction accuracy.

[0077] In one embodiment, constructing a joint optimization objective function using the memory-enhanced feature set as input and the voltage value as output includes:

[0078] The joint optimization objective function is L = 0.5y T ×K -1 ×y+0.5ln|K|+(n / 2)×ln(2π)+β×∑||m i -μ|| 2 Where L is the joint optimization objective function value, y is the actual voltage value vector of all training data points, and K is the kernel matrix constructed based on the memory-enhanced input feature set. -1 Let |K| be the inverse of the kernel matrix, |K| be the determinant of the kernel matrix, n be the number of training data points, β be the regularization coefficient, and m be the inverse of the kernel matrix. i Let be the memory cell parameter vector of the i-th battery sample, and μ be the mean of all memory cell parameter vectors.

[0079] In one implementation, ln(2π) is an inherent constant term of the probability density function of the Gaussian distribution, derived from the mathematical derivation of the negative log marginal likelihood in Gaussian process regression. K is used as the kernel matrix for constructing the memory-enhanced feature set to quantify the similarity between features. The value of β needs to be adjusted according to the number of samples and the feature dimension to balance the fitting accuracy and regularization strength, and can be 0.05.

[0080] In one implementation, the function quantifies the model's predictive performance and generalization ability into a single objective value, providing a clear minimization objective for subsequent gradient coupling optimization algorithms, ensuring that the trained model is both accurate and stable.

[0081] In one embodiment, the working principle of the preset gradient coupling optimization algorithm includes:

[0082] Calculate the first gradient vector of the hyperparameters and the second gradient vector of all memory cell parameter vectors based on the joint optimization objective function;

[0083] The first gradient vector and the second gradient vector are concatenated to obtain the joint gradient vector;

[0084] The attention weight vector is obtained by performing a linear transformation on the joint gradient vector;

[0085] The first gradient vector and the second gradient vector are weighted and combined according to the attention weight vector to obtain the second coupling vector and the first coupling vector, respectively.

[0086] The first updated gradient is obtained by weighting the first gradient vector and the first coupling vector;

[0087] The second updated gradient is obtained by weighted summation of the second gradient vector and the second coupling vector.

[0088] Based on the minimization objective of the joint optimization objective function, the optimal hyperparameters and optimal memory unit parameter vectors are obtained by performing gradient descent operations on the hyperparameters and memory unit parameter vectors according to the first update gradient and the second update gradient, respectively.

[0089] In one implementation, the attention weight vector is calculated based on the joint gradient representation through a feedforward neural network layer and a Softmax activation function. Each element of the attention weight vector is used to measure the coupling strength between the first gradient vector and the second gradient vector in different dimensions. The weighting coefficient of the second gradient vector and the second coupling vector is the adaptive step size coefficient of the current iteration.

[0090] In one implementation, this process achieves efficient global optimization of hyperparameters and memory unit parameter vectors through the coupling and collaborative updating of gradient information, solving the problem of fragmented parameter updates and low optimization efficiency in traditional gradient descent. First, for each term in the joint optimization objective function related to hyperparameters, partial derivatives are calculated with respect to each hyperparameter: First, it is determined that the hyperparameter is a parameter of the kernel function; then, the partial derivative of the kernel matrix with respect to that hyperparameter is calculated using a formula. Combining the inverse of the kernel matrix, the actual voltage value vector, etc., the partial derivative value corresponding to each hyperparameter is obtained. Finally, these partial derivative values ​​are arranged in order to form the first gradient vector. For the regularization term of the objective function, the partial derivatives of each memory cell parameter vector are calculated: considering the difference between the memory cell parameter vector and its mean, and combining the regularization coefficient and the number of samples, the partial derivative value corresponding to each memory cell parameter vector is calculated. Then, all partial derivative values ​​are arranged in order to form the second gradient vector. The two gradient vectors are then concatenated and linearly transformed to generate an attention weight vector. The core function of this vector is to dynamically determine the importance of the first and second gradient vectors in the current optimization step based on features such as the absolute value and direction of the gradient, thus achieving adaptive allocation of gradient weights. Next, the two original gradient vectors are weighted and combined according to the attention weight vector to obtain the first and second coupling vectors. Essentially, this allows the hyperparameter gradient to absorb effective information from the memory unit parameter gradients, and vice versa, achieving cross-parameter transfer and fusion of gradient information and avoiding the one-sidedness of updating a single parameter gradient. Finally, the original gradient and the corresponding coupling vector are weighted and summed to obtain the updated gradient, where... ω1 and ω2 are weight coefficients, ensuring that the updated gradient retains the core information of the original gradient while incorporating the complementary information of the coupled gradient. Finally, based on the gradient descent method, the hyperparameters are iteratively updated with the first update gradient and the memory unit parameter vector is iteratively updated with the second update gradient along the direction of minimizing the objective function L, until L converges. Through gradient coupling, the hyperparameters and memory unit parameters are co-optimized, avoiding local optima caused by single parameter updates, improving the efficiency and accuracy of model training, and finally obtaining optimal parameters with strong adaptability and good generalization ability.

