Power lithium battery thermal runaway risk prediction method based on multivariate degradation information

Abstract

The invention discloses a power lithium battery thermal runaway risk prediction method based on multivariate degradation information, and the method comprises the following steps: carrying out the cleaning, denoising and feature extraction of historical operation data of a battery, and obtaining a multivariate feature vector used for degradation characterization; inputting the multivariate feature vector into a fusion model, establishing a mapping relation between the feature vector and a battery thermal stability parameter through the fusion model, and obtaining a degradation characterization parameter capable of characterizing the battery thermal stability; performing a linear transformation for a non-linear relationship between the degradation characterization parameter and the health state parameter to obtain a degradation index that changes approximately linearly in the feature space; and based on the linearized degradation index, establishing a random process model of degradation evolution on a selected time scale, and estimating model parameters by using degradation data.

Description

Technical Field

[0001] This invention relates to a method for predicting the thermal runaway risk of power lithium batteries based on multi-dimensional degradation information, and belongs to the field of battery testing technology. Background Technology

[0002] Lithium-ion batteries are the core components of new transportation vehicles such as electric vehicles and drones. Numerous cases have shown that thermal runaway of lithium-ion batteries is a significant cause of spontaneous combustion in electric vehicles and other vehicles. Concerns about the risk of thermal runaway in lithium-ion batteries have become a bottleneck restricting their wider application. Current research on lithium-ion battery thermal runaway mainly focuses on battery heat generation characteristics, thermal runaway mechanisms, and high-safety battery materials, with little attention paid to changes in thermal runaway risk caused by degradation in charge-discharge performance over cycles.

[0003] To date, research on lithium-ion battery thermal runaway has primarily focused on materials and thermal management strategies, such as developing highly stable electrode materials, optimizing electrolyte formulations, and improving the efficiency of heat dissipation systems. Furthermore, many researchers have dedicated themselves to elucidating the chain reaction mechanism and critical conditions of thermal runaway, using accelerated calorimetry (ARC) to determine key parameters such as the self-exothermic initiation temperature T1. However, most existing studies are based on fresh batteries or standard aged samples, lacking systematic modeling and quantitative analysis of how performance degradation caused by complex operating conditions in real-world use affects thermal stability evolution. Current research on lithium-ion battery thermal runaway mainly focuses on battery heat generation characteristics, thermal runaway mechanisms, and high-safety battery materials, with little attention paid to changes in thermal runaway risk caused by performance degradation during cyclic charging and discharging. Of particular note is that, due to the random operating conditions during lithium battery use, the degradation behavior of lithium batteries will manifest as a multi-mechanism coupled, path-dependent stochastic process. This leads to uncertainty in the "strength" of degradation that allows the battery to withstand external stress without losing control. Furthermore, the external triggering stress is also random, making it difficult to effectively quantify the probability of exceeding this "strength" (thermal runaway risk). This characteristic makes it difficult for traditional deterministic models to accurately capture the risk dynamics in real-world scenarios, thus failing to quantify the gradually increasing thermal runaway risk of lithium batteries with increasing cycle count.

[0004] The current research has the following main shortcomings: First, it is necessary to extract a fusion index from multi-source monitoring data that can reflect both the degradation state and the thermal stability; second, it is necessary to establish a probabilistic model that can simultaneously characterize the uncertainty of the degradation process and the randomness of external triggering stress.

[0005] For example, Chinese invention application CN202510819827.3 discloses a method and system for assessing the health status of lithium batteries, specifically relating to the field of battery health status assessment technology. It involves time-series alignment and structured preprocessing of multi-source operating data from multiple historical operating cycles of a target lithium battery to construct a structured dataset; constructing a spatiotemporal feature residual map based on residual mapping analysis to extract spatial heterogeneity indicators; generating regional degradation feature maps by combining spatial heterogeneity indicators; identifying heterogeneous aging patterns and generating classification results through high-dimensional feature embedding and evolution path clustering; assessing the health status level of the target lithium battery based on the regional degradation feature maps and heterogeneous aging pattern classification results; determining whether the battery has local potential thermal runaway risks based on the assessment results, and generating corresponding risk warning signals and safety handling suggestions. This method can accurately identify the nonlinear effects of lithium battery aging heterogeneity, effectively improving the accuracy of health status assessment and safety risk warning capabilities. However, it does not reference and fuse lithium battery performance degradation parameters, and the parameters considered in the dynamic mapping from performance degradation to thermal runaway risk are still insufficient, failing to fully realize reliable application throughout the entire life cycle of lithium batteries. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and to provide a method for predicting the thermal runaway risk of power lithium batteries based on multi-dimensional degradation information.

