A power battery thermal runaway prediction method, electronic equipment and storage medium

Abstract

This application discloses a method, electronic device, and storage medium for predicting thermal runaway in power batteries, relating to the field of power battery safety monitoring technology. The method includes: acquiring sparse sensing signals from the battery system and preprocessing them to generate standardized input data packets; based on the input data packets, inverting unmeasurable boundary thermal excitations into parameterized equivalent thermal boundary driving quantities; using the equivalent thermal boundary driving quantities as input, using a multi-physics constraint prediction model to output the internal temperature field and thermal runaway-related internal state quantities; performing online adaptive calibration of boundary parameters or output deviations based on the residual between the predicted and measured surface temperatures; constructing risk criteria based on the calibrated internal state trajectory, and generating graded early warning signals and control linkage outputs. This application, through a closed-loop mechanism of equivalent boundary inversion and online calibration, achieves stable and early prediction of thermal runaway under sparse observation, effectively suppressing drift caused by cell differences and aging.

Description

Technical Field

[0001] This application relates to the field of power battery safety monitoring technology, and in particular to a method for predicting thermal runaway of power batteries, electronic equipment, and storage medium. Background Technology

[0002] Power batteries undertake high-power, long-cycle energy throughput tasks in vehicle and energy storage systems. During charge-discharge cycles, the battery cells continuously generate heat, and their thermal state is influenced by a combination of factors, including ambient temperature, cooling circuit status, assembly thermal resistance, and operating condition sequences. When the system encounters overcharging, internal short circuits, external thermal abuse, or reduced heat dissipation capacity, the cell temperature may gradually rise from a slow climb to a self-heating acceleration phase, triggering a chain reaction of exothermic reactions and eventually leading to thermal runaway. In engineering practice, the critical evolution process in the early stages of thermal runaway mainly occurs inside the battery cell. However, the signals that can be stably obtained over a long period in actual operation are limited to surface temperature, ambient temperature, and electrical parameters such as voltage and current at the casing or module level. The low spatial coverage and limited information of external observations mean that traditional early warning methods relying solely on threshold judgments often only respond when thermal runaway has already entered the acceleration phase, resulting in severely insufficient early warning lead time. Therefore, how to transform sparse external observations into reliable estimates of the internal thermal state and reaction process, and on this basis, form an early warning capability with sufficient lead time, is a core direction for continuous research in this field.

[0003] The closest existing technology to this invention is a technical approach that combines a first-principles-based mechanistic model with a data-driven surrogate model. Its basic approach is as follows: A temperature rise and heat release model of the battery cell under thermal abuse conditions is established based on the heat conduction equation and reaction kinetic equation. High-fidelity numerical simulation is used to obtain the evolution data of the battery cell's temperature field and reaction process. Then, different operating conditions (such as heating rate, heating start and end times, and changes in heat transfer intensity) are systematically scanned within the parameter space to generate training samples in batches. Finally, the surrogate model is trained using these samples to replace the high-fidelity simulation solver, enabling online inference. Some improved schemes further design the input as a time-history description of thermal excitation (such as heating curves and equivalent heat flux), expand the output to include internal temperature distribution and internal state parameters characterizing the reaction process, and introduce physical constraint terms during training to improve the model's cross-condition generalization ability.

[0004] However, the above-mentioned technical approach still has the following key drawbacks when deployed in engineering: First, existing methods typically assume that the boundary thermal excitation is a known input during inference. However, in real-world scenarios, factors such as external heat abuse and time-varying heat dissipation capacity make it difficult to directly measure boundary conditions. If fixed assumptions are used, the prediction results will produce significant systematic biases when the operating conditions deviate.

[0005] Second, the proxy model relies on offline simulation data for training, which cannot cover individual differences in battery cells and parameter drift caused by long-term aging. It also lacks an online correction mechanism during operation, and the reliability of the early warning continues to decline over time.

[0006] Third, the research focus of existing technologies is on reproducing the thermal runaway process under the premise of given accurate boundary input, but it has failed to incorporate the actual engineering requirement of constructing input from sparse observations into the overall solution. It lacks a complete closed-loop link from sparse observations to equivalent input inversion, and then to online calibration and risk warning, making it difficult to stably adapt to the limited sensing resources and real-time control architecture of the battery management system.

[0007] In summary, existing technologies have significant technical gaps in terms of input construction capabilities under unknown boundary thermal excitation conditions, runtime drift suppression capabilities, and end-to-end closed-loop early warning architecture. Summary of the Invention

[0008] In real-world applications, the signals available to battery systems are often sparse and unstable, typically consisting of only surface temperatures, ambient temperatures, and electrical parameters such as voltage, current, and state of charge (SOC) at a few locations. The true critical process of thermal runaway occurs inside the cell, where the internal temperature field and reaction process are difficult to measure directly. Furthermore, external thermal excitation or boundary heat transfer conditions are not naturally known in the field; they may arise from external thermal abuse, localized heat dissipation failures, changes in structural thermal resistance, or fluctuations in the cooling circuit state, manifesting as time-varying equivalent heat flux or equivalent heat transfer intensity. Because these boundary conditions cannot be directly obtained, many prediction methods must rely on assumptions in their input construction. If operating conditions change or these assumptions fail, false alarms, missed alarms, or insufficient prediction time windows can easily occur.

[0009] Building upon this, individual cell variations and aging drift during use further amplify the uncertainty of predictions. Even for the same cell model, factors such as material batches, assembly differences, thermal contact conditions, and internal resistance variations can cause the system to exhibit different thermal response characteristics at different times. If the model relies solely on offline training or fixed parameters, it can easily deviate from reality after running for a period of time, causing the prediction results to gradually become distorted. What the engineering side truly needs is a method that can operate online for extended periods: one that can stably estimate the internal thermal state and risk evolution even under conditions of limited input signals, unclear boundary conditions, and continuously changing system parameters, and convert the results into actionable early warning outputs.

[0010] Therefore, the problem this application focuses on is not simply to make the thermal runaway process more realistic, but to solve the key gap between sparse observations and risk conclusions: on the one hand, to construct reasonable equivalent inputs when the boundary thermal excitation is unknown, and on the other hand, to suppress drift through continuous calibration during operation, so that the prediction can remain consistent and deployable in different cells, different environments, and different aging stages, thereby achieving more reliable and earlier real-time prediction and early warning of thermal runaway.

