A lithium battery thermal runaway state identification method based on a physical information neural network

Abstract

The present application relates to the technical field of electric variable measurement test, in particular to a lithium battery thermal runaway state identification method based on physical information neural network. The specific implementation process includes: obtaining wide frequency fluctuation current, voltage response trajectory and surface temperature signal, generating multi-dimensional electric measurement data stream and injecting battery physical information network; stripping out dynamic polarization internal resistance sequence and deep thermal source tensor, and constructing electric-thermal cross-domain residual model to quantify charge heating residual; extracting residual gradient flow to correct the electric characteristic reference, and online evaluating the dynamic polarization internal resistance sequence; when the step electric parameter shock wave is identified, the thermal runaway critical blocking signal is output. The present application uses physical information neural network and electric-thermal cross-domain residual model, fuses electric-thermal coupling physical law for online calculation, and continuously corrects the reference drift caused by battery aging by extracting residual gradient flow, so as to realize high-precision, low-delay online measurement and reliable early warning of lithium battery thermal runaway critical state.

Description

Technical Field

[0001] This invention relates to the field of electrical variable measurement and testing technology, specifically to a method for identifying the thermal runaway state of lithium batteries based on a physical information neural network. Background Technology

[0002] With the widespread application of lithium-ion batteries in new energy vehicles, the thermal safety of batteries has become increasingly prominent. Under complex real-world operating conditions, lithium batteries may trigger complex chain-like exothermic side reactions due to overcharging, over-discharging, internal short circuits, or external high temperatures, eventually leading to thermal runaway. To prevent thermal runaway, it is necessary to measure the battery's external electrical variables (such as terminal voltage and charging / discharging current) and surface temperature using conventional sensors to estimate and identify the core thermal state of the battery, which cannot be directly measured, and to provide early warnings. Currently, existing technologies can be broadly divided into two categories: The first category is based on pure mechanistic models. These methods construct complex physical models containing multiple partial differential equations based on the law of conservation of energy, heat transfer, and electrochemical kinetics, and deduce the internal thermal state of the battery through the measurement input of external electrical and thermal parameters. The second category is based on purely data-driven methods, which directly utilize the massive amounts of external electrical variables and temperature features obtained from sensor measurements to train deep learning networks, attempting to fit a nonlinear mapping relationship between external measurement variables and the internal thermal runaway state through a large amount of data.

[0003] However, existing technologies have inherent limitations in practical applications. In purely mechanistic model-based methods, the partial differential equations describing the thermal runaway evolution and internal electrochemical reactions of a battery typically exhibit strong nonlinearity and rigidity. Given the extremely limited embedded computing power of onboard battery systems, solving these equations in real-time online is extremely difficult, resulting in very poor real-time performance of internal state measurements. Furthermore, purely mechanistic models heavily rely on accurate initial physical property parameters, which drift significantly with the degradation of electrical properties caused by battery aging, leading to a substantial decrease in measurement and identification accuracy in the later stages of the battery's lifespan. On the other hand, purely data-driven methods completely lack the constraints of fundamental thermodynamic and electrochemical laws, essentially representing only mathematical probability fitting. When faced with extreme electrical conditions or complex and variable environments not covered in the training dataset, the model is prone to outputting results that violate physical principles such as energy conservation, leading to missed detections or frequent false alarms. Since thermal runaway is a dangerous event with an extremely low probability, obtaining high-quality destructive thermal runaway tag data covering the entire battery life cycle is extremely costly and dangerous. Conventional neural networks cannot establish reliable state assessment and testing benchmarks without sufficient extreme operating condition data.

[0004] In summary, existing technologies, lacking massive amounts of destructive test data, cannot construct an identification network that possesses both signal processing capabilities and ensures that the estimation results strictly adhere to the physical laws of battery electro-thermal coupling. Consequently, it is difficult to achieve high-precision, low-latency online measurement and reliable identification of the critical state of thermal runaway in lithium batteries.

[0005] To address this, a method for identifying the thermal runaway state of lithium batteries based on physical information neural networks is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying the thermal runaway state of lithium batteries based on physical information neural networks, for identifying the thermal runaway state of lithium batteries in new energy vehicles.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for identifying the thermal runaway state of a lithium battery based on a physical information neural network includes:

[0009] The system acquires the spontaneously superimposed broadband fluctuating current in the charging circuit of a new energy vehicle battery, and simultaneously collects the voltage response trajectory and surface temperature signal at the battery terminal; the broadband fluctuating current, voltage response trajectory and surface temperature signal are interleaved and reconstructed in the time domain to generate a multidimensional electrical measurement data stream.

[0010] The multidimensional electrical measurement data stream is injected into the battery physical information network, and the dynamic polarization internal resistance sequence and deep heat source tensor characterizing the current battery polarization characteristics are extracted; an electrothermal cross-domain residual model is constructed, and boundary tests are performed based on the dynamic polarization internal resistance sequence and deep heat source tensor to quantify the charge heating residual reflecting the thermoelectric coupling imbalance state.

[0011] The charge heating residual is forced to collapse to zero as a convergence boundary, and the residual gradient flow is extracted. Based on the residual gradient flow, the electrical characteristic benchmark of the battery physical information network is continuously corrected, and the dynamic polarization internal resistance sequence after being constrained by the deep heat source tensor is evaluated online.

[0012] When a step electrical parameter shock wave is identified in the electrical characteristic reference, the battery charging circuit is cut off and a thermal runaway critical blocking signal is output, thus converging the risk of thermal runaway of the battery within a safe threshold.

[0013] Preferably, the specific implementation process of acquiring the spontaneously superimposed broadband fluctuating current in the charging circuit of a new energy vehicle battery, and simultaneously collecting the voltage response trajectory and surface temperature signal at the battery terminal includes:

[0014] A wideband current transformer is used to continuously monitor the battery charging circuit and capture the broadband fluctuating current that spontaneously superimposes during the charge transfer process. A voltage sensing channel sharing a global clock with the broadband fluctuating current is established to continuously sample the potential difference between battery terminals and generate a voltage response trajectory. Based on the temperature measurement node attached to the battery casing, the surface temperature signal diffused from the battery's internal heat conduction to the external surface is captured. The broadband fluctuating current, voltage response trajectory, and surface temperature signal are synchronously imported into a buffer queue for baseline calibration and anti-aliasing filtering.