[0007] A method for predicting the thermal runaway risk of power lithium batteries based on multi-dimensional degradation information includes the following steps: The historical operating data of the battery is cleaned, denoised, and feature extracted to obtain a multivariate feature vector for degradation characterization; The multivariate feature vectors are input into the fusion model, and the mapping relationship between the feature vectors and the battery thermal stability parameters is established through the fusion model to obtain degradation characterization parameters that can characterize the battery thermal stability. To address the nonlinear relationship between degradation characterization parameters and health status parameters, a linearization transformation is performed to obtain degradation indices that exhibit approximately linear changes in the feature space. Based on the linearized degradation index, a stochastic process model of degradation evolution is established at a selected time scale, and the model parameters are estimated using degradation data. Based on the stochastic process model, thermal runaway is defined as the event when the degradation index first reaches or crosses a preset threshold. The corresponding first-reach time distribution is calculated to obtain the cumulative probability of thermal runaway risk over time. Treating the preset threshold as a random variable, the cumulative probability is marginalized or numerically solved based on its statistical distribution to obtain the overall thermal runaway risk that comprehensively considers internal degradation and external triggering stress.

[0008] This technical solution, through a multivariate degradation information fusion method, enables the model to indirectly obtain estimated values ​​of characterization parameters from readily available degradation data. This method effectively solves the problem that degradation characterization parameters of battery thermal stability cannot be directly measured online, laying a solid foundation for predicting the probability of thermal runaway risk based on the degradation trajectory of battery thermal stability characterization parameters. The introduction of a fusion model to stochastically model battery degradation fully characterizes the uncertainties in the lithium battery cycle aging process. Furthermore, considering the nondeterministic nature of external triggering stresses for thermal runaway, a risk model more consistent with engineering practice is established, enabling dynamic prediction of the probability of thermal runaway. Based on multivariate degradation information during actual lithium battery use, the probability of battery thermal runaway is effectively predicted. The resulting thermal runaway risk curve can provide a reference for battery health management and has significant practical application value.

[0009] Preferably, the steps for establishing the fusion model include: Feature filtering is performed on the operating data of power lithium batteries to construct multivariate feature vectors, and a neural network structure or other fusion model structure with physical constraints is designed. The model is trained and optimized based on multi-cycle degradation data; The correspondence between the health status parameters and the thermal stability parameters was established through experimental calibration. The fusion model is a physical information neural network. During the training process, the electrochemical-thermal coupling mechanism equation is incorporated as a physical constraint into the loss function to learn the mapping relationship by jointly minimizing the observation error and the physical residual.

[0010] This technical solution utilizes multidimensional feature vectors to filter features from the operating data of power lithium batteries, fully leveraging information such as voltage, current, temperature, and time to comprehensively reflect the battery's degradation state and operating conditions. This provides rich and highly relevant input data for subsequent model training, significantly improving the model's input representativeness and information completeness. The model not only learns from historical data but also follows the internal thermodynamic laws of the battery. This design effectively prevents overfitting and significantly improves the stability and physical consistency of predictions. By training and optimizing the model based on multi-cycle degradation data, the fusion model can capture the dynamic characteristics of the battery at different degradation stages, exhibiting good time-scale adaptability and maintaining prediction accuracy throughout the battery's entire lifespan.

[0011] Furthermore, the multivariate feature vector includes, but is not limited to: average charging current, current standard deviation, charging time, accumulated charge, curve slope, curve entropy, kurtosis, and skewness. These features are filtered through correlation analysis and / or feature selection methods and then used as input to the fusion model.

[0012] This technical solution can reflect the degradation characteristics of batteries at different charging and discharging stages from multiple dimensions such as electrical, temporal and statistical features. By using correlation analysis and feature selection methods to screen candidate features, redundant or low-correlation indicators are eliminated, and key features that contribute most to thermal stability parameters and health status are retained. This can improve the effectiveness of the fusion model input, reduce computational complexity and training time, and improve modeling efficiency.

[0013] Preferably, the linearization transformation step includes: A feature space linearization transformation is performed on the nonlinear correspondence between degradation characterization parameters and health status parameters, mapping the original nonlinear curve to an approximate straight line in the linear space to obtain a linearized degradation index. The time-scale transformation of the linearized degradation index sequence over time or cycle number is performed to obtain a degradation trajectory that changes approximately linearly over the transformed time scale. Wherein, the feature space linearization transformation and the time scale transformation are both reversible monotonic transformations, and the transformation form is selected from at least one of logarithmic, power function, square, log-time or affine linear mapping.

[0014] This technical solution enables the linearized degradation index to exhibit monotonically changing characteristics, effectively reducing the uncertainty of parameter estimation and avoiding model convergence difficulties caused by multiple solutions or local extrema of nonlinear curves, thereby improving the accuracy of model parameter identification and fitting stability. By transforming the time or cyclic axis of the degradation index, the originally non-uniform degradation process becomes approximately linear under the new time scale, facilitating the modeling of degradation evolution using linear stochastic processes (such as the Wiener process), and improving the model's adaptability and computational efficiency.