[0011] The first technical solution adopted in this application is: providing a method for predicting thermal runaway of power batteries based on sparse sensing inversion and online calibration, including: Acquire sparse sensing signals from the battery system, including surface temperature, ambient temperature, and electrical parameters at multiple locations. Perform time alignment, cleaning, and confidence marking on the sparse sensing signals to generate standardized input data packets. Based on the input data packet, the battery boundary thermal excitation that cannot be directly measured is equivalent to a parameterized equivalent thermal boundary expression, and the parameters of the equivalent thermal boundary expression are inverted using the surface temperature information in the input data packet to obtain the time-varying equivalent thermal boundary driving quantity. Input the equivalent thermal boundary driving quantity, use the pre-trained multi-physics constraint prediction model to predict the internal temperature field distribution of the cell and the internal state quantity used to characterize the thermal runaway evolution stage, and simultaneously output the surface temperature estimate at the sensor location. Based on the residual between the estimated surface temperature and the measured surface temperature in the sparse sensing signal, the parameters of the equivalent thermal boundary driving quantity, the output bias term of the prediction model, and / or the key scaling factor are adaptively updated to suppress system drift. Based on the calibrated internal temperature field and internal state variables, a risk criterion is constructed, a risk level and / or remaining time estimate are generated, and an early warning signal for control linkage is output.

[0012] In an optional embodiment, the time alignment of the sparse sensing signal includes: establishing a unified master time axis, performing linear interpolation processing on each channel signal based on adjacent valid sample values, and marking the data reliability according to the interpolation span; the cleaning includes processing signal missing, signal jump, signal delay and signal noise, wherein the signal noise processing adopts a weighted fusion method based on the current observation value and the smoothed value of the previous time step to suppress high-frequency jitter.

[0013] In an optional embodiment, the equivalent thermal boundary expression includes at least one of the following: a sequence of equivalent heat flux over time, a sequence of equivalent heat transfer coefficient over time, and a sequence of equivalent boundary temperature over time. The process of equating the non-measurable battery boundary thermal excitation to a parameterized equivalent thermal boundary expression further includes parameterization: expressing the equivalent thermal boundary as a combination of a set of low-dimensional parameters and a preset time basis function, using a finite number of control coefficients to characterize the boundary change trend over the entire time interval, and adjusting the smoothness and expressive power of the boundary change by adjusting the basis function type and the number of control coefficients.

[0014] In an optional embodiment, the step of using the surface temperature information in the input data packet to invert the parameters expressed by the equivalent thermal boundary includes: constructing and solving an objective function to achieve the inversion, wherein the objective function is used to measure the weighted error between the predicted and measured surface temperature values ​​obtained based on the current boundary parameters, and applying smoothing constraints and physical rationality constraints to the rate of change or amplitude of the parameters through regularization terms to suppress high-frequency jitter and overfitting of the inversion results.

[0015] In an optional embodiment, the multi-physics constraint prediction model adopts a mapping structure that encodes boundary driving information and spatiotemporal coordinate information separately, fuses them in the latent space, and outputs the prediction results at the corresponding coordinates. The prediction model is trained by a total loss function that includes a data fitting term and a physical consistency term. The physical consistency term is used to constrain the model output to satisfy the consistency of the heat conduction equation, reaction kinetic equation, boundary conditions, and initial conditions.

[0016] In an optional embodiment, the adaptive update of the parameters of the equivalent thermal boundary driving quantity, the output bias term of the prediction model, and / or the key scaling factor based on the residual between the estimated surface temperature and the measured surface temperature in the sparse sensing signal includes: calculating a weighted comprehensive error index within a sliding window by comparing the measured values ​​and predicted values ​​of each temperature channel, and triggering a calibration action when the index continuously exceeds a threshold; the calibration action includes adaptive updates of the boundary parameters, the output bias term, or the scaling factor.

[0017] In an optional embodiment, when updating the parameters of the equivalent thermal boundary driving quantity, the parameter update magnitude is limited by constructing an objective function that includes the fitting error term within the current window and a penalty term for the parameter change relative to the previous time step; and / or, when updating the output deviation term, a step-limited recursive correction method based on the current residual direction is adopted, and the correction magnitude is dynamically adjusted according to the data reliability and operating conditions through a gating function.

[0018] In an optional embodiment, the construction of risk criteria to generate risk levels and / or remaining time estimates includes: generating a comprehensive risk score by fusing internal temperature levels, temperature change trends, internal state quantity levels and their change trends, and prediction confidence indices; the remaining time estimation includes calculating the estimated time to reach the danger threshold based on the difference between the current maximum internal temperature and the danger threshold and the current rate of temperature rise, and outputting a conservative estimate when the rate of temperature rise is too low or the data confidence is low; the risk levels include at least a warning level, an alert level, and a high-risk level, and a hysteresis comparison and minimum duration maintenance mechanism is used to suppress level jumps.

[0019] The second technical solution adopted in this application is: providing an electronic device, the electronic device including: a memory and a processor coupled to each other, the processor being used to execute program instructions stored in the memory to implement the power battery thermal runaway prediction method of any of the above.

[0020] The third technical solution adopted in this application is: providing a computer-readable storage medium that stores program data, which can be executed by a processor to implement the power battery thermal runaway prediction method described above.

[0021] Due to the adoption of the above technical solution, this application has at least one of the following beneficial effects compared with the prior art:

[0022] 1. Even with unknown boundary thermal excitation and only sparse sensing input, the method can still construct equivalent inputs and achieve thermal runaway prediction, significantly improving its deployability in actual vehicles and energy storage systems.

[0023] 2. The introduction of an online calibration mechanism can effectively suppress prediction drift caused by individual cell differences, aging, and changes in heat dissipation conditions, significantly improving long-term operational stability and cross-cell consistency.

[0024] 3. By adopting a fast inference and lightweight online update strategy based on a proxy model, the prediction accuracy is guaranteed while maintaining computational efficiency, which can meet the response time requirements of real-time early warning and control linkage of the battery management system. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 A schematic flowchart of a power battery thermal runaway prediction method provided in an embodiment of this application; Figure 2 This is a schematic diagram of an equivalent thermal boundary inversion module provided in an embodiment of this application; Figure 3 This is a schematic diagram of a multi-physics constraint prediction module provided in an embodiment of this application; Figure 4 This is a schematic diagram of an online calibration closed loop provided in one embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only for explaining this application and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all structures. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0027] The terms "first," "second," etc., used in this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] Existing methods for predicting thermal runaway in power batteries typically rely on known boundary thermal excitation inputs. However, in practical engineering, external thermal excitations or boundary heat transfer conditions are not directly measurable. Furthermore, models trained on offline data lack operational correction capabilities and struggle to address cell variations and aging drift. Therefore, this application provides a method for predicting thermal runaway in power batteries based on sparse sensing inversion and online calibration. The core idea is to: equate the unmeasurable boundary thermal excitation to a parameterizable expression; invert this equivalent expression using sparse observations; transform the equivalent boundary drive into an internal state estimate using a multi-physics constraint prediction model; and finally suppress operational drift through an online calibration mechanism, forming a complete closed loop from sparse observation to risk warning.