[0091] In one embodiment, an initial threshold set is obtained by calculating a uniform safety threshold for the battery and a cell-specific threshold based on temperature, current, state of charge, power, and cell aging, respectively.

[0092] A two-layer time-series graph network model is constructed with the battery as the global node and each individual cell as the local node.

[0093] The dynamic coupling weight set is obtained by calculating the dynamic coupling weights between the global node and each local node, and between each local node, based on the temperature value, current value, charge state, and power.

[0094] In the time-series graph network model, the voltage confidence threshold is input into the first risk quantization function to calculate the initial risk value of the global node, and the cell aging degree of each cell is input into the second risk quantization function to calculate the initial risk value of each local node.

[0095] Based on the dynamic coupling weight set, the initial risk values ​​of all nodes in the time sequence graph network model are calculated in multiple rounds of coupled iterations until the absolute value of the change in the risk value of all nodes in two adjacent iterations is less than the preset convergence threshold.

[0096] Risk values ​​are extracted and transformed from global nodes and local nodes in the time series graph network model to obtain an initial threshold set.

[0097] In one implementation, the unified safety thresholds in the initial threshold set are voltage threshold, temperature threshold, and current threshold, the individual cell-specific threshold is the upper limit of the charging cut-off voltage for each cell, and the preset convergence threshold is set by the technician, for example, 0.05.

[0098] In one implementation, the entire battery is defined as a global node, its core function being to integrate and transmit system-level safety states, serving as the core calculation carrier for the safety threshold of the entire battery system. Each individual battery cell is defined as a local node to accurately capture the individual characteristics of each cell due to aging, manufacturing process differences, etc., enabling a separate characterization of the individual cell's safety state. A pre-defined similarity weight model based on a Gaussian kernel function is used to transform the real-time values ​​of each parameter into the correlation strength between the global node and each local node, and between each local node: for example, during high-current discharge, the current interaction between adjacent cells increases, corresponding to an increase in the coupling weight between local nodes; under high battery health conditions, the constraint of the global node on local nodes increases, corresponding to an increase in the coupling weight between the global and local nodes; the resulting dynamic coupling weight set ensures that the correlation strength between nodes can be adjusted in real time according to operating conditions, allowing... Subsequent risk iterations are more aligned with actual operating scenarios. The next iteration focuses on risk transmission and mutual correction. The global node integrates system-level risks and distributes them to local nodes based on dynamic coupling weights, correcting the risk values ​​of local nodes. For example, when the system as a whole is under high load, the global node transmits a high-risk benchmark to all local nodes, raising the starting point for individual risk assessment of each cell. Simultaneously, each local node transmits its own individual risks through dynamic coupling weights between nodes. For instance, if a cell is severely aged, its high-risk value is transmitted to adjacent cells and aggregated and fed back to the global node, correcting the system-level risk value until a stable state is reached. This avoids risk values ​​deviating from reality due to insufficient iteration, ensuring that the final node risk value is the result of global and local interaction and dynamic adaptation. Finally, it is converted into an initial threshold through a preset mapping relationship between risk values ​​and safety thresholds.