[0015] Preferably, the steps for establishing the stochastic process model include: The linearization degradation index is treated as a random variable that evolves over time or the number of cycles, and its evolution is assumed to follow a continuous-time stochastic process. Construct a stochastic process model that includes drift and diffusion terms to describe the average trend and stochastic fluctuations of degradation; The degradation increment is calculated based on the degradation observation data, and the drift coefficient and diffusion coefficient are obtained by maximum likelihood estimation or Bayesian estimation method. The independence and distribution characteristics of the model residuals are tested to verify the rationality of the model; The stochastic process model is preferably a linear Wiener process model, or can be a diffusion process, a Gamma process, or other stochastic process models suitable for degradation modeling.

[0016] This technical solution enables the dynamic evolution of battery degradation within a probabilistic framework, providing a unified and quantifiable mathematical description for risk prediction. By introducing drift and diffusion terms into the model, the deterministic trends and stochastic fluctuations of battery degradation can be described respectively, thus balancing long-term trends and short-term uncertainties in modeling and improving the model's fitting accuracy and dynamic response capability to actual degradation behavior. The stochastic process model can flexibly select forms such as linear Wiener processes, Gamma processes, or general diffusion processes according to different battery types and degradation mechanisms, thereby being compatible with different degradation paths and temporal characteristics and improving the model's applicability in multiple application scenarios.

[0017] Preferably, the first crossing time distribution is derived from an inverse Gaussian distribution when the stochastic process model is a linear Wiener process, and the first arrival time distribution is obtained by corresponding analytical or numerical methods under other specific forms of the stochastic process model.

[0018] This technical solution allows the first-crossing time distribution to be analytically derived as an inverse Gaussian distribution when the stochastic process model is a linear Wiener process. This provides a clear mathematical expression for the time distribution of thermal runaway risk, facilitating the calculation of cumulative risk probability and remaining safe lifetime indicators within an analytical framework, thus improving the physical interpretability and computational transparency of the model results. Beyond Wiener processes, appropriate analytical or numerical methods can be selected to solve the first-crossing time distribution based on the characteristics of the adopted stochastic process model (such as Gamma processes, diffusion processes, etc.), thereby establishing a general risk calculation framework. This avoids dependence on a single model form and enhances the universality of the method. By employing analytical solutions based on inverse Gaussian distributions or numerical methods for nonlinear processes, the probability distribution of degradation indicators reaching thresholds can be accurately obtained. This approach is applicable to theoretical modeling scenarios and can also be extended to real-time risk calculation of actual operational data, ensuring the accuracy and real-time nature of the prediction results.

[0019] Preferably, the preset threshold is treated as a random variable and modeled according to a normal distribution or other appropriate distribution; The overall thermal runaway risk is obtained by Monte Carlo sampling of the threshold distribution and averaging or marginalizing the first-arrival time distribution for each sampling threshold. The sample size for Monte Carlo sampling can be adjusted according to the required accuracy and computing resources.

[0020] This technical solution effectively characterizes the randomness of external triggering factors such as ambient temperature fluctuations, differences in operating conditions, and manufacturing consistency deviations, making the risk model closer to actual working conditions and significantly improving the credibility and engineering applicability of the assessment results. It unifies and integrates internal degradation randomness and external triggering uncertainty within a probabilistic framework, thereby obtaining a more comprehensive and robust overall thermal runaway risk distribution.

[0021] Preferably, the method further includes an online update step: periodically or under event triggering in the BMS to retrain or incrementally identify the parameters of the fusion model and the stochastic process model online based on a sliding window or an incremental dataset, and update the estimation and early warning decision of the thermal runaway risk in real time accordingly.

[0022] This technical solution enables the model to continuously absorb new operational data and correct parameter deviations, ensuring that predictions consistently reflect the battery's current true degradation state and avoiding risk assessment biases caused by model aging. Batteries are affected by multiple factors during actual operation, including temperature, load, and aging rate variations. Through an online update mechanism, the model can continuously adapt to changing operating conditions, maintaining stability and generalization ability in risk predictions across different usage scenarios, thus exhibiting higher engineering applicability. As model parameters are updated, the calculated thermal runaway risk curve, risk level, or remaining safe life index are also dynamically adjusted, allowing the BMS to trigger appropriate safety strategies such as cooling, load reduction, and power-off in real time, improving the system's proactive safety capabilities.