[0030] like Figure 1 As shown, Figure 1 The flowchart of a power battery thermal runaway prediction method provided in an embodiment of this application is shown, including the following steps executed sequentially: acquiring sparse sensor signals and performing preprocessing, equivalent thermal boundary inversion, multi-physics constraint prediction, online calibration, and risk judgment and early warning. Specifically, the output of the sparse sensor signal acquisition and preprocessing step is sent to the equivalent thermal boundary inversion step and the online calibration step, respectively; the output of the equivalent thermal boundary inversion step is sent to the multi-physics constraint prediction step; the output of the multi-physics constraint prediction step is sent to the online calibration step and the risk judgment and early warning step, respectively; and the output of the online calibration step is fed back to the equivalent thermal boundary inversion step and the multi-physics constraint prediction step, forming a closed loop.

[0031] Sparse sensing signals refer to the set of temperature sensor signals that can be stably acquired over a long period of time in the actual engineering deployment of battery systems, whose spatial coverage is much lower than the total number of observed signals of individual battery cells. Typically, when the ratio of the number of temperature sensors in the battery pack to the total number of individual battery cells is no higher than 1 / 10, it is considered to meet the sparse condition; in more constrained engineering scenarios, this ratio can be as low as 1 / 40 or even lower. Sparse sensing signals usually only originate from a limited number of temperature measurement points on the surface of the casing or module, lacking direct measurement of the internal temperature of the battery cells. At the same time, the signal channels may be incomplete in the time dimension, such as being asynchronous, missing, jumping, or having noise.

[0032] The workflow of the method in this embodiment is described in detail below: The system receives raw sparse sensing signals from the battery management system and thermal management controller as input. These raw sparse sensing signals include surface temperatures, ambient temperatures, and electrical parameters such as voltage, current, and SOC (State of Charge) at multiple locations. Temperature signals include one or more surface temperature points selected from one or more of the following: near the cell casing, near the tabs, on the module surface, or near the cooling plate. Ambient signals include one or more of the following: internal temperature, coolant inlet temperature, coolant outlet temperature, or duct temperature. Electrical parameters include at least voltage and current, and preferably, one or more of the following: SOC, estimated internal resistance, or power command. To reduce bias caused by point differences, each temperature channel is bound to a corresponding physical location category and thermal coupling strength level. During input generation, either single-point sequences or aggregated sequences based on location categories can be used to avoid misinterpreting fluctuations near the cooling plate as an overall temperature rise trend. Simultaneously, metadata such as sensor range, resolution, update cycle, and whether hardware filtering has been applied is retained on the acquisition side to determine the reliability boundaries of subsequent channel data.

[0033] Because the timestamps of different signals are not completely synchronized, the original signals are preprocessed. First, a main time axis is selected, with a time step set between 0.5 and 5 seconds. For each time step, the alignment value of each channel at that moment is calculated. Linear interpolation is used for alignment, and the interpolation span is recorded. When the interpolation span exceeds a preset upper limit, the alignment value is marked as low confidence so that its weight can be reduced in subsequent stages or protection logic can be triggered.

[0034] The linear interpolation method is as follows: Let the effective sampling times of two consecutive signals in a certain signal channel be... and , The corresponding sample value is and Then at the intermediate moment Alignment estimate at [location] Calculate using the following formula (1)

[0035] After alignment, signal loss, signal jump, signal delay, and signal noise are processed separately. For signal loss, if it is a short-term missing packet, interpolation or keeping the most recent valid value is used to fill it in, and the duration of continuous loss is recorded; if it is a long-term disconnection, it is marked as "unavailable", and subsequent modules switch to a conservative strategy or only use the still available channels within this window. For signal jump, the signal change rate is calculated, and when the change rate exceeds the preset reasonable upper limit, the point is marked as an anomaly. For signal delay, it is identified by comparing the deviation between the timestamp and the main time axis, and the data age information is retained in the input to avoid splicing expired electrical parameters with the latest temperature to avoid false correlation. For signal noise, the exponential moving average method is used for smoothing, as shown in the following formula (2):

[0036] in For the first The original observations at each sampling time point, This is the smoothed value from the previous time step. The result after smoothing. This is a smoothing coefficient. A smaller coefficient is used under normal operating conditions where the temperature rise is relatively slow. A larger value is used during the rapid heating phase. value.

[0037] Preprocessing also includes the calculation of derived quantities. These derived quantities include one or more of the following: smoothed temperature value, short-window heating rate, and heating acceleration. These are used to distinguish between hysteresis caused by enhanced external thermal excitation and weakened cooling during boundary inversion. The electrical parameter sequence is aligned into interpretable segments according to operating conditions, including one or more of the following: charge / discharge switching, power step, and prolonged static state. These states are entered as conditional variables into subsequent modules to avoid misclassifying normal heating caused by current changes as anomalies. For environmental and cooling signals, flag quantities are formed. These flag quantities include one or more of the following: cooling level change events, fan start events, and coolant temperature abrupt change events. These serve as constraints or priors during subsequent inversion of the equivalent thermal boundary to reduce unreasonable high-frequency fluctuations in the inversion results.

[0038] Input data is organized using a sliding window approach. The window length is set between 30 seconds and 10 minutes, and the window step size is the same as or an integer multiple of the main time step. Each window outputs a structured data packet, which includes: aligned multi-channel time series, confidence weight or availability flag for each channel, key event markers within the window, and a statistical summary within the window. The statistical summary includes one or more of the following: maximum temperature within the window, quantile of the temperature rise rate, missing data ratio, and data age distribution.

[0039] After preprocessing, the output is a standardized input data packet. After the standardized input data packet is formed, the equivalent thermal boundary is inverted. The battery boundary thermal excitation, which cannot be directly measured, is equivalent to a parameterized equivalent thermal boundary expression. The parameters of the equivalent thermal boundary expression are inverted using the surface temperature information in the input data packet to obtain the time-varying equivalent thermal boundary driving quantity.

[0040] The equivalent thermal boundary expression includes at least one of the following: the equivalent heat flux over time, the equivalent heat transfer coefficient over time, and the equivalent boundary temperature over time.

[0041] The equivalent thermal boundary is parameterized. It is expressed as a combination of a set of low-dimensional parameters and a predefined time basis function. Let the equivalent thermal boundary be... Expanding it into the following equation (3):

[0042] in The pre-selected time basis functions can be piecewise linear bases, spline bases, or a set of locally supported basis functions. The coefficients to be inverted. Formula (3) means using a finite number of coefficients to control the boundary change trend over the entire time period. This determines the smoothness of boundary changes and the expressible complexity. The degree of freedom is determined; when a smoother boundary change is desired, a more smooth basis function can be selected and the number of basis functions can be reduced. When capturing events such as cooling level switching or sudden external heat source entry, denser nodes can be placed near the event or local supporting basis functions can be used to enhance the expressive power of the boundary at critical moments. For implementations requiring more intuitive control, piecewise constants or piecewise linear sequences can be used directly as parameters. For example, several control points can be selected within a window and linear interpolation can be performed between the control points. This approach facilitates the application of rate of change constraints and amplitude constraints, and is convenient for recursive updates in online scenarios.