[0099] In one embodiment, the first risk quantification function and the second risk quantification function include:

[0100] The first risk quantification function is: Where R1 is the initial risk potential value of the global node, V1 is the rated battery voltage value, V2 is the voltage confidence threshold, V3 is the real-time voltage value, and k1 is the first normalization coefficient.

[0101] The second risk quantification function is: Where S represents the health status of a single battery cell and its value ranges from (0,1). S represents the decay rate of a single cell, α represents the aging rate sensitivity coefficient, and k2 represents the first normalization coefficient.

[0102] In one implementation, α is set by the technician, for example, 0.25, k1 is used to adjust the output to the target numerical range, for example, 3, and the value of k2 is a coefficient dynamically determined according to the actual system's magnitude matching requirements to match the numerical range of R2, which can be 2.

[0103] In one implementation, the first risk quantification function constructs a nonlinear mapping through the relative relationship of voltage parameters, amplifying the risk impact of voltage approaching the safety boundary. First, V2-V3 is calculated; this difference directly quantifies the safety margin between the real-time voltage and the safety boundary. The smaller the difference, the closer the voltage is to or beyond the normal operating range, and the higher the safety risk. Then, V1 is divided by this difference. Essentially, this utilizes the benchmark characteristics of the rated voltage to transform the absolute difference of the safety margin into a relative proportion, amplifying the degree of voltage deviation. For example, when V2-V3 approaches 0, the ratio increases sharply, intuitively reflecting the sharp increase in risk. Finally, it is multiplied by k1 for normalization, transforming the amplified ratio into a risk value R1 within the standard range, ensuring that the risk value meets both the quantitative assessment requirements and has intuitive risk level identification.

[0104] In one implementation, the second risk quantification function integrates static aging state and dynamic decay trend to construct a multi-dimensional aging risk quantification model. First, 1 / (1-S) is calculated, and a nonlinear mapping is constructed through the reciprocal relationship to amplify the impact of static aging degree. For example, when S decreases from 0.9 to 0.8, 1 / (1-S) increases from 10 to 12.5, significantly improving the risk contribution and accurately characterizing the decrease in safety redundancy caused by cell aging. Second, 1+α×(dS / dt) is calculated, and the impact of dynamic decay rate is amplified through α. For example, when the absolute value of dS / dt increases, this term increases, and the risk value increases synchronously, realizing a forward-looking consideration of aging trend and avoiding focusing only on the current aging state while ignoring future risks. Finally, the two results are multiplied together and multiplied by k2 to achieve coupled quantification of static and dynamic aging risks. At the same time, through normalization, it is transformed into a risk value R2 within a standard range to ensure that R2 can comprehensively and accurately reflect the aging-related safety risk potential of a single cell.

[0105] In one embodiment, calibrating the initial threshold set to obtain a calibration threshold set includes:

[0106] Obtain the actual safety boundary offset under historical operating conditions;

[0107] Calculate the uniform calibration factor and the differential calibration factor based on the actual safety boundary offset;

[0108] The calibrated unified security threshold is obtained by calibrating the unified security threshold in the initial threshold set according to the unified calibration factor;

[0109] The calibration differential threshold is obtained by calibrating the differential threshold of each cell in the initial threshold set according to the differential calibration factor;

[0110] The calibration threshold set is obtained by statistically calibrating the unified safety threshold and all calibration-differentiated thresholds.

[0111] In one implementation, the calibration factor calculation process involves first classifying historical offset data, grouping offsets corresponding to a globally unified safety threshold into one category, and classifying offsets corresponding to individual cell-specific thresholds separately for each cell. Then, the mean of the global offset data is calculated, and combined with a preset error tolerance range, a unified calibration factor is determined. If the mean offset is positive, it indicates the theoretical threshold is too strict, and the calibration factor is greater than 1; if the mean is negative, it indicates the theoretical threshold is too lenient, and the calibration factor is less than 1. Statistical analysis is then performed on the individual offset data of each cell, and combined with the cell's aging degree and historical operating characteristics, a differentiated calibration factor is determined to ensure that the calibration factor accurately matches the individual error characteristics of each cell. The core of the entire process is to achieve a mapping between historical errors and calibration factors through statistical analysis, allowing the calibration factor to correct both system-level errors and individual-level errors.