[0023] The beneficial effects of this invention are as follows: Through a multivariate degradation information fusion method, the model can indirectly obtain estimated values ​​of characterization parameters from easily accessible degradation data. This method effectively solves the problem that degradation characterization parameters of battery thermal stability cannot be directly measured online, laying a solid foundation for predicting the probability of thermal runaway risk based on the degradation trajectory of battery thermal stability characterization parameters. The introduction of a fusion model to stochastically model battery degradation fully characterizes the uncertainties in the lithium battery cycle aging process. Furthermore, considering the nondeterministic nature of external triggering stress for thermal runaway, a risk model more consistent with engineering practice is established, achieving dynamic prediction of the probability of thermal runaway. Based on multivariate degradation information during actual lithium battery use, the probability of battery thermal runaway is effectively predicted. The resulting thermal runaway risk curve can provide a reference for battery health management and has significant practical application value. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0025] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is the relationship curve between T1 and SOH; Figure 3It is a degradation curve of thermal runaway risk characterization parameters obtained based on multi-source information fusion; Figure 4 It is the result of linearization of the thermal runaway risk characterization parameters; Figure 5 This is a schematic diagram of threshold distribution; Figure 6 This is the thermal runaway risk curve. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities or parameters with the same name but different names. It is clear that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of the present invention. Subsequent embodiments will not explain this in detail.

[0028] The directional and positional terms used in this invention, such as "up," "down," "front," "back," "left," "right," "inner," "outer," "top," "bottom," and "side," are merely for reference to the accompanying drawings. Therefore, the directional and positional terms used are for illustrating and understanding this invention, and not for limiting the scope of protection of this invention.

[0029] like Figure 1-6 The image shows an embodiment of the method for predicting the thermal runaway risk of a power lithium battery based on multi-dimensional degradation information according to the present invention. The method for predicting the thermal runaway risk of a power lithium battery based on multi-dimensional degradation information includes the following steps: The historical operating data of the battery is cleaned, denoised, and feature extracted to obtain a multivariate feature vector for degradation characterization; The multivariate feature vectors are input into the fusion model, and the mapping relationship between the feature vectors and the battery thermal stability parameters is established through the fusion model to obtain degradation characterization parameters that can characterize the battery thermal stability. To address the nonlinear relationship between degradation characterization parameters and health status parameters, a linearization transformation is performed to obtain degradation indices that exhibit approximately linear changes in the feature space. Based on the linearized degradation index, a stochastic process model of degradation evolution is established at a selected time scale, and the model parameters are estimated using degradation data. Based on the stochastic process model, thermal runaway is defined as the event when the degradation index first reaches or crosses a preset threshold. The corresponding first-reach time distribution is calculated to obtain the cumulative probability of thermal runaway risk over time. Treating the preset threshold as a random variable, the cumulative probability is marginalized or numerically solved based on its statistical distribution to obtain the overall thermal runaway risk that comprehensively considers internal degradation and external triggering stress.

[0030] This technical solution, through a multivariate degradation information fusion method, enables the model to indirectly obtain estimated values ​​of characterization parameters from readily available degradation data. This method effectively solves the problem that degradation characterization parameters of battery thermal stability cannot be directly measured online, laying a solid foundation for predicting the probability of thermal runaway risk based on the degradation trajectory of battery thermal stability characterization parameters. The introduction of a fusion model to stochastically model battery degradation fully characterizes the uncertainties in the lithium battery cycle aging process. Furthermore, considering the nondeterministic nature of external triggering stresses for thermal runaway, a risk model more consistent with engineering practice is established, enabling dynamic prediction of the probability of thermal runaway. Based on multivariate degradation information during actual lithium battery use, the probability of battery thermal runaway is effectively predicted. The resulting thermal runaway risk curve can provide a reference for battery health management and has significant practical application value.

[0031] The steps for establishing the fusion model include: Feature filtering is performed on the operating data of power lithium batteries to construct multivariate feature vectors, and a neural network structure or other fusion model structure with physical constraints is designed. The model is trained and optimized based on multi-cycle degradation data; The correspondence between the health status parameters and the thermal stability parameters was established through experimental calibration. The fusion model is a physical information neural network. During the training process, the electrochemical-thermal coupling mechanism equation is incorporated as a physical constraint into the loss function to learn the mapping relationship by jointly minimizing the observation error and the physical residual.

[0032] This technical solution utilizes multidimensional feature vectors to filter features from the operating data of power lithium batteries, fully leveraging information such as voltage, current, temperature, and time to comprehensively reflect the battery's degradation state and operating conditions. This provides rich and highly relevant input data for subsequent model training, significantly improving the model's input representativeness and information completeness. The model not only learns from historical data but also follows the internal thermodynamic laws of the battery. This design effectively prevents overfitting and significantly improves the stability and physical consistency of predictions. By training and optimizing the model based on multi-cycle degradation data, the fusion model can capture the dynamic characteristics of the battery at different degradation stages, exhibiting good time-scale adaptability and maintaining prediction accuracy throughout the battery's entire lifespan.