[0015] Preferably, the specific implementation process of reconstructing the broadband fluctuating current and voltage response trajectories with the surface temperature signal through time-domain interleaving includes:

[0016] The timestamps of the baseline-calibrated broadband fluctuating current, voltage response trajectories, and surface temperature signals are extracted. Based on the global clock, the broadband fluctuating current, voltage response trajectories, and surface temperature signals are uniformly interpolated to the same time resolution. Following a feature fusion mechanism, the time-aligned broadband fluctuating current, voltage response trajectories, and surface temperature signals are concatenated into a multi-channel time series. A time-domain interleaving reconstruction operation is performed on the multi-channel time series to establish a joint state matrix representing the time-delay correlation between variables. The features of each dimension of the joint state matrix are normalized and tensorized for encapsulation to generate a multi-dimensional electrical measurement data stream.

[0017] Preferably, the specific implementation process of injecting the multidimensional electrical measurement data stream into the battery physical information network and extracting the dynamic polarization internal resistance sequence and deep heat source tensor characterizing the current battery polarization features includes:

[0018] A battery physical information network is established that integrates prior knowledge of electrothermal coupling physics with a deep feature extraction architecture. The multidimensional electrical measurement data stream is forward-injected into the hidden layer of the battery physical information network via the input layer. The battery physical information network is driven to perform online calculations using the internally embedded equivalent circuit physical laws to decouple the dynamic polarization internal resistance sequence containing ohmic impedance and polarization capacitance from the multidimensional electrical measurement data stream. The implicit solution operator of the heat transfer equations is triggered synchronously to map the residual time series features into a deep heat source tensor reflecting the internal heat generation rate distribution.

[0019] Preferably, the specific implementation process of constructing an electrothermal cross-domain residual model, and quantifying the charge-heating residual reflecting the thermoelectric coupling imbalance state by performing boundary tests based on the dynamic polarization internal resistance sequence and the deep heat source tensor, includes:

[0020] Using the dynamic polarization internal resistance sequence and the deep heat source tensor as mutually verifying independent and dependent variables, an electrothermal cross-domain residual model characterizing the mismatch between Joule heat dissipation and actual temperature rise is constructed. Thermoelectric conversion boundary conditions characterizing safe operating conditions are set, and the dynamic polarization internal resistance sequence and the deep heat source tensor are input into these boundary conditions for boundary testing. The difference between the excessive ohmic heat generation caused by abnormal polarization and the theoretical heat transfer distribution in the deep heat source tensor is calculated. This difference is then transformed into an error index through high-dimensional spatial projection, quantifying the charge-heating residual reflecting the internal critical state.

[0021] Preferably, the specific implementation process of forcing the charge-generating residual to collapse towards zero as a convergence boundary and extracting the residual gradient flow includes:

[0022] The charge heating residual is introduced into the loss function of the battery physical information network, and a target optimization equation with the charge heating residual as the penalty term is constructed. The automatic differentiation engine is started to perform backpropagation calculation on the target optimization equation, and the charge heating residual is forced to collapse to zero as the convergence boundary. During the iterative optimization process of network weight update, the direction vector representing the state evolution trend on the loss plane is continuously collected. The residual gradient flow containing nonlinear mapping relationship is extracted along the trajectory direction of the fastest descent of the penalty term.

[0023] Preferably, the specific implementation process for online evaluation of the dynamic polarization internal resistance sequence after deep heat source tensor constraint, based on the electrical characteristic benchmark of the residual gradient flow continuously correcting battery physical information network, includes:

[0024] The extracted residual gradient flow is fed back to the parameter update controller of the battery physical information network; the error compensation direction indicated by the residual gradient flow is used to continuously correct the electrical characteristic reference drift caused by the increase in impedance due to battery aging; the dynamic polarization internal resistance sequence after thermodynamic cross-validation constraint by the deep heat source tensor is obtained, and the degree of deviation of the dynamic polarization internal resistance sequence from the normal polarization trajectory is evaluated online in combination with the dynamically updated electrical characteristic reference; an evaluation index sequence reflecting the dynamic evolution of internal interface impedance and thermal runaway evolution characteristics is output.

[0025] Preferably, when a step electrical parameter shock wave is identified in the electrical characteristic reference, the specific implementation process of cutting off the battery charging circuit and outputting a thermal runaway critical blocking signal to converge the battery's thermal runaway risk within a safety threshold includes:

[0026] The mutation detection logic runs continuously within the monitoring window of the electrical characteristic benchmark. When a nonlinear mutation occurs in the evaluation index sequence input to the decision node, and a step electrical parameter shock wave is identified in the electrical characteristic benchmark, a hardware protection interruption command is triggered. According to the hardware protection interruption command, the battery charging circuit is cut off by a physical switching element, and the injection of external power is blocked. Simultaneously, a thermal runaway critical blocking signal is output through the communication interface. Based on the thermal runaway critical blocking signal, the charging and discharging process of the battery is intervened, and the thermal runaway risk of the battery is converged within a safe threshold.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention collects the spontaneously superimposed broadband fluctuating current in the battery charging circuit and combines it with the battery terminal voltage response trajectory and surface temperature signal for time-domain interleaving reconstruction to construct a multi-dimensional electrical measurement data stream. This enables external measurable electrical parameters to form a unified data expression structure in the time dimension and physical correlation dimension, thereby more completely reflecting the coupling characteristics between charge transfer and thermal diffusion processes inside the battery and improving the ability to perceive abnormal states inside the battery.

[0029] 2. This invention constructs a battery physical information neural network that integrates prior knowledge of electrothermal coupling physics. It embeds equivalent circuit laws and heat transfer mechanisms into the deep learning network structure, so that the network can simultaneously satisfy electrical and thermal constraints when performing state estimation. This enables the decoupled acquisition of dynamic polarization internal resistance sequence and deep heat source distribution information from external measurement data, effectively improving the problems of computational complexity and poor real-time performance of pure mechanism models and the lack of physical constraints in pure data-driven models. It can still achieve internal thermal state estimation under limited on-board embedded computing power.

[0030] 3. This invention constructs an electrothermal cross-domain residual model and introduces charge heating residual as the network convergence boundary to continuously verify the thermoelectric coupling consistency between the dynamic polarization internal resistance sequence and the deep heat source tensor. It also uses residual gradient flow to correct the network's electrical characteristic benchmark online. This enables the invention to maintain stable state identification capability even under battery aging or complex operating conditions. When a step electrical parameter shock wave is detected, the charging circuit is promptly interrupted, thereby achieving early identification and active safety control of the critical state of thermal runaway of lithium batteries, and improving the operational safety and reliability of new energy vehicle batteries. Attached Figure Description

[0031] Figure 1 This is a flowchart of a lithium battery thermal runaway state identification method based on physical information neural network proposed in this invention.