[0112] In one embodiment, the safety risk value obtained by performing a predictive risk assessment of the battery based on a calibration threshold set, a predicted temperature value, and a predicted voltage value includes:

[0113] If both the predicted voltage and predicted temperature values ​​are greater than the unified safety threshold in the calibration threshold set, then the degree of deviation is calculated for the voltage and temperature respectively to obtain the voltage deviation risk value and the temperature deviation risk value.

[0114] The individual cell consistency risk coefficient is obtained by calculating the parameter dispersion between cells based on the numerical distribution characteristics of the differential thresholds of each cell in the calibration threshold set.

[0115] The safety risk value is obtained by weighting and fusing the voltage deviation risk value, temperature deviation risk value, and individual unit consistency risk coefficient.

[0116] In one implementation, the process achieves precise monitoring of the core safety indicators of the battery while also taking into account the long-term safety hazard of cell consistency, making risk assessment more comprehensive and accurate. It can provide early warning of potential safety issues while avoiding over-warning, effectively improving the scientific nature and efficiency of battery safety management.

[0117] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A method for dynamically adjusting the safety threshold of a reliable battery management system, characterized in that, The method includes: Acquire real-time operating data and cell aging status of the battery; the real-time operating data includes voltage, current, state of charge, temperature, and power. The voltage confidence threshold is obtained by calculating the normal operating threshold range of the battery voltage using a preset Gaussian process regression model. Using the voltage confidence threshold as the safety boundary, an initial threshold set is obtained by calculating the battery's unified safety threshold and the cell-specific threshold based on the temperature value, the current value, the charge state, the power, and the cell aging degree. The initial threshold set is calibrated to obtain a calibration threshold set; The predicted temperature and voltage values ​​are obtained by using an LSTM model. A safety risk value is obtained by performing a predictive risk assessment on the battery based on the calibration threshold set, the predicted temperature value, and the predicted voltage value. The safety threshold set is obtained by adjusting the calibration threshold set according to the safety risk value.

2. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 1, characterized in that, The training process of the preset Gaussian process regression model includes: Obtain historical operating datasets of multiple independent battery samples in a healthy state, and construct a training data point set based on the dataset; each training data point in the training data point set includes voltage value, current value, charge state and temperature value; Assign a trainable memory unit parameter vector to each battery sample; For each data point in the training data point set, the current value, charge state and temperature value contained therein are concatenated with the memory cell parameter vector of the battery sample corresponding to that data point to obtain the memory enhancement input feature set; Using the memory enhancement feature set as input and voltage value as output, a joint optimization objective function is constructed; the joint optimization objective function is a fusion function of the negative log marginal likelihood term of all training data points and the distribution consistency regularization term of all memory unit parameter vectors; The optimal hyperparameters and the optimal memory unit parameter vector are obtained by minimizing the joint optimization objective function using a preset gradient coupling optimization algorithm. The optimal hyperparameters and the optimal memory cell parameter vector are input into the Gaussian process regression model to obtain the preset Gaussian process regression model.

3. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 2, characterized in that, Using the memory-enhanced feature set as input and voltage values ​​as output, a joint optimization objective function is constructed, comprising: The joint optimization objective function is L = 0.5y T ×K -1 ×y+0.5ln|K|+(n / 2)×ln(2π)+β×∑||m i -μ|| 2 Where L is the joint optimization objective function value, y is the actual voltage value vector of all training data points, and K is the kernel matrix constructed based on the memory-enhanced input feature set. -1 Let |K| be the inverse of the kernel matrix, |K| be the determinant of the kernel matrix, n be the number of training data points, β be the regularization coefficient, and m be the inverse of the kernel matrix. i Let be the memory cell parameter vector of the i-th battery sample, and μ be the mean of all memory cell parameter vectors.

4. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 2, characterized in that, The working principle of the preset gradient coupling optimization algorithm includes: Calculate the first gradient vector of the hyperparameters and the second gradient vector of all memory cell parameter vectors according to the joint optimization objective function; The first gradient vector and the second gradient vector are concatenated to obtain the joint gradient vector; A linear transformation is performed on the joint gradient vector to obtain the attention weight vector; The first gradient vector and the second gradient vector are weighted and combined according to the attention weight vector to obtain the second coupling vector and the first coupling vector, respectively. The first updated gradient is obtained by weighting the first gradient vector and the first coupling vector. The second updated gradient is obtained by weighted summation of the second gradient vector and the second coupling vector. Based on the minimization objective of the joint optimization objective function, the optimal hyperparameters and the optimal memory cell parameter vector are obtained by performing gradient descent operations on the hyperparameters and the memory cell parameter vector according to the first update gradient and the second update gradient, respectively.

5. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 1, characterized in that, Based on the temperature value, current value, state of charge, power, and cell aging degree, an initial threshold set is obtained by calculating the battery's uniform safety threshold and the cell-specific threshold, including: A two-layer time-series graph network model is constructed with the battery as the global node and each individual battery cell as the local node. The dynamic coupling weight set is obtained by calculating the dynamic coupling weights between the global node and each local node, and between each local node, based on the temperature value, the current value, the charge state, and the power. In the time-series network model, the voltage confidence threshold is input into the first risk quantization function to calculate the initial risk value of the global node, and the cell aging degree of each cell is input into the second risk quantization function to calculate the initial risk value of each local node. Based on the dynamic coupling weight set, the initial risk values ​​of all nodes in the time sequence graph network model are calculated in multiple rounds of coupling iteration until the absolute value of the change in the risk value of all nodes in two adjacent iterations is less than the preset convergence threshold. Risk values ​​are extracted and transformed from the global nodes and local nodes in the time-series graph network model to obtain an initial threshold set.

6. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 4, characterized in that, The first risk quantification function and the second risk quantification function include: The first risk quantification function is Where R1 is the initial risk potential value of the global node, V1 is the rated battery voltage value, V2 is the voltage confidence threshold, V3 is the real-time voltage value, and k1 is the first normalization coefficient. The second risk quantification function is Where S represents the health status of a single battery cell and its value ranges from (0,1). S represents the decay rate of a single cell, α represents the aging rate sensitivity coefficient, and k2 represents the first normalization coefficient.

7. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 1, characterized in that, The calibration threshold set is obtained by calibrating the initial threshold set, including: Obtain the actual safety boundary offset under historical operating conditions; Calculate the uniform calibration factor and the differential calibration factor based on the actual safety boundary offset; The unified security threshold in the initial threshold set is calibrated according to the unified calibration factor to obtain the calibrated unified security threshold; The calibration differential threshold is obtained by calibrating the differential threshold of each individual cell in the initial threshold set according to the differential calibration factor; The calibration threshold set is obtained by statistically analyzing the uniform safety threshold and all differential calibration thresholds.

8. The method for dynamically adjusting the safety threshold of a reliable battery management system according to claim 1, characterized in that, A predictive risk assessment of the battery based on the calibration threshold set, the predicted temperature value, and the predicted voltage value yields a safety risk value, including: If both the predicted voltage value and the predicted temperature value are greater than the unified safety threshold of the calibration threshold set, then the degree of deviation is calculated for the voltage and temperature respectively to obtain the voltage deviation risk value and the temperature deviation risk value. The individual cell consistency risk coefficient is obtained by calculating the parameter dispersion between cells based on the numerical distribution characteristics of the differential thresholds of each cell in the calibration threshold set. The safety risk value is obtained by weighting and fusing the voltage deviation risk value, the temperature deviation risk value, and the individual unit consistency risk coefficient.

Images