[0033] The multivariate feature vector includes, but is not limited to: average charging current, standard deviation of current, charging time, accumulated charge, curve slope, curve entropy, kurtosis and skewness. The features are filtered through correlation analysis and / or feature selection methods and then used as input to the fusion model.

[0034] This technical solution can reflect the degradation characteristics of batteries at different charging and discharging stages from multiple dimensions such as electrical, temporal and statistical features. By using correlation analysis and feature selection methods to screen candidate features, redundant or low-correlation indicators are eliminated, and key features that contribute most to thermal stability parameters and health status are retained. This can improve the effectiveness of the fusion model input, reduce computational complexity and training time, and improve modeling efficiency.

[0035] The linearization transformation step includes: A feature space linearization transformation is performed on the nonlinear correspondence between degradation characterization parameters and health status parameters, mapping the original nonlinear curve to an approximate straight line in the linear space to obtain a linearized degradation index. The time-scale transformation of the linearized degradation index sequence over time or cycle number is performed to obtain a degradation trajectory that changes approximately linearly over the transformed time scale. Wherein, the feature space linearization transformation and the time scale transformation are both reversible monotonic transformations, and the transformation form is selected from at least one of logarithmic, power function, square, log-time or affine linear mapping.

[0036] This technical solution enables the linearized degradation index to exhibit monotonically changing characteristics, effectively reducing the uncertainty of parameter estimation and avoiding model convergence difficulties caused by multiple solutions or local extrema of nonlinear curves, thereby improving the accuracy of model parameter identification and fitting stability. By transforming the time or cyclic axis of the degradation index, the originally non-uniform degradation process becomes approximately linear under the new time scale, facilitating the modeling of degradation evolution using linear stochastic processes (such as the Wiener process), and improving the model's adaptability and computational efficiency.

[0037] The steps for establishing the stochastic process model include: The linearization degradation index is treated as a random variable that evolves over time or the number of cycles, and its evolution is assumed to follow a continuous-time stochastic process. Construct a stochastic process model that includes drift and diffusion terms to describe the average trend and stochastic fluctuations of degradation; The degradation increment is calculated based on the degradation observation data, and the drift coefficient and diffusion coefficient are obtained by maximum likelihood estimation or Bayesian estimation method. The independence and distribution characteristics of the model residuals are tested to verify the rationality of the model; The stochastic process model is preferably a linear Wiener process model, or can be a diffusion process, a Gamma process, or other stochastic process models suitable for degradation modeling.

[0038] This technical solution enables the dynamic evolution of battery degradation within a probabilistic framework, providing a unified and quantifiable mathematical description for risk prediction. By introducing drift and diffusion terms into the model, the deterministic trends and stochastic fluctuations of battery degradation can be described respectively, thus balancing long-term trends and short-term uncertainties in modeling and improving the model's fitting accuracy and dynamic response capability to actual degradation behavior. The stochastic process model can flexibly select forms such as linear Wiener processes, Gamma processes, or general diffusion processes according to different battery types and degradation mechanisms, thereby being compatible with different degradation paths and temporal characteristics and improving the model's applicability in multiple application scenarios.

[0039] The first crossing time distribution is derived by an inverse Gaussian distribution when the stochastic process model is a linear Wiener process, and the first arrival time distribution is obtained by corresponding analytical or numerical methods under other specific forms of the stochastic process model.

[0040] This technical solution allows the first-crossing time distribution to be analytically derived as an inverse Gaussian distribution when the stochastic process model is a linear Wiener process. This provides a clear mathematical expression for the time distribution of thermal runaway risk, facilitating the calculation of cumulative risk probability and remaining safe lifetime indicators within an analytical framework, thus improving the physical interpretability and computational transparency of the model results. Beyond Wiener processes, appropriate analytical or numerical methods can be selected to solve the first-crossing time distribution based on the characteristics of the adopted stochastic process model (such as Gamma processes, diffusion processes, etc.), thereby establishing a general risk calculation framework. This avoids dependence on a single model form and enhances the universality of the method. By employing analytical solutions based on inverse Gaussian distributions or numerical methods for nonlinear processes, the probability distribution of degradation indicators reaching thresholds can be accurately obtained. This approach is applicable to theoretical modeling scenarios and can also be extended to real-time risk calculation of actual operational data, ensuring the accuracy and real-time nature of the prediction results.