[0032] Figure 2 This is a schematic diagram of the battery physical information network proposed in this invention;

[0033] Figure 3 This is a schematic diagram of the electrothermal cross-domain residual model proposed in this invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It must be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to constitute any limitation on the scope of protection of this invention. Therefore, all equivalent changes or modifications conceived by those skilled in the art based on the content disclosed in this invention without inventive effort should fall within the scope of protection claimed by this invention.

[0035] Reference Figures 1 to 3 This invention provides a method for identifying the thermal runaway state of lithium batteries based on physical information neural networks. The technical solution is as follows:

[0036] Example 1:

[0037] Reference Figure 1 This embodiment proposes a method for identifying the thermal runaway state of lithium batteries based on a physical information neural network, including:

[0038] The system acquires the spontaneously superimposed broadband fluctuating current in the charging circuit of a new energy vehicle battery, and simultaneously collects the voltage response trajectory and surface temperature signal at the battery terminal; the broadband fluctuating current, voltage response trajectory and surface temperature signal are interleaved and reconstructed in the time domain to generate a multidimensional electrical measurement data stream.

[0039] The multidimensional electrical measurement data stream is injected into the battery physical information network, and the dynamic polarization internal resistance sequence and deep heat source tensor characterizing the current battery polarization characteristics are extracted; an electrothermal cross-domain residual model is constructed, and boundary tests are performed based on the dynamic polarization internal resistance sequence and deep heat source tensor to quantify the charge heating residual reflecting the thermoelectric coupling imbalance state.

[0040] The charge heating residual is forced to collapse to zero as a convergence boundary, and the residual gradient flow is extracted. Based on the residual gradient flow, the electrical characteristic benchmark of the battery physical information network is continuously corrected, and the dynamic polarization internal resistance sequence after being constrained by the deep heat source tensor is evaluated online.

[0041] When a step electrical parameter shock wave is identified in the electrical characteristic reference, the battery charging circuit is cut off and a thermal runaway critical blocking signal is output, thus converging the risk of thermal runaway of the battery within a safe threshold.

[0042] Furthermore, the specific implementation process for acquiring the spontaneously superimposed broadband fluctuating current in the charging circuit of a new energy vehicle battery, and simultaneously collecting the voltage response trajectory and surface temperature signal at the battery terminal, includes:

[0043] A wideband current transformer is used to continuously monitor the battery charging circuit and capture the broadband fluctuating current that spontaneously superimposes during the charge transfer process. A voltage sensing channel sharing a global clock with the broadband fluctuating current is established to continuously sample the potential difference between battery terminals and generate a voltage response trajectory. Based on the temperature measurement node attached to the battery casing, the surface temperature signal diffused from the battery's internal heat conduction to the external surface is captured. The broadband fluctuating current, voltage response trajectory, and surface temperature signal are synchronously imported into a buffer queue for baseline calibration and anti-aliasing filtering.

[0044] Specifically, the wideband fluctuating current is acquired by deploying a wideband current transformer in series in the battery charging circuit. The measurement frequency band of the wideband current transformer covers 0.01Hz to 10kHz, with a sampling rate set to 20kSa / s. This ensures the complete capture of the wideband fluctuating current components spontaneously superimposed on the DC charging baseline, which accompany multiple electrochemical processes such as lithium-ion interfacial migration, solid electrolyte interfacial film relaxation, and current collector ohmic voltage drop. In actual testing, the fluctuating current amplitude captured by the transformer during the 0.5C constant current charging phase is approximately 0.3% to 2.1% of the base charging current, with frequency components concentrated in the 5Hz to 500Hz range. This range precisely corresponds to the characteristic frequency range of the charge transfer relaxation process within the battery, providing an effective excitation signal source for subsequent thermal runaway state identification.

[0045] The voltage response trajectory is acquired by establishing a voltage sensing channel that shares a global clock with the wideband fluctuating current. This voltage sensing channel employs a 24-bit high-precision analog-to-digital converter with a resolution of 0.1 μV, continuously sampling the instantaneous potential difference between the positive and negative terminals of the battery at a sampling rate of 20 kSa / s. The global clock is provided by a GPS timing module, with a clock synchronization error not exceeding 100 nanoseconds, thus ensuring accurate preservation of the phase relationship between the current excitation signal and the voltage response signal. In a preferred embodiment, when the battery is in a normal charging state, the steady-state terminal voltage is approximately 3.35V, and the dynamic voltage response amplitude superimposed on it is approximately 0.5mV to 3.2mV. However, when the battery enters the pre-thermal runaway stage (internal temperature exceeding 75°C), the voltage response amplitude in the same frequency band shifts significantly, increasing by approximately 18% to 34% compared to the normal state. This shift constitutes a key input feature for the subsequent physical information neural network identification model.

[0046] The surface temperature signal is acquired by attaching temperature sensing nodes to the surface of the battery casing. These nodes utilize thin-film platinum resistance temperature sensors with a sensitivity of 0.385 Ω / ℃, a measurement accuracy of ±0.1℃, and a response time constant not exceeding 200 ms. The temperature sensing nodes are arranged in a regular array on the six outer surfaces of the battery module, with a node spacing not exceeding 50 mm, ensuring complete capture of the temperature distribution field diffused from the internal heat source to the external casing surface via heat conduction. In a thermal runaway induced by a thermal needle penetration test on the aforementioned battery module, the temperature sensing nodes detected a rapid increase in surface temperature from the initial value of 28.3℃ to 47.6℃ within 8 seconds of triggering thermal runaway, a deviation of only 0.4℃ from the infrared thermal imager reference value, verifying the effective capture capability of the temperature sensing nodes for precursory signals of internal thermal runaway.

[0047] The synchronous import of the wideband fluctuating current, voltage response trajectory, and surface temperature signals is achieved by connecting the outputs of the three sensors in parallel to a multi-channel data acquisition unit sharing a clock reference. The multi-channel data acquisition unit has a built-in 512MB deep buffer queue and uses a first-in-first-out (FIFO) mechanism to sequentially buffer the three raw signals. Baseline calibration is achieved by estimating the mean value of each channel during the silent period before acquisition begins. The DC bias compensation accuracy for the current channel is no less than 0.5mA, the common-mode rejection ratio for the voltage channel is no less than 100dB, and the zero-point drift compensation accuracy for the temperature channel is no less than 0.05℃, ensuring that the subsequent feature extraction stage is not affected by the sensor's own zero-bias error. Anti-aliasing filtering is achieved by deploying a high-order elliptic low-pass filter with a cutoff frequency of 8kHz at the front end of the analog-to-digital converter. Its attenuation at the Nyquist frequency (10kHz) can reach over 60dB, effectively suppressing spectral aliasing distortion introduced by the limited sampling rate and ensuring that the amplitude and phase characteristics of each frequency component in the wideband fluctuating current range from 0.01Hz to 8kHz are accurately reproduced.