[0043] After parameterization, the goal of the inversion is to find a set of parameters. This ensures that the observables obtained based on the equivalent boundary drive are as consistent as possible with the measured data. Typically, a forward mapping is needed to represent how the predicted surface temperature is generated under given equivalent boundaries and known operating conditions. This forward mapping can be a fast, low-order thermal response model or a simplified version of the subsequent prediction module; the key is to quickly output the estimated surface temperature during the iterative solution process. And it allows multiple temperature points to participate in the fitting together. Based on this, the inversion objective function can be constructed as shown in equation (4) below:

[0044] in, For the first Each temperature channel at time The measured value, The forward mapping represents the predicted values ​​at the same time. The weights are used to reflect the data's reliability, the importance of sensor location, or the impact of data age. For error measurement function, For regularization terms, This is the regularization weight. The purpose of this formula is to prioritize fitting the inversion results to reliable data, while using the regularization term to suppress unreasonable high-frequency fluctuations or overfitting; when jump points or occasional anomalies exist... A more robust approach to anomalies can be chosen to prevent a small number of outliers from skewing the boundary inversion. Weights The credibility markers from the previous stage can be directly inherited. For example, the weight of points with large missing interpolation spans can be reduced, the weight of points with older data can be reduced, and the weight of points near the cooling plate and susceptible to local disturbances can be reduced in certain scenarios, thus making the inversion more in line with engineering intuition.

[0045] Regular terms The design determines the interpretability and stability of the equivalent boundary. A common and effective approach is to add smoothing constraints, limiting the rate of change or second-order transformation of the boundary parameters, so that the boundary does not exhibit physically inconsistent fluctuations over time. For example, when When dealing with piecewise linear control points, a smoothing term can be constructed using adjacent differences. The meaning of this is to penalize excessively rapid changes; if a greater emphasis is placed on the overall smoothness of the curve, the second-order difference can be further penalized, making the boundary changes more closely resemble a continuously accelerating pattern. The smoothing constraint is not merely to make the curve look better, but to pull the inversion problem from noise-driven to mechanism-driven: surface temperature changes have thermal inertia; the external boundary cannot oscillate at high frequencies for an extremely short period while the surface temperature remains stable. Therefore, if the inverted boundary exhibits high-frequency jitter, it often indicates overfitting or data quality issues. In addition, amplitude constraints and rate-of-change limits can be added, such as limiting... Alternatively, the value of the equivalent heat transfer coefficient should fall within a reasonable range, and its change per unit time should be limited to the upper limit. This allows for boundary changes during cooling switching or heat source intrusion, but does not allow unreasonable spikes.

[0046] Besides temperature error fitting, equivalent boundary inversion can also leverage electrical parameter information to enhance identifiability, especially in scenarios with significant current changes and non-negligible endogenous heating. In such cases, information such as current, voltage, and state of charge (SOC) can be used to construct prior terms for internal heating or to distinguish the contributions of boundary heat input and endogenous heating. A more robust engineering approach is usually not to force electrical parameters into the boundary representation, but rather to treat them as conditional variables in the forward mapping. This allows the forward mapping to consider the influence of endogenous heating when estimating surface temperature, thus making the inverted boundary closer to the combined effect of external input and heat dissipation capacity. The advantage of this approach is that the output meaning of the boundary inversion is clearer: it primarily reflects changes in external heat input and heat dissipation, avoiding misinterpreting normal heating caused by current steps as external heat abuse. Furthermore, when the system is at rest or with low current, the boundary inversion is more sensitive to external heat abuse, enabling earlier identification of non-electrically driven abnormal temperature rises.

[0047] In terms of the solution strategy, a recursive update method is adopted in the online scenario: each time a new data segment is acquired, an update is performed based on the inversion result of the previous window. The initial value of the recursive update is taken as the boundary value of the previous window or its extrapolation. The stopping conditions include any one of the following: the objective function value drops to a preset threshold, the maximum number of iterations is reached, or the update magnitude exceeds a preset upper limit. The inversion frequency is decoupled from the window step: temperature sampling is updated on a second-by-second basis, while boundary inversion is updated every few seconds to tens of seconds; the inversion frequency is reduced when the risk is low and the operating conditions are stable, and increased when the temperature rise rate increases or the cooling state switches.

[0048] To reduce the risk of multiple solutions, one or more of the following measures are adopted: reduce the degrees of freedom during the parameterization stage; enhance smoothness constraints and amplitude constraints; introduce event priors, and treat cooling speed switching, fan start / stop, or coolant temperature change events as allowed nodes for boundary changes; for cases with multiple temperature points, strengthen the uniqueness of the solution through point weights and consistency constraints.

[0049] Output the inversion confidence information. The fitting residual statistic is calculated according to the following formula (5):

[0050] in The number of effective samples after weighting; when As the temperature continues to rise, an online calibration mechanism is triggered. The inversion output includes the boundary sequence within this window. Or one or more of the following: its equivalent parameter sequence, corresponding timestamp alignment information, residual statistics, missing proportion, average data age, and a marker indicating whether the inversion iteration is stable and convergent.

[0051] For scenarios requiring prediction of short-term risks, a short-term extrapolation strategy for the boundaries is proposed: maintain constant boundary values ​​within the future prediction window, or use the terminal slope for linear extrapolation, and limit the extrapolation length to a preset range. Extrapolation is used to provide continuous driving force for subsequent predictions within the short-term window; when new observations arrive, updates are triggered by the new observations first, avoiding long-term reliance on extrapolation that could lead to error accumulation.

[0052] When a large number of temperature channels are offline, the data in key channels is too old, there are unexplained jumps within the window that occur simultaneously in multiple unrelated channels, or the inversion iteration fails to converge, a conservative boundary is output, that is, the boundary of the previous window remains unchanged or is updated by a very small amount, and the inversion confidence is marked as low.

[0053] The above equivalent thermal boundary inversion process is as follows: Figure 2 As shown. Figure 2 This is a schematic diagram of an equivalent thermal boundary inversion module provided in an embodiment of this application. Figure 2 As shown, the standardized input data packet is sent to the equivalent thermal boundary inversion module. Inside the module, the parameterization unit expresses the equivalent thermal boundary as a combination of a set of low-dimensional parameters and a preset time basis function; the forward mapping unit outputs the surface temperature estimate based on the current boundary parameters and condition variables; the objective function solving unit solves for the optimal parameters that minimize the objective function shown in formula (4) through iterative optimization; and the confidence assessment unit calculates the fitting residual statistics based on formula (5) and outputs the inversion confidence label. The module outputs the equivalent thermal boundary driving quantity and its confidence information.