[0041] The preset threshold is treated as a random variable and modeled according to a normal distribution or other appropriate distribution; The overall thermal runaway risk is obtained by Monte Carlo sampling of the threshold distribution and averaging or marginalizing the first-arrival time distribution for each sampling threshold. The sample size for Monte Carlo sampling can be adjusted according to the required accuracy and computing resources.

[0042] This technical solution effectively characterizes the randomness of external triggering factors such as ambient temperature fluctuations, differences in operating conditions, and manufacturing consistency deviations, making the risk model closer to actual working conditions and significantly improving the credibility and engineering applicability of the assessment results. It unifies and integrates internal degradation randomness and external triggering uncertainty within a probabilistic framework, thereby obtaining a more comprehensive and robust overall thermal runaway risk distribution.

[0043] The method further includes an online update step: in the BMS, the parameters of the fusion model and the stochastic process model are periodically retrained or incrementally identified online based on a sliding window or an incremental dataset, or under event triggering, and the estimation and early warning decision of the thermal runaway risk are updated in real time accordingly.

[0044] This technical solution enables the model to continuously absorb new operational data and correct parameter deviations, ensuring that predictions consistently reflect the battery's current true degradation state and avoiding risk assessment biases caused by model aging. Batteries are affected by multiple factors during actual operation, including temperature, load, and aging rate variations. Through an online update mechanism, the model can continuously adapt to changing operating conditions, maintaining stability and generalization ability in risk predictions across different usage scenarios, thus exhibiting higher engineering applicability. As model parameters are updated, the calculated thermal runaway risk curve, risk level, or remaining safe life index are also dynamically adjusted, allowing the BMS to trigger appropriate safety strategies such as cooling, load reduction, and power-off in real time, improving the system's proactive safety capabilities.

[0045] First, multiple current-related feature values ​​are extracted from battery degradation data. These features constitute a multivariate information vector reflecting changes in the battery's internal state. Then, a Physical Information Neural Network (PINN) is used to deeply fuse this multivariate information. After obtaining a reliable SOH estimate through PINN, the process moves to the extraction stage of the SOH-T1 correspondence. There is an inherent coupling relationship between the battery's macroscopic performance degradation (SOH) and its microscopic thermal stability (T1). This relationship is usually calibrated and verified through offline adiabatic thermal runaway calorimetry experiments. However, the inherent relationship between SOH and T1 often exhibits nonlinear characteristics, making it unsuitable for direct dynamic prediction. Therefore, the next step is to linearize the extracted correspondence, transforming the nonlinear SOH-T1 relationship curve into an approximate straight line in a linear space. Within this linear space, a linear prediction model can be used to dynamically extrapolate the current T1 value based on the SOH estimate.

[0046] Through this series of sequential data fusion and transformation steps, the model can obtain online estimates of the characterization parameter T1 from readily available degradation data. This method effectively solves the problem that T1 cannot be directly measured online, laying a solid foundation for subsequent prediction of thermal runaway risk probability based on T1 degradation trajectories.

[0047] Thermal runaway risk prediction model based on nonlinear Wiener process: The degradation process of degradation comprehensive characterization parameter T1 obtained through multi-parameter fusion By using nonlinear transformation and linear regression models, it is transformed into a monotonically increasing battery degradation process that follows a linear Wiener process: in, The drift coefficient characterizes the average rate of degradation; The diffusion coefficient characterizes the random fluctuations of the degradation process; B(t) It is standard Brownian motion. Therefore, at any time t, X(t) follows a normal distribution, that is: Battery thermal runaway is defined as the moment when the health indicator X(t) first crosses a critical failure threshold L. Let T be the time of the first thermal runaway crossing; then the risk before time t can be defined as: For a linear Wiener process, the first crossover time T follows an inverse Gaussian distribution.

[0048] If the external triggering stress for thermal runaway, representing parameter T1, is a constant L, then the cumulative distribution function (CDF) of the inverse Gaussian distribution, i.e., the thermal runaway risk function, is: in, The CDF is a standard normal distribution.

[0049] However, in reality, due to individual battery differences, the complexity of operating conditions, and the inherent uncertainty of thermal runaway, the external triggering stress that induces thermal runaway is not a fixed value, but a random variable. Assume L follows a normal distribution. Its probability density function is: At this point, the overall thermal runaway risk F(t) can be calculated using the following formula: The analytical solution to the above integral expression is complex. When the standard deviation of the external triggering stress... When the value is relatively small, an approximate result can be obtained through mathematical methods such as Taylor expansion: This approximation combines mathematical simplicity with engineering practicality. It is worth noting that when... When the value is 0, the equation degenerates into the case of fixed external triggering stress, verifying its consistency.