[0048] This embodiment utilizes the spontaneously generated broadband fluctuating current during the battery charging process directly as the excitation signal in the battery charging circuit, eliminating the need for any additional active excitation source. This improves upon the requirement of interrupting the charging process and injecting a dedicated excitation current, as is necessary for electrochemical impedance spectroscopy (EIS) techniques. The synchronous acquisition architecture, with three signals sharing a global clock reference, enhances the accuracy of subsequent impedance feature extraction. The high-density array arrangement and high-precision acquisition of the surface temperature signal provide ample response window for subsequent thermal runaway early warning. The quality indicators of the three signals, after baseline calibration and anti-aliasing filtering, provide a high-quality data foundation for the training and inference of the physical information neural network model.

[0049] Furthermore, the specific implementation process of reconstructing the broadband fluctuating current and voltage response trajectories with the surface temperature signal through time-domain interleaving to generate a multi-dimensional electrical measurement data stream includes:

[0050] The timestamps of the baseline-calibrated broadband fluctuating current, voltage response trajectories, and surface temperature signals are extracted. Based on the global clock, the broadband fluctuating current, voltage response trajectories, and surface temperature signals are uniformly interpolated to the same time resolution. Following a feature fusion mechanism, the time-aligned broadband fluctuating current, voltage response trajectories, and surface temperature signals are concatenated into a multi-channel time series. A time-domain interleaving reconstruction operation is performed on the multi-channel time series to establish a joint state matrix representing the time-delay correlation between variables. The features of each dimension of the joint state matrix are normalized and tensorized for encapsulation to generate a multi-dimensional electrical measurement data stream.

[0051] Specifically, the timestamp is extracted by reading the GPS timestamp carried in the header of the data frames output by each sensor. The wideband fluctuating current channel and voltage response channel are both acquired at a rate of 20 kSa / s, with a timestamp resolution of 50 microseconds. The surface temperature signal, limited by the 200 ms response time constant of the thin-film platinum resistance sensor, is acquired at a rate of 5 Sa / s, with a timestamp resolution of 200 ms. During actual data frame inspection, it was found that due to a ±5 ppm frequency drift in the internal clock crystal of the analog-to-digital converter, the cumulative deviation between the timestamps of each channel could reach a maximum of 18 ms after 3600 seconds of continuous acquisition. Without correction, this deviation will introduce a non-negligible phase error in the subsequent impedance feature extraction. Therefore, extracting accurate timestamps is a prerequisite for time-domain interleaving reconstruction.

[0052] The unified interpolation to the same time resolution is achieved by performing hierarchical interpolation on the three signals based on a global clock. The global clock uses a 1PPS (pulses per second) signal output from the GPS timing module as a reference, forcibly aligning the original timestamps of each channel to a unified time axis. For the broadband fluctuating current and voltage response trajectories, since both have a sampling rate of 20 kSa / s, only time axis realignment is required, without introducing interpolation errors. For the surface temperature signal, since its sampling rate is only 5 Sa / s, the interpolation uses cubic spline interpolation to upsample it to 20 kSa / s to match the time resolution of the current and voltage channels. The splicing of the multi-channel time series is achieved by cascading the time-aligned three signals along the channel dimension according to a feature fusion mechanism. The feature fusion mechanism specifies that an original multi-channel matrix is ​​constructed using time steps as row indices and signal channels as column indices. The first column stores broadband fluctuating current sampling values, the second column stores voltage response trajectory sampling values, and the third to eighteenth columns store temperature sampling values ​​distributed on the six outer surfaces of the battery, forming a total of eighteen-channel time sequence.

[0053] The time-domain interleaving reconstruction operation is achieved by sliding a time window across the multi-channel time series and extracting the cross-correlation delays between channels, thereby establishing a joint state matrix characterizing the time-delay correlations between variables. The time window length is set to 1024 sampling points, corresponding to a time span of 51.2 ms. This length covers the main time constants (typically 5 ms to 50 ms) of the charge transfer relaxation process in lithium iron phosphate batteries. Specifically, the time-domain interleaving reconstruction involves calculating the cross-correlation function between the broadband fluctuating current and the voltage response trajectory within each sliding window, and, based on a long-term historical buffer containing multiple consecutive periods, calculating the cross-correlation functions between the broadband fluctuating current and each temperature node, and between the voltage response trajectory and each temperature node, respectively. The time delay corresponding to the cross-correlation peak is extracted, and this time delay is added as an additional feature dimension to the state description of the current window. In actual calculations, the peak delay of the current-voltage cross-correlation under normal charging conditions is stably distributed in the range of 0.15ms to 0.22ms. However, in the pre-thermal runaway stage (internal temperature exceeding 75°C), this delay significantly increases to the range of 0.38ms to 0.61ms, an increase of approximately 1.5 to 2.8 times, indicating that the time-delay correlation characteristics in the joint state matrix are highly sensitive to the thermal runaway state. The joint state matrix is ​​ultimately composed of the original sequence of eighteen channels within each sliding window plus the cross-correlation delay characteristics of each pair of channels.

[0054] The dimensional normalization and tensor encapsulation operations are achieved by performing maximum and minimum normalization on each dimension of the joint state matrix and reshaping it into a four-dimensional tensor. The maximum and minimum normalization uses the extreme value range of each dimension's features obtained statistically from a battery lifecycle test dataset (covering 500 complete charge-discharge cycles) as the normalization boundary, linearly mapping each dimension's feature values ​​to the zero-to-one interval. Specifically, the normalization boundary for the wideband fluctuating current dimension is -15.3A to +15.3A, for the voltage response trajectory dimension it is 2.50V to 3.65V, for the temperature node dimension it is -20℃ to 85℃, and for the cross-correlation delay feature dimension it is 0ms to 2.0ms. After dimensional normalization, the joint state matrix is ​​zero-padded and aligned, and then reshaped into a four-dimensional tensor equal to the batch size multiplied by 1024, 18, and 10. The first dimension is the batch index, the second dimension is the time step dimension, the third dimension is the signal channel dimension, and the fourth dimension is the time delay feature sub-dimension. This four-dimensional tensor constitutes the final output multidimensional electrical measurement data stream, which can be directly used as the standardized input for subsequent physical information neural network models. In a preferred embodiment, the size of a single batch of multidimensional electrical measurement data stream is approximately 8.5MB, and the end-to-end processing latency (from the original signal input to the tensor encapsulation output) is measured to be 23.6ms, meeting the engineering requirement of no more than 50ms response latency for onboard real-time thermal runaway early warning.