[0054] After the equivalent thermal boundary is inverted, multi-physics constraint prediction is performed. Using the equivalent thermal boundary driving quantity as input, the pre-trained multi-physics constraint prediction model is used to predict the internal temperature field distribution of the cell and the internal state quantity used to characterize the thermal runaway evolution stage, and simultaneously outputs the surface temperature estimate of the sensor location.

[0055] The input to the prediction model includes the equivalent thermal boundary representation and its parameterized coefficients, as well as one or more of the following as boundary-related condition variables: ambient temperature, cooling medium temperature, and cooling state; optionally, it also includes one or more of the following as condition variables: current, voltage, state of charge (SOC), and internal resistance estimates. The output of the prediction model includes the spatial and temporal distribution of the internal temperature, and one or more internal state variables characterizing the thermal runaway evolution stage. These internal state variables include one or more of the following: equivalent exothermic intensity, equivalent reaction progress, or equivalent decomposition rate.

[0056] The training data for the prediction model comes from virtual samples under multiple operating conditions. The virtual samples are generated by numerically solving a coupled model of heat conduction and reaction kinetics under various boundary driving combinations to obtain the internal temperature field and internal state variables, and extracting surface temperature as an observable. These various boundary driving combinations include different heating rates, different boundary abrupt events, different heat dissipation intensity changes, and different initial temperature ranges. The sampling density is increased during the critical phase of sample collection. Each sample corresponds to the boundary driving and conditional variables within a time window, and the output is the field variable value at any time and spatial location within that window. If the residuals in certain intervals remain consistently high during training, the sampling density or loss weight for these points is increased to concentrate training resources on the challenging regions.

[0057] The prediction model employs a mapping structure that encodes boundary driving information and spatiotemporal coordinate information separately, fuses them in the latent space, and outputs the prediction results at the corresponding coordinates. Let the equivalent boundary driving sequence or its parameterization at the cutoff time t within the window be as shown in equation (6):

[0058] in Indicates the cutoff time within the window The equivalent boundary-driven sequence or its parameterized representation. It is a function that encodes boundary-driven operations. Spatial location With time The function for encoding, This represents conditional variables such as ambient temperature, electrical parameters, and cooling status. This represents the coordinates The predicted output at that point can include a combination of temperature and internal state variables. This formula means treating the boundary-driven mechanism as the input that determines the global response, and... Using this as a query point, the same set of parameterized mappings outputs results at different locations and times, thereby enabling continuous prediction of internal field variables. Compared to outputting discrete points at only a few moments, this representation is more suitable for subsequent trend judgment and remaining time estimation, and it is also easier to interface with different cell structures or grid divisions.

[0059] Relying solely on data fitting carries a risk: in boundary combinations or critical stages not covered by the training samples, the model may produce seemingly smooth but physically inconsistent predictions. Therefore, it is necessary to introduce multi-physics constraints into the training process, ensuring that the output not only fits the data but also satisfies the relationship between the thermal equation and reaction kinetics, and remains consistent with the initial and boundary conditions. The training objective can be written as a loss function composed of data terms and physical terms, as shown in equation (7):

[0060] in This is used to constrain the predicted output to be consistent with the simulation or experimental labeled data at the sample points. Used to constrain the residuals of the heat conduction equation and the reaction kinetic equation, and These are used to constrain the consistency between boundary conditions and initial conditions, respectively. These are the weighting coefficients. The weighting coefficients are dynamically adjusted based on the residual magnitude or convergence speed during training. The calculation points of the physical residuals spatially cover the cell region, temporally cover the vicinity of boundary events and the self-heating acceleration interval, and increase the sampling density near the boundary or in rapidly changing intervals.

[0061] The prediction module outputs reliability auxiliary quantities. These include the predicted surface temperature at the sensor location and the corresponding residual statistics, as well as a confidence index that varies with the window, formed by combining the residuals with a data confidence label. When the confidence is low, the warning trigger threshold is increased, and the deterministic expression of the remaining time estimate is reduced. When the magnitude or rate of change of the equivalent boundary exceeds the historical range of the training samples, the prediction module will mark the output as an extrapolated state and limit the growth rate of the internal state quantities or trigger an encrypted update for online calibration.

[0062] During online inference, the prediction module takes the equivalent thermal boundary driver and condition variables at the window level as input, outputs the internal temperature field and internal state variables at predetermined spatiotemporal query points, and simultaneously outputs the surface temperature estimate at the sensor location. The spatiotemporal query points are selected from one or more of the following: near the casing, in the central region of the cell, or near the tab. The inference frequency is consistent with the data window step, or the update frequency is increased when the risk increases or the boundary changes rapidly.

[0063] The above multi-physics constraint prediction process is as follows: Figure 3 As shown. Figure 3 This is a schematic diagram of a multi-physics constraint prediction module provided in one embodiment of this application. Figure 3 As shown, the equivalent thermal boundary driving quantity and its reliability information, along with condition variables, are fed into the multi-physics constraint prediction module. Inside the module, the boundary drive encoder encodes the equivalent boundary driving sequence within the window, outputting a boundary feature vector; the spatiotemporal coordinate encoder encodes the spatial location and time of the query point, outputting a spatiotemporal feature vector; the latent space fusion unit fuses the two feature vectors in the latent space; and the decoding output unit outputs the predicted temperature or state variable at that spatiotemporal coordinate. The module outputs the internal temperature field distribution, internal state variables, and estimated surface temperatures at the sensor locations.

[0064] After the prediction results are generated, online calibration is performed. Based on the residual between the estimated surface temperature and the measured surface temperature, the parameters of the equivalent thermal boundary driving quantity, the output bias term of the prediction model, and / or the key scaling factor are adaptively updated.

[0065] For the There are several temperature channels, let them be at time... The measured value is The prediction module provides an estimate for the same location as follows: The residual is then defined as follows (8):

[0066] if A slight fluctuation around zero within a window indicates that the boundary inversion and prediction modules are consistent with reality within that window; if the residuals show a continuous shift in the same direction, it indicates a systematic bias that needs calibration; if the residuals suddenly change and only occur in individual channels, it is often more like a sensor malfunction or local noise, and should not drive a large update of the overall model. To make the triggering logic more stable, it is usually not only necessary to look at a single point residual, but to summarize the residuals within a sliding window to form a comparable drift index. For example, the weighted mean square residual can be defined as follows (9):

[0067] in For The sliding window at the end, As weights, they inherit the credibility markers from the data preprocessing stage. The number of effective samples after weighting; when The calibration action is triggered when the residual exceeds a preset threshold for multiple consecutive windows; the calibration action is also triggered when the sign and magnitude of the residual show a trend change and the conditional variables such as cooling status and ambient temperature do not change in a way that can explain the trend. When the cooling level is switched, the fan is started, or the coolant temperature changes abruptly, the residual increases within a predetermined time but does not immediately trigger calibration; the calibration is performed after the thermal inertial response stabilizes.