[0050] In this study, X(t) is the health index obtained by linearizing T1, which is derived from multivariate information fusion. The distribution parameters of external triggering stress ( The value of ) can be determined through statistical analysis of the ambient temperature under actual operating conditions of the power battery. This model successfully characterizes the randomness of the performance degradation process simultaneously. ) and the randomness of thermal runaway triggering conditions ( This allows for a more comprehensive and accurate probabilistic prediction of the risk of thermal runaway in lithium batteries.

[0051] Fusion of thermal runaway risk characterization parameters based on PINN: The modeling used in this application employs charge-discharge data from the XJTU battery dataset, with a charge-discharge cycle of 2C (charging to 4.2V at a constant current of 2.0C (4A), then maintaining the voltage until the current drops to 0.05C (0.1A); resting for 5 minutes; discharging to 2.5V at a current of 1.0C (2A); and then resting for another 5 minutes). The SOH comparison results obtained by fusing the mean, standard deviation, kurtosis, skewness, charging time, accumulated charge, curve slope, and curve entropy of the voltage and current curves are as follows: Figure 2 As shown.

[0052] By calculating the Pearson correlation coefficient between features, the correlation between each feature was analyzed. Strong correlations were found between eight current-related features and State of Health (SOH). Therefore, this paper extracts eight parameters, including the average current and standard deviation, which are highly correlated with lithium battery performance degradation, as multi-source information to fuse and obtain SOH, and then... Figure 3 The approximate relationship between T1 and SOH shown below yields the following results: Figure 4 The T1 degradation curve is shown.

[0053] Given the fundamental requirement of linearizing the degradation process in Wiener process modeling, this paper employs time scale transformation to linearize the degradation process of the thermal runaway risk characterization parameters. The time scale transformation model used is shown in equation (9): A cycle² transformation is used to map the original number of iterations to a new feature space, making... T 1 and t They exhibit a highly linear relationship, with a coefficient of determination R² of 0.977, and the degradation trajectory is as follows: Figure 4 As shown.

[0054] Parameter estimation for thermal runaway risk prediction model: The Wiener process is a continuous-time stochastic process commonly used to describe battery degradation behavior. In this model, the drift coefficient of the Wiener process ( ) and diffusion coefficient ( ) represent the average degradation rate and the degree of fluctuation in the degradation process, respectively. These parameters are estimated based on incremental degradation data of the battery.

[0055] Extract the degradation increment for each time step from the battery degradation data. Assume the battery is at time step... Each time step T 1. Degradation amount is X i Then the degradation increment Defined as: in, This represents the step size of the transformed loop.

[0056] Using the maximum likelihood estimation method, the mean and standard deviation of the degradation increment are calculated and used as the drift coefficient and diffusion coefficient of the Wiener process, respectively. Where n is the number of iterations for the degradation increment. Estimated model parameters. =-6.177×10⁻ 5 ; =0.03524.

[0057] Thermal runaway risk curve: The battery degradation process is modeled as a Wiener process. Thermal runaway is considered to occur when the degradation first reaches a certain threshold. Due to the uncertainty of the threshold, it is assumed that the random failure threshold follows a normal distribution. in The mean, Let the standard deviation be the distribution parameters of the randomly triggered stress. =68.65, =1.00, the threshold distribution curve is shown in the figure. Figure 5 As shown: The random failure threshold makes analytical calculations for predicting thermal runaway risk difficult. Since the Monte Carlo method can approximate integral calculations under random failure threshold conditions through numerical simulation, it avoids complex analytical derivations and effectively ensures accuracy in long-term predictions. This paper uses Monte Carlo simulation to randomly sample 100 threshold samples from the threshold distribution for calculation, obtaining the thermal runaway risk curve for this lithium battery as shown below. Figure 6 As shown.

[0058] This figure illustrates the evolution curve of the probability of thermal runaway risk in lithium-ion batteries as the number of cycles increases, based on the aforementioned predicted model. This curve reveals the intrinsic link between battery aging and the degradation of thermal safety performance. Analysis of the thermal runaway risk probability curve shows an S-shaped growth characteristic, indicating that the risk of thermal runaway in lithium-ion batteries increases non-linearly with the number of cycles. Specifically, when the number of cycles is below 500, the risk probability is close to zero, which can be considered a safe stage; when the number of cycles exceeds 500, the risk probability begins to gradually increase, reflecting the potential hazards caused by accelerated battery aging; when the number of cycles reaches approximately 1000, the probability of thermal runaway risk exceeds 10%, and targeted intervention measures should be introduced to prevent thermal runaway from occurring.

[0059] In summary, the curve reveals the cumulative effect of thermal runaway risk during battery cycling, highlighting the need to strengthen battery health monitoring and safety management at high cycle counts to prevent thermal runaway events.

[0060] In summary, the lithium battery thermal runaway risk prediction model proposed in this application can effectively predict the probability of thermal runaway based on multi-dimensional degradation information during actual use of lithium batteries. The resulting thermal runaway risk curve can provide a reference for battery health management and has significant practical application value.