[0055] This embodiment improves the time alignment error problem caused by asynchronous acquisition in multimodal battery state monitoring by uniformly interpolating heterogeneous multimodal sensing signals with different sampling rates to the same time resolution. The time-domain interleaving reconstruction operation embeds the time delay correlation information implicit in the electrothermal coupling dynamics into the joint state matrix in the form of numerical features by explicitly modeling the cross-correlation time delay relationship between variables. This enables the model to capture time delay anomalies in the early stage of thermal runaway, improving the sensitivity of early warning.

[0056] Furthermore, the specific implementation process of injecting the multidimensional electrical measurement data stream into the battery physical information network and extracting the dynamic polarization internal resistance sequence and deep heat source tensor characterizing the current battery polarization features includes:

[0057] A battery physical information network is established that integrates prior knowledge of electrothermal coupling physics with a deep feature extraction architecture. The multidimensional electrical measurement data stream is forward-injected into the hidden layer of the battery physical information network via the input layer. The battery physical information network is driven to perform online calculations using the internally embedded equivalent circuit physical laws to decouple the dynamic polarization internal resistance sequence containing ohmic impedance and polarization capacitance from the multidimensional electrical measurement data stream. The implicit solution operator of the heat transfer equations is triggered synchronously to map the residual time series features into a deep heat source tensor reflecting the internal heat generation rate distribution.

[0058] Reference Figure 2Specifically, the battery physical information network is established by embedding the constraints of the equivalent circuit model and the heat transfer equations into the training objective of a deep neural network in the form of a physical residual loss function. The overall architecture of the battery physical information network consists of four functional modules: an input layer, four layers of temporal feature extraction hidden layers, a physical constraint decoupling branch, and a heat source mapping branch. The input layer receives a four-dimensional tensor (the number of batches multiplied by 1024, 18, and 10) as standardized input. The four layers of temporal feature extraction hidden layers adopt a bidirectional long short-term memory network structure, with the number of hidden units in each layer set to 256, 256, 128, and 128 respectively. The activation function is a hyperbolic tangent function, and the dropout ratio is set to 0.15 to suppress overfitting. The physical constraint decoupling branch and the heat source mapping branch share the feature outputs of the first two layers of the bidirectional long short-term memory network and are respectively connected to independent fully connected decoders to output their respective physical quantity estimation results. The electrothermal coupling physical prior knowledge of the battery physical information network is implemented through the embedding of the following two sets of physical constraints: the first is the voltage-current relationship residual derived based on the second-order equivalent circuit model (including ohmic internal resistance, the first polarization resistor and polarization capacitor pair, and the second polarization resistor and polarization capacitor pair); the second is the spatiotemporal evolution residual of the temperature field derived based on the three-dimensional Fourier heat transfer equation. In a preferred embodiment, the physical information network is optimized on a training dataset containing four operating conditions: normal charging and discharging, overcharging, overheating, and needle-induced thermal runaway. The total number of samples in the training set is 12,800 time windows, and the number of samples in the validation set is 3,200 time windows. The Adam optimizer is used, with an initial learning rate set to 0.001, which is decayed to 0.00005 through a cosine annealing strategy. The total number of training rounds is 200, and the optimal validation loss reaches 0.0083 in the 147th round.

[0059] The forward injection of the multidimensional electrical measurement data stream from the input layer to the hidden layer is achieved through a standard forward propagation mechanism. The eighteen-channel original sequence and nine-dimensional time-delay features in the multidimensional electrical measurement data stream are mapped to a unified 512-dimensional embedding space through independent linear projection layers at the input layer and then merged. Subsequently, they are injected into the first-layer bidirectional long short-term memory network in time-step order. The forward computation direction of the bidirectional long short-term memory network captures the electrochemical historical state before the current moment, while the reverse computation direction captures short-term trend information after the current moment. The hidden states in both directions are concatenated at each time step to form a 512-dimensional joint hidden state vector. In the aforementioned embodiment, the forward propagation computation time for a single batch (batch size of 32 time windows) was measured to be 11.3 ms, enabling real-time inference on an onboard computing unit equipped with a quad-core ARM Cortex-A76 processor.

[0060] The decoupling of the dynamic polarization resistance sequence is achieved through online calculation using embedded second-order equivalent circuit physical laws, driven by a physical constraint decoupling branch. This physical constraint decoupling branch consists of two fully connected layers with an output dimension of 5, corresponding to the estimated ohmic resistance, first polarization resistance, first polarization capacitance, second polarization resistance, and second polarization capacitance, respectively. The embedded equivalent circuit physical laws constrain the output of the decoupling branch as follows: In each training step, the wideband fluctuating current input within the current time window and the estimated values ​​of the five circuit parameters output by the decoupling branch are substituted into the time-domain response equation of the second-order equivalent circuit. The root mean square error between the model-predicted voltage and the measured voltage response trajectory is calculated as the circuit physical residual loss. This residual loss and the data-driven loss are weighted and summed at a ratio of 0.6 to 0.4 to form the final training loss. The dynamic polarization resistance sequence is defined as the time series of the sum of the output ohmic resistance and the two polarization resistances over a continuous time window. In a preferred embodiment, under normal charging conditions with an initial state of charge of 80% and a temperature of 25°C, the ohmic internal resistance of the decoupled branch output is stable in the range of 0.82mΩ to 0.89mΩ, the first polarization resistance is stable in the range of 1.14mΩ to 1.31mΩ, and the second polarization resistance is stable in the range of 0.63mΩ to 0.71mΩ, with a total dynamic polarization internal resistance sequence value of approximately 2.59mΩ to 2.91mΩ. However, when the battery enters the pre-thermal runaway stage (internal temperature exceeding 75°C), the total value of the dynamic polarization internal resistance sequence rapidly increases to the range of 3.87mΩ to 5.43mΩ, an increase of 49.4% to 86.5% compared to the normal state. This increase rate deviates from the offline electrochemical impedance spectroscopy measurement results by only 4.2%, verifying the accuracy of the decoupling method.