[0068] The calibration process includes adaptive updates to boundary parameters, output bias terms, or scaling factors. The choice among these three is as follows: when the bias originates from inaccurate boundary drivers such as changes in heat dissipation efficiency, the boundary parameters are updated; when the bias exhibits slow, continuous drift, the output bias term is updated; and when the bias shows a proportional deviation in the same direction, reflecting changes in the overall thermal response, the scaling factor is updated. A dual-timescale strategy is employed: a fast-scale layer performs conservative compensation, while the slow-scale layer adjusts the boundary parameters or scaling factor when sufficient evidence is available, preventing a single update from leading the system in the wrong direction.

[0069] When updating the parameters of the equivalent thermal boundary driving quantity, let the boundary parameters at the previous moment be... The current update is The calibration target with inertial constraints is constructed as shown in equation (10):

[0070] in This indicates that the boundary parameters are The surface temperature estimate provided by the time prediction module For robust error measurement, It is a regularized strength; at the same time, amplitude constraints and rate of change constraints are applied to the parameters.

[0071] When updating the output bias term, the calibrated surface temperature estimate is written as the model output. Deviation Compensation Item The sum is shown in equation (11) below:

[0072] The deviation compensation term adopts a step-limited recursive correction method based on the current residual direction, as shown in equation (12):

[0073] in To update the step size, A function for gating and limiting residuals; gating function The correction magnitude is dynamically adjusted based on data reliability and operating events: the response magnitude is reduced when the data is too old, the missing percentage is high, or the cooling state undergoes abrupt changes; the response magnitude is increased when the residuals are consistently in the same direction and stable in magnitude across multiple windows. The deviation term establishes both the surface temperature channel and the internal state variables.

[0074] When updating the scaling factor, a global scaling factor is introduced. Correct the thermal response. When the residuals of multiple temperature channels exhibit a proportional deviation in the same direction and this deviation does not match the reasonable range of the boundary inversion, multiply the internal heating term or equivalent thermal resistance term by... And driven by residual trends Changes are made, and an update speed limit is set.

[0075] Online calibration also includes a miscalibration prevention mechanism and a rollback mechanism. The miscalibration prevention mechanism works as follows: when the residual of a certain temperature channel deviates abnormally from other channels, but the electrical parameters do not change correspondingly with the cooling state, and the data quality marker for that channel shows anomalies, the weight of that channel in the weighted comprehensive error index is reduced, or it is removed from the residual summary; when multiple channels with large spatial differences simultaneously exhibit consistent deviations that persist within the window, calibration is allowed. The rollback mechanism works as follows: when the residual deteriorates significantly within a predetermined time after calibration, or calibration causes the equivalent boundary parameters to exceed limits, or calibration causes unreasonable changes in internal state quantities, it rolls back to the previous set of stable parameters, and the next update step size is reduced or the regularization strength is increased, narrowing the update amplitude in subsequent windows.

[0076] In terms of project timing, deviation updates are performed on a second-by-second or window-step basis; boundary parameter updates and scaling factor updates are performed at a lower frequency or when trigger conditions are met. During periods of rapid risk increase, the calibration frequency is increased and the window length is shortened; during periods of low risk and stable operating conditions, the window is lengthened and the update frequency is reduced.

[0077] The above online calibration process is as follows: Figure 4 As shown. Figure 4 This is a schematic diagram of an online calibration closed loop provided in one embodiment of this application. Figure 4 As shown, the measured surface temperature and its confidence mark, along with the estimated predicted surface temperature, are fed into the online calibration closed loop. Inside the closed loop, the residual calculation unit calculates the residual of each channel according to formula (8) and calculates the weighted comprehensive error index within the sliding window according to formula (9); the trigger judgment unit decides whether to trigger calibration based on the continuous deviation of the comprehensive error index; the parameter update unit selects to update the boundary parameters, output the deviation term, or the scaling factor according to the trigger signal, wherein the boundary parameter update is solved according to formula (10), and the deviation term update is recursively corrected according to formulas (11) and (12); the backoff control unit performs parameter backoff when the residual deteriorates after calibration. The closed loop outputs the updated boundary parameters, deviation term, or scaling factor, which are sent back to the equivalent thermal boundary inversion module and the multi-physics constraint prediction module through the feedback loop.

[0078] After calibration, risk criteria and early warnings are performed. Based on the calibrated internal temperature field and internal state variables, risk criteria are constructed, risk levels and / or remaining time estimates are generated, and early warning signals for control linkage are output.

[0079] The risk criteria use internal temperature levels, temperature rise trends, and growth trends of internal state quantities as indicators. Internal temperature aspects include the absolute level of the maximum internal temperature and its rate of change; internal state quantities include the values ​​and rates of change of state indicators characterizing equivalent exothermic intensity or equivalent reaction progress. Criterion calculations are performed within a sliding window, the window length of which is consistent with or an integer multiple of the sliding window used in the online calibration process.

[0080] A comprehensive risk score is generated by integrating multiple indicators. As shown in equation (13):

[0081] in For the internal temperature field at time The maximum value, The rate of change estimated for it within the window. As an internal state indicator, Its rate of change, For safe reference temperature, For temperature normalization scale, and These are the normalized scales for the rate of temperature rise and the rate of change of state, respectively. As a consistency / confidence index, These are the weighting coefficients. For a monotonically compressible function, make It falls within the range of 0 to 1. Each normalization scale and weighting coefficient is a configurable parameter, calibrated according to cell type, heat dissipation capacity, and application scenario.

[0082] The early warning output employs a tiered mechanism, with tiering based on risk score magnitude, duration of risk score, and whether trend indicators meet acceleration characteristics. The tiering mechanism includes hysteresis comparison and a minimum duration maintenance mechanism. The tiers include at least a warning level, an alert level, and a high-risk level. The warning level is defined as follows: the risk score exceeds a low threshold and does not decline within several consecutive windows, while the rate of temperature rise or internal state change rate shows an upward trend. The alert level is defined as: the risk score exceeds a medium threshold and the trend item meets the persistence condition; the high-risk level is defined as: the risk score exceeds a high threshold, or is triggered by the remaining time estimate. Hysteresis comparison requires a lower decline threshold to be met and maintained for a predetermined time before the risk level falls from a high level to a low level. The minimum duration maintenance mechanism ensures that once a higher level is reached, the system maintains the output of that level for at least a predetermined time.