[0061] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

[0062] While the invention has been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for predicting thermal runaway risk of power lithium battery based on multi-degradation information, characterized in that: The method comprises the following steps: performing cleaning, denoising and feature extraction on historical operation data of the battery to obtain a multivariate feature vector for degradation characterization; inputting the multivariate feature vector into a fusion model to establish a mapping relationship between the feature vector and a battery thermal stability parameter through the fusion model, and obtaining a degradation characterization parameter capable of representing the battery thermal stability; performing linearization transformation on a nonlinear relationship between the degradation characterization parameter and a health state parameter to obtain a degradation index that changes approximately linearly in a feature space; based on the linearized degradation index, establishing a stochastic process model of degradation evolution on a selected time scale, and estimating model parameters by using degradation data; defining thermal runaway occurrence as an event in which the degradation index first reaches or crosses a preset threshold according to the stochastic process model, calculating a corresponding first-reach time distribution, and thus obtaining a cumulative probability of thermal runaway risk over time; treating the preset threshold as a random variable, marginalizing or numerically solving the cumulative probability according to the statistical distribution of the random variable, and obtaining an overall thermal runaway risk considering internal degradation and external trigger stress.

2. The method of claim 1, wherein the method is based on multivariate degradation information. The establishment step of the fusion model comprises: performing feature screening on the operation data of the power lithium battery to construct a multivariate feature vector, and designing a neural network structure or other fusion model structure with physical constraints; training and optimizing the model based on multi-cycle degradation data; establishing a corresponding relationship between the health state parameter and the thermal stability parameter through experimental calibration; wherein the fusion model is a physical information neural network, and in the training process, the electrochemical-thermal coupling mechanism equation is taken as a physical constraint and is incorporated into a loss function, so as to jointly minimize the observation error and the physical residual to learn the mapping relationship.

3. The method for predicting the thermal runaway risk of power lithium batteries based on multi-dimensional degradation information as described in claim 2, characterized in that: The multivariate feature vector includes but is not limited to: average value of charging segment current, current standard deviation, charging time, cumulative charge, curve slope, curve entropy, kurtosis and skewness. The features are screened by correlation analysis and / or feature selection method and used as inputs of the fusion model.

4. The method of claim 1, wherein the method is based on multivariate degradation information. The linearization transformation step comprises: performing feature space linearization transformation on the nonlinear corresponding relationship between the degradation characterization parameter and the health state parameter, mapping the original nonlinear curve to an approximate straight line in the linear space, to obtain a linearized degradation index; performing time scale transformation on the change sequence of the linearized degradation index over time or cycle number, to obtain a degradation trajectory that changes approximately linearly on the transformed time scale; wherein the feature space linearization transformation and the time scale transformation are both reversible monotonic transformations, and the transformation form is selected from at least one of logarithm, power function, square, logarithmic time or affine linear mapping.

5. The method of claim 1, wherein the method is based on multivariate degradation information of the power lithium battery. The establishment step of the stochastic process model comprises: regarding the linearized degradation index as a random variable evolving over time or cycle number, and assuming that its evolution satisfies a continuous-time stochastic process; constructing a stochastic process model containing a drift term and a diffusion term to describe the average change trend and random fluctuation of degradation; calculating the degradation increment based on the degradation observation data, and using maximum likelihood estimation or Bayesian estimation method to obtain the drift coefficient and diffusion coefficient; The independence and distribution characteristics of the model residuals are tested to verify the rationality of the model; The random process model is preferably a linear Wiener process model, or alternatively a diffusion process, a Gamma process or other suitable random process model for degradation modeling.

6. The method of claim 1, wherein the method is based on multivariate degradation information of the power lithium battery. The first passage time distribution is derived as an inverse Gaussian distribution when the random process model is a linear Wiener process, and is obtained by using corresponding analytical or numerical methods when the random process model has other specific forms.

7. The method of claim 1, wherein the method is based on multivariate degradation information of the power lithium battery. The preset threshold is considered as a random variable and is modeled as a normal distribution or other appropriate distribution; The overall thermal runaway risk is obtained by Monte Carlo sampling of the threshold distribution and averaging or marginalizing the first passage time distribution for each sampled threshold; The sample size of the Monte Carlo sampling can be adjusted according to the required accuracy and computing resources.

8. The method of claim 1, wherein the method is based on multivariate degradation information of the power lithium battery. The method further comprises an online updating step: periodically or upon event triggering, the parameters of the fusion model and the random process model are retrained or incrementally identified based on a sliding window or an incremental data set in the BMS, and the estimation of the thermal runaway risk and the early warning decision are updated in real time accordingly.

Images