[0061] The generation of the deep heat source tensor is achieved by synchronously triggering the implicit solution operator of the heat transfer equations embedded in the heat source mapping branch. The heat source mapping branch receives residual temporal features output from the third and fourth layer bidirectional long short-term memory networks. These residual temporal features contain nonlinear dynamic components that cannot be fully explained by the equivalent circuit model, mainly reflecting the spatiotemporal distribution information of the local heat generation rate inside the battery. The heat source mapping branch consists of a three-layer fully connected network with an output dimension of 64, corresponding to the instantaneous heat generation rate estimate of each unit after discretizing the battery module into 64 volumetric units of 4x4x4. This 64-dimensional output vector constitutes the deep heat source tensor. The implicit solver of the heat transfer equations constrains the output of the heat source mapping branch in the following manner: the estimated 64-dimensional heat generation rate is substituted into the finite difference discretization scheme of the three-dimensional unsteady Fourier heat transfer equation of the battery, and the predicted temperature at each temperature measurement node at the next moment is calculated by forward integration. The root mean square error between the predicted temperature and the measured surface temperature signal is calculated as the heat transfer physical residual loss and propagated back to the heat source mapping branch. In a preferred embodiment, under normal charging conditions, the estimated heat generation rate of the 64 volumetric units of the deep heat source tensor is distributed in the range of 0.8 kW / m³ to 2.3 kW / m³, with a relatively uniform distribution and a standard deviation of 0.31 kW / m³. However, in the needle-induced thermal runaway experiment, 6 seconds after the thermal runaway is triggered, the estimated heat generation rate of the local extreme units in the deep heat source tensor rapidly jumps to 187.4 kW / m³, which is about 81 times higher than the normal state. Moreover, the location of the extreme units matches the location of the hotspot captured by the thermal imager with a 91.7% match rate. The heat transfer physical residual loss triggered simultaneously is also about 23 times higher than the normal state at this moment.

[0062] This embodiment embeds the constraints of the second-order equivalent circuit model and the three-dimensional Fourier heat transfer equation into a deep neural network in the form of physical residual loss. This allows the battery physical information network to be subject to strict physical constraints while extracting features driven by data. This improves the poor generalization ability of pure data-driven models due to insufficient training samples or deviations in operating condition distribution. It provides a feature representation with both physical interpretability and high information density for subsequent thermal runaway state identification, thereby improving the overall reliability of battery thermal runaway identification.

[0063] Furthermore, an electrothermal cross-domain residual model is constructed. Boundary tests are performed based on the dynamic polarization internal resistance sequence and the deep heat source tensor. The specific implementation process of quantifying the charge-heating residual reflecting the thermoelectric coupling imbalance state includes:

[0064] Using the dynamic polarization internal resistance sequence and the deep heat source tensor as mutually verifying independent and dependent variables, an electrothermal cross-domain residual model characterizing the mismatch between Joule heat dissipation and actual temperature rise is constructed. Thermoelectric conversion boundary conditions characterizing safe operating conditions are set, and the dynamic polarization internal resistance sequence and the deep heat source tensor are input into these boundary conditions for boundary testing. The difference between the excessive ohmic heat generation caused by abnormal polarization and the theoretical heat transfer distribution in the deep heat source tensor is calculated. This difference is then transformed into an error index through high-dimensional spatial projection, quantifying the charge-heating residual reflecting the internal critical state.

[0065] Reference Figure 3 Specifically, the electrothermal cross-domain residual model is constructed by establishing a baseline relationship between the dynamic polarization internal resistance sequence as the independent variable and the deep heat source tensor as the dependent variable under normal operating conditions. This baseline relationship is based on Joule's law and the principles of electrochemical thermodynamics: under normal charge and discharge conditions, the theoretical heat generation rate of any volume cell within the battery mainly consists of three parts: ohmic heat, polarization heat, and reversible entropy heat. Ohmic heat is determined by the product of the square of the local current and the ohmic internal resistance of the cell, while polarization heat is determined by the product of the polarization resistance and the square of the polarization current. These two, along with the reversible heat generated by the entropy change due to the electrochemical reaction, together constitute the theoretical total heat generation power of the cell. The electrothermal cross-domain residual model uses Gaussian process regression as the core modeling method. It takes the statistical characteristics (mean, standard deviation, and rate of change, totaling 3 dimensions) of the dynamic polarization internal resistance sequence within 100 consecutive time windows as input features and the theoretical predicted value of the heat generation rate of the 64 volume units of the deep heat source tensor at the corresponding time as the output target. The Gaussian process regression model is fitted on 8000 time window samples including normal charging and discharging conditions. The kernel function used is a weighted combination of radial basis function and linear kernel function. The noise variance hyperparameter is determined to be 0.0042 by maximum likelihood estimation.

[0066] The thermoelectric conversion boundary conditions are set by statistically analyzing the confidence boundaries of the joint distribution of the dynamic polarization internal resistance sequence and the deep heat source tensor under normal operating conditions. Specifically, the thermoelectric conversion boundary conditions are defined as follows: within the full range of state of charge (20% to 100%) and temperature (-20°C to 45°C), operating condition cells are divided with a 0.5% state of charge step and a 5°C temperature step, respectively. Within each operating condition cell, the mean and three standard deviations of the total value of the dynamic polarization internal resistance sequence in the normal operating sample are used as the polarization internal resistance boundary. Similarly, the mean and three standard deviations of the sum of the heat generation rates of the 64 units of the deep heat source tensor are used as the total heat generation rate boundary. These two types of boundaries together constitute a thermoelectric conversion boundary condition lookup table covering the entire operating condition domain. In the measured data of the aforementioned battery module, the dynamic polarization internal resistance boundary for an 80% state of charge and a temperature of 25 degrees Celsius ranged from 2.21 mΩ to 3.18 mΩ, and the total heat generation boundary ranged from 68.4 W to 127.6 W. For a 50% state of charge and a temperature of 40 degrees Celsius, the dynamic polarization internal resistance boundary ranged from 2.47 mΩ to 3.52 mΩ, and the total heat generation boundary ranged from 74.1 W to 138.9 W. The real-time output dynamic polarization internal resistance sequence and the deep heat source tensor were compared with the boundary conditions corresponding to the current operating condition. If either quantity exceeded the boundary, a boundary test anomaly flag was triggered, and the excess quantity and direction were recorded for subsequent residual calculation. In the needle-induced thermal runaway experiment, the boundary test anomaly flag was first activated 9.3 seconds before thermal runaway was triggered, approximately 5.1 seconds earlier than the traditional method that relies solely on the surface temperature threshold, verifying the boundary test's sensitive ability to capture early signs of thermal runaway.