[0083] The remaining time estimate includes the estimated time for the maximum internal temperature to reach the danger threshold and / or the estimated time for internal condition indicators to reach critical levels. For the maximum internal temperature, the estimated time to reach the danger threshold is... As shown in equation (14):

[0084] in The dangerous threshold temperature, This is a lower limit constant used to prevent the denominator from being too small and causing the estimate to diverge when the temperature rise rate is very small or affected by noise. The remaining time estimate is output in conjunction with the confidence index: when the confidence index decreases or the data quality marker indicates an increased risk of extrapolation, a conservative remaining time estimate is output or only the risk level is output without outputting a specific value.

[0085] The control linkage output is provided to the battery management system or the thermal management controller in the form of interface fields. The output fields include a risk level, a risk score, remaining time estimation and a confidence label, accompanied by a trigger basis label. The trigger basis label is jointly determined by the trend term and the consistency term, and is used to indicate the dominant evidence type of the current risk: when the temperature rise rate and the internal state change rate rise together with good consistency, it is marked as self-heating acceleration dominated; when the equivalent boundary thermal excitation increases significantly within a predetermined time and multi-point temperatures respond synchronously, it is marked as boundary heat input enhancement dominated; when the cooling state change does not match the temperature response and the residual error continuously deviates, it is marked as heat dissipation capacity degradation or system drift dominated.

[0086] The linkage action is enhanced step by step according to the risk level. At the prompt level, a lightweight linkage signal is output, for driving the upper-layer control logic to increase the monitoring frequency, limit the upper limit of charging power or raise the priority of the cooling strategy. At the early warning level, a protection suggestion signal is output, for driving the upper-layer control logic to perform power reduction, rate limiting, and forcing the thermal management to enter the enhanced heat dissipation mode. At the high-risk level, a forced protection signal is output, for driving the upper-layer control logic to perform rapid power reduction, isolate the faulty branch, trigger a safety shutdown or instruct manual disposal. The graded, continuous and confidence-bearing decision basis output by this method upgrades the upper-layer control logic from single-point threshold alarm to internal trend-driven graded protection.

[0087] A set of end-to-end embodiments is given below.

[0088] In one embodiment, the system is deployed on a module comprising a plurality of cells, and collects at least three surface temperature channels, an ambient temperature or cooling medium temperature channel, and electrical parameters such as voltage and current in real time. The main time axis is updated at a fixed step size, and data is organized within a sliding window. In the start-up phase, the system first completes alignment and smoothing with a short window, establishes confidence labels for each temperature channel, and records event changes of the cooling strategy. Subsequently, the system inverts the equivalent thermal boundary (such as Figure 2 shown in the figure) within each window; the prediction module (such as Figure 3 shown in the figure) outputs the key position values of the internal temperature field and internal state indicators within the same window, and outputs surface temperature estimated values at the sensor positions; the online calibration (such as Figure 4 shown in the figure) takes the residual error between the measured surface temperature value and the predicted value as the entry, and updates the boundary parameters or outputs a deviation term when the residual error presents continuous deviation; the risk criterion uses the calibrated internal maximum temperature, temperature rise trend and internal state growth trend as indicators to form a risk score and a risk level.

[0089] In this embodiment, the system initially operates under stable conditions, with the equivalent thermal boundary inversion outputting a slowly changing boundary sequence. The predicted internal temperature matches the surface observation, and the residual fluctuates slightly around zero. When an external heat input enhancement event occurs, the surface temperature rises synchronously at multiple points, and the upward trend does not match the current change. During the boundary inversion stage, the system interprets this change as an equivalent heat input enhancement, resulting in a boundary-driven sequence with an increased amplitude and a limited rate of change. Under this boundary drive, the internal hotspot temperature output by the prediction module rises faster than the surface temperature, and the growth rate of the internal state indicators increases synchronously. The risk score begins to increase before the temperature approaches the danger threshold, moving from the alert level to the warning level. Online calibration observes a continuous unidirectional deviation between the predicted and measured surface temperatures, with the deviation consistent at multiple points and normal data quality markings. This is determined to be a systematic drift rather than a sensing anomaly. Therefore, the boundary parameters are updated with a limited amplitude, and a slowly changing deviation term is superimposed to align the predicted and measured surface temperatures. The calibrated internal temperature trajectory stabilizes, the remaining time estimate converges, and the upper-level control logic executes enhanced cooling and power limiting accordingly. As the control action takes effect, the surface temperature rise rate decreases, the internal state growth rate decreases, and the risk score gradually decreases after satisfying the hysteresis and minimum hold time requirements.

[0090] In another embodiment, a surface temperature channel experiences continuous missing data and excessively large interpolation spans during operation. During the preprocessing stage, the system marks this channel as low-confidence and reduces its contribution to residual summarization and boundary inversion weights, ensuring that boundary inversion primarily relies on still reliable temperature channels and operating condition information. At this point, the prediction module still outputs the internal state trajectory, but the confidence index decreases as effective observations diminish, and the risk output becomes more conservative. Online calibration reduces the update step size and increases regularization constraints to avoid updates driven by unreliable observations, preventing false alarms or missed alarms due to single-channel outages. Once the temperature channel recovers and the data age and missing data ratio return to normal, the system gradually restores its weight.

[0091] In another embodiment, a short-lived spike occurs in a channel. This spike is short-lived and does not occur synchronously in other channels, while the electrical parameters and cooling status show no corresponding changes. The system adopts an anomaly-insensitive approach in its error measurement, suppressing the impact of this spike on boundary inversion and calibration targets. In terms of triggering logic, calibration is only triggered when the weighted comprehensive error index deviates continuously within a window; a single spike does not trigger calibration. If the spike occurs near a cooling switch, the system uses a cooling event marker to activate a gating mechanism, reducing the calibration update intensity for a predetermined time. The system then determines whether a systematic deviation exists after the thermal inertial response stabilizes. If the residual deteriorates after calibration, a rollback mechanism restores the parameters to the previous stable state and reduces the update amplitude in subsequent windows, ensuring that the system does not enter uncontrollable oscillations due to a single abnormal data point.

[0092] In another embodiment, the cooling loop efficiency decreases as the ambient temperature rises, and the residual index continues to deviate in the same direction over a longer period. The system uses drift parameters or deviation terms with a slow time scale to absorb the trend, while simultaneously updating the boundary parameters under constraints. The change in risk score exhibits a smooth trend. When a real external heat input enhancement event is superimposed on this long-term drift, the system distinguishes the main cause of the risk increase by the accelerating characteristics of the internal state growth rate and outputs the corresponding risk level and linkage signal.

[0093] As can be seen from the above embodiments, the system forms a stable boundary drive and internal state trajectory under normal operating conditions. During the stage of enhanced external heat input or self-heating acceleration, the risk level is raised in advance and the control linkage is activated. Under conditions such as sensor disconnection, noise jumps, and cooling switching, stable output is maintained by relying on reliability marking, robust error measurement, gating update, and rollback mechanisms. Regarding the above embodiments, this application provides an electronic device; please refer to [link / reference]. Figure 5 , Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The computer device includes a memory and a processor, wherein the memory and the processor are coupled to each other. The memory stores program data, and the processor executes the program data to implement the steps of any embodiment of the power battery thermal runaway prediction method described above.

[0094] In this embodiment, the processor may also be referred to as a CPU (Central Processing Unit). The processor may be an integrated circuit chip with signal processing capabilities. The processor may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0095] The methods described in the above embodiments can be implemented as computer programs; therefore, this application proposes a computer-readable storage medium. Please refer to [link to relevant documentation]. Figure 6 , Figure 6 This is a schematic diagram of a computer-readable storage medium provided in one embodiment of this application. The computer-readable storage medium stores program data that can be executed by a processor to implement the steps of any embodiment of the power battery thermal runaway prediction method described above.

[0096] In this embodiment, the computer-readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or a medium capable of storing program data. Alternatively, it can be a server storing the program data, which can send the stored program data to other devices for execution or run the stored program data itself.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0098] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0099] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0100] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting thermal runaway of a power battery, characterized in that, include: Acquire sparse sensing signals from the battery system, including surface temperature, ambient temperature, and electrical parameters at multiple locations. Perform time alignment, cleaning, and confidence marking on the sparse sensing signals to generate standardized input data packets. Based on the input data packet, the battery boundary thermal excitation that cannot be directly measured is equivalent to a parameterized equivalent thermal boundary expression, and the parameters of the equivalent thermal boundary expression are inverted using the surface temperature information in the input data packet to obtain the time-varying equivalent thermal boundary driving quantity. Input the equivalent thermal boundary driving quantity, use the pre-trained multi-physics constraint prediction model to predict the internal temperature field distribution of the cell and the internal state quantity used to characterize the thermal runaway evolution stage, and simultaneously output the surface temperature estimate at the sensor location. Based on the residual between the estimated surface temperature and the measured surface temperature in the sparse sensing signal, the parameters of the equivalent thermal boundary driving quantity, the output bias term of the prediction model, and / or the key scaling factor are adaptively updated to suppress system drift. Based on the calibrated internal temperature field and internal state variables, a risk criterion is constructed, a risk level and / or remaining time estimate are generated, and an early warning signal for control linkage is output.

2. The method for predicting thermal runaway of a power battery according to claim 1, characterized in that, The time alignment of the sparse sensing signal includes: establishing a unified main time axis, performing linear interpolation processing on each channel signal based on adjacent valid sample values, and marking the data reliability according to the interpolation span; The cleaning process includes processing signal loss, signal jumps, signal delays, and signal noise. The signal noise processing employs a weighted fusion method based on the current observation value and the smoothed value of the previous moment to suppress high-frequency jitter.

3. The method for predicting thermal runaway of a power battery according to claim 1, characterized in that, The equivalent thermal boundary expression includes at least one of the following: the equivalent heat flux over time, the equivalent heat transfer coefficient over time, and the equivalent boundary temperature over time. The method of equating the non-measurable battery boundary thermal excitation to a parameterized equivalent thermal boundary expression also includes parameterization processing: expressing the equivalent thermal boundary as a combination of a set of low-dimensional parameters and a preset time basis function, using a finite number of control coefficients to characterize the boundary change trend over the entire time interval, and adjusting the smoothness and expressive power of the boundary change by adjusting the basis function type and the number of control coefficients.

4. The method for predicting thermal runaway of a power battery according to claim 1, characterized in that, The method of inverting the parameters expressed by the equivalent thermal boundary using the surface temperature information in the input data packet includes: constructing and solving an objective function to achieve the inversion. The objective function is used to measure the weighted error between the predicted and measured surface temperature values ​​obtained based on the current boundary parameters. A smoothing constraint and a physical rationality constraint are applied to the rate of change or amplitude of the parameters through a regularization term to suppress high-frequency jitter and overfitting of the inversion results.

5. The method for predicting thermal runaway of a power battery according to claim 1, characterized in that, The multi-physics constraint prediction model adopts a mapping structure that encodes boundary driving information and spatiotemporal coordinate information separately, fuses them in the latent space, and outputs the prediction results at the corresponding coordinates. The prediction model is trained using a total loss function that includes a data fitting term and a physical consistency term. The physical consistency term is used to constrain the model output to satisfy the consistency of the heat conduction equation, reaction kinetic equation, boundary conditions, and initial conditions.

6. The method for predicting thermal runaway of a power battery according to claim 1, characterized in that, The step of adaptively updating the parameters of the equivalent thermal boundary driving quantity, the output bias term of the prediction model, and / or the key scaling factor based on the residual between the estimated surface temperature and the measured surface temperature in the sparse sensing signal includes: calculating the weighted comprehensive error index within the sliding window by comparing the measured values ​​and predicted values ​​of each temperature channel, and triggering a calibration action when the index continuously exceeds a threshold. The calibration action includes adaptive updates to boundary parameters, output bias terms, or scaling factors.

7. The method for predicting thermal runaway of a power battery according to claim 6, characterized in that, When updating the parameters of the equivalent thermal boundary driving quantity, the objective function is solved by constructing a function that includes the fitting error term within the current window and the penalty term for the parameter change relative to the previous time step, so as to limit the parameter update magnitude. And / or, when updating the output deviation term, a step-limited recursive correction method based on the current residual direction is adopted, and the correction magnitude is dynamically adjusted according to the data reliability and operating events through a gating function.

8. The method for predicting thermal runaway of a power battery according to claim 1, characterized in that, The construction of risk criteria to generate risk level and / or remaining time estimate includes: generating a comprehensive risk score by integrating internal temperature level, temperature change trend, internal state quantity level and its change trend, and prediction confidence index; The remaining time estimation includes calculating the estimated time to reach the danger threshold based on the difference between the current maximum internal temperature and the danger threshold and the current rate of temperature rise, and outputting a conservative estimate when the rate of temperature rise is too low or the data confidence is low. The risk levels include at least a warning level, an alert level, and a high-risk level, and a hysteresis comparison and minimum duration maintenance mechanism are used to suppress level jumps.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor coupled to each other, the processor being used to execute program instructions stored in the memory to implement the power battery thermal runaway prediction method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program data that can be executed by a processor to implement the power battery thermal runaway prediction method as described in any one of claims 1-8.

Images