[0067] The quantification of the charge-heating residual is achieved by calculating the difference between the excessive ohmic heat generation caused by abnormal polarization and the theoretical heat transfer distribution of the deep heat source tensor, and then transforming it into a scalar error index through high-dimensional spatial projection. The calculation of the excessive ohmic heat generation is based on the excess amount in the dynamic polarization internal resistance sequence that exceeds the upper limit of the corresponding operating condition grid: the difference between the total value of the dynamic polarization internal resistance sequence at the current moment and the upper limit of the boundary is multiplied by the square of the current charging current, and combined with the local heating weight allocation vector, to obtain the excessive ohmic heat generation power distribution vector in 64 volume unit dimensions, with a dimension of 64. The theoretical heat transfer distribution is provided by the theoretical value of the deep heat source tensor predicted and output by the electrothermal cross-domain residual model based on the current dynamic polarization internal resistance sequence, also with a dimension of 64. The difference vector is obtained by calculating the element-wise difference between the excess ohmic heat generation power distribution vector and the residual vector obtained by subtracting the theoretical value from the measured value of the deep heat source tensor. This difference vector has a dimension of 64, and each element reflects the degree of local electrothermal coupling imbalance of the corresponding volume unit. The high-dimensional spatial projection is achieved by projecting a 64-dimensional difference vector onto the directions of the first eight principal components determined in advance by principal component analysis of normal operating condition samples, and then calculating the weighted sum of squares of the scores of each principal component after projection. The weights are determined by the proportion of variance explained by each principal component, ultimately yielding a scalar charge-heating residual index. The principal component analysis is performed on the 64-dimensional difference vector of 8000 time window samples under normal operating conditions. The first eight principal components cumulatively explain 93.7% of the variance, which serves as the basis for the high-dimensional spatial projection. In a preferred embodiment, under normal charging conditions, the charge heating residual index is stably distributed in the range of 0.0 to 0.18, with an average value of 0.07. In the pre-overcharge induced thermal runaway stage (internal temperature exceeding 65 degrees Celsius), the charge heating residual index rapidly rises to the range of 0.43 to 0.87, with an average value of 0.61, which is about 7.6 times higher than the normal state. In the needle-induced thermal runaway experiment, 7.2 seconds before the thermal runaway is triggered, the charge heating residual index exceeds the preset judgment threshold of 0.35, triggering the pre-thermal runaway warning, which is 4.8 seconds earlier than the warning method based on a single temperature threshold. In the full life cycle test of 500 complete charge-discharge cycles, the false alarm rate of the charge heating residual index exceeding the 0.35 threshold under normal conditions is only 0.23%.

[0068] This embodiment constructs an electrothermal cross-domain residual model, incorporating the dynamic polarization internal resistance sequence and the deep heat source tensor into a unified electrothermal coupling consistency evaluation framework. This improves the problem of independent analysis of electric domain features and thermal domain features, which makes it impossible to quantify the degree of coupling deviation between the two. It provides core feature inputs with high signal-to-noise ratio and high physical interpretability for subsequent final state determination based on thermal runaway identification networks, thereby improving the reliability and timeliness of thermal runaway early warning of the overall system.

[0069] Furthermore, the specific implementation process of forcing the charge-heating residual to collapse towards zero as a convergence boundary and extracting the residual gradient flow includes:

[0070] The charge heating residual is introduced into the loss function of the battery physical information network, and a target optimization equation with the charge heating residual as the penalty term is constructed. The automatic differentiation engine is started to perform backpropagation calculation on the target optimization equation, and the charge heating residual is forced to collapse to zero as the convergence boundary. During the iterative optimization process of network weight update, the direction vector representing the state evolution trend on the loss plane is continuously collected. The residual gradient flow containing nonlinear mapping relationship is extracted along the trajectory direction of the fastest descent of the penalty term.

[0071] Specifically, the objective optimization equation is constructed by weighting the charge heating residual penalty term and adding it to the original loss function of the battery physical information network. The original loss function of the battery physical information network is composed of circuit physical residual loss and heat transfer physical residual loss weighted at a ratio of 0.6 to 0.4. After introducing the charge heating residual penalty term, the objective optimization equation consists of a weighted sum of three parts: circuit physical residual loss, heat transfer physical residual loss, and charge heating residual penalty term, with weights of 0.4, 0.3, and 0.3, respectively. The charge heating residual penalty term directly adopts the scalar charge heating residual index quantized in the aforementioned steps. Its physical meaning is a measure of the degree of electrothermal coupling imbalance under the current network parameter state. Incorporating it into the loss function explicitly drives the network to converge towards electrothermal coupling self-consistency during gradient descent optimization. The convergence boundary is set such that the charge-heating residual penalty term approaches zero. This requires that the network's prediction results of the dynamic polarization internal resistance sequence and the deep heat source tensor under any input conditions satisfy the Joule heat conservation constraint, and does not allow any unexplained electrothermal coupling deviations. In a preferred embodiment, after introducing the charge-heating residual penalty term, the value of the objective optimization equation in the initial training round is 0.1847, an increase of 67.5% compared to the original loss function value of 0.1103 without the penalty term. This indicates that the charge-heating residual penalty term imposes significant additional constraint pressure on the network parameters, pushing the network towards more stringent physical consistency optimization.

[0072] The backpropagation computation is achieved by initiating a computational graph tracing mechanism based on the PyTorch automatic differentiation engine. During the forward propagation phase, the automatic differentiation engine constructs a complete computational graph, recording all operator links from the input multidimensional electrical measurement data through a four-layer bidirectional long short-term memory network, a physical constraint decoupling branch, a heat source mapping branch, and finally to the charge-heating residual penalty term. During the backpropagation phase, the automatic differentiation engine traverses the computational graph backwards, calculating the partial derivatives of the objective optimization equation with respect to each network parameter layer by layer according to the chain rule. The convergence boundary forcing the charge-heating residual to collapse to zero is achieved by preserving the complete gradient contribution of the charge-heating residual penalty term to all network parameters in the backpropagation path of the automatic differentiation engine, without any truncation or scaling of the gradient of this penalty term, ensuring that the gradient signal of the electrothermal coupling conservation constraint can be transmitted to the lower-level network parameters near the input layer without attenuation. In the aforementioned embodiment, the gradient norm of the charge heating residual penalty term with respect to the weight matrix of the first layer bidirectional long short-term memory network was measured to be 0.0031, which is 1 to 2.87 compared with the gradient norm of the circuit physical residual loss with respect to the weight matrix of the same layer, which is 0.0089. This indicates that the gradient signal of the charge heating residual penalty term still maintains sufficient strength at the low-level network parameters, and can effectively influence the adjustment of the low-level feature extraction method towards electrothermal self-consistency.

[0073] The continuous acquisition of the direction vector is achieved by recording the gradient vector of the target optimization equation with respect to the output of the intermediate layers of the network during each iteration of network weight updates. Specifically, the direction vector is defined as the partial derivative of the target optimization equation with respect to the 128-dimensional hidden state vector output by the third-layer bidirectional long short-term memory network. This layer's output is located before the bifurcation point between the physical constraint decoupling branch and the heat source mapping branch, and its gradient direction comprehensively reflects the evolution trend of the network state under the combined influence of circuit physical constraints and heat transfer physical constraints. After each training batch completes backpropagation, the 128-dimensional direction vector of the current batch is appended and stored in the direction vector history buffer. This buffer uses a sliding window mechanism to retain the direction vector records of the most recent 500 training batches, with a total storage size of approximately 32MB. In a preferred embodiment, principal component analysis of the direction vector history buffer revealed that the direction vectors of samples corresponding to normal charging conditions are mainly concentrated in the subspace spanned by the first three principal components, explaining 78.3% of the cumulative variance. In contrast, the projection energy of the direction vectors of samples corresponding to thermal runaway precursor conditions is significantly enhanced in the directions of the 4th to 9th principal components, with the cumulative variance explained in this interval jumping from 11.2% in normal conditions to 34.7% in thermal runaway precursor conditions. This indicates that the high-dimensional distribution structure of the direction vectors has a high discriminative power for battery operating states.

[0074] The residual gradient flow is extracted by integrating and aggregating the direction vectors in successive iterations along the trajectory direction of the fastest descent of the charge-heating residual penalty term. The fastest descent trajectory direction is determined in each iteration by the negative gradient direction output by the objective optimization equation at the intermediate layer of the network, i.e., the direction indicated by the negative direction vector. The integration and aggregation employs an exponential moving average mechanism to weight and accumulate the negative direction vectors in successive iterations, with a decay coefficient set to 0.97. This filters out the interference of single-batch stochastic gradient noise on the trajectory direction estimation, ensuring that the extracted residual gradient flow reflects the true descent trend on the loss plane rather than random perturbations. The residual gradient flow is ultimately represented as a 128-dimensional vector sequence, with each vector corresponding to a training iteration step. The entire sequence describes the state evolution path of the network parameters under the constraints of the objective optimization equation, embedding the electrothermal coupling nonlinear mapping relationship revealed during the collapse of the charge-heating residual from its initial non-zero value to zero. In a preferred embodiment, the residual gradient flow sequence extracted from training rounds 1 to 147 (the optimal validation loss rounds) has a length of 4700 iterations. The L2 norm of this sequence monotonically decreases from 0.0412 at the beginning of training to 0.0071 at convergence, a decrease of 82.8%. The convergence curve shows a sudden change in slope between rounds 89 and 121. Analysis shows that this sudden change corresponds to the phase transition point in the electrothermal coupling representation of the network from an underconstrained state to a fully self-consistent state. Before and after this phase transition point, the F1 score of the network on the validation set of thermal runaway precursor conditions jumps from 0.841 to 0.907, an increase of 7.8 percentage points.

[0075] This embodiment explicitly incorporates the charge-heating residual penalty term into the objective optimization equation and forces it to collapse towards zero. This ensures that the battery physics information network is always rigidly driven by the Joule thermal conservation physical constraint during parameter optimization. This improves the physical uninterpretability problem in deep learning methods where network outputs may violate the fundamental law of energy conservation, elevating physical rationality from a soft constraint to a mandatory guarantee at the convergence boundary level. The automatic differentiation engine fully preserves the gradient of the charge-heating residual penalty term, ensuring that the electrothermal coupling constraint signal can be effectively transmitted to the parameters of the underlying feature extraction network. This provides a high-dimensional feature input with both dynamic evolution information and physical constraint information for the subsequent thermal runaway state identification network.

[0076] Furthermore, based on the electrical characteristic benchmark of the residual gradient flow continuously correcting battery physical information network, the specific implementation process for online evaluation of the dynamic polarization internal resistance sequence after deep heat source tensor constraint includes:

[0077] The extracted residual gradient flow is fed back to the parameter update controller of the battery physical information network; the error compensation direction indicated by the residual gradient flow is used to continuously correct the electrical characteristic reference drift caused by the increase in impedance due to battery aging; the dynamic polarization internal resistance sequence after thermodynamic cross-validation constraint by the deep heat source tensor is obtained, and the degree of deviation of the dynamic polarization internal resistance sequence from the normal polarization trajectory is evaluated online in combination with the dynamically updated electrical characteristic reference; an evaluation index sequence reflecting the dynamic evolution of internal interface impedance and thermal runaway evolution characteristics is output.

[0078] Specifically, the operation of feeding the residual gradient flow back to the parameter update controller is achieved by inputting the extracted 128-dimensional residual gradient flow vector sequence into the online adaptive parameter update controller of the battery physical information network. The parameter update controller is constructed using a meta-learning framework, with its core being a two-layer gated recurrent unit network. The input dimension is 128, the hidden layer dimension is 64, and the output dimension is consistent with the dimension of the subset of parameters that need to be adaptively updated online in the battery physical information network, totaling 1847 parameters. This parameter subset covers the weights and biases of the fully connected layers in the physical constraint decoupling branch, as well as the weights and biases of the first two fully connected layers in the heat source mapping branch. In the offline phase, the parameter update controller uses simulation data simulating battery aging trajectories (covering aging levels from 0 to 500 cycles) as the training scenario, defining battery adaptation tasks at different aging levels as a set of independent tasks under the meta-learning framework. The outer-layer meta-optimization objective is: given a residual gradient flow sequence, after applying the parameter increments from the controller output to the battery physical information network, to ensure that the charge heating residual penalty term converges to a preset threshold of 0.05 within the minimum number of iterations (with a maximum of 5 steps). This threshold is used as the meta-loss function. The Adam optimizer is employed for outer-layer updates of the controller weights. The meta-learning rate is set to 0.0003, the meta-training batch size is 16 aging tasks, and the total number of meta-training rounds is 500. Simulated aging trajectory data is generated based on a semi-empirical cyclic aging model, with capacity decay to 80% of the initial value as the lifetime termination condition. This covers three charge / discharge rate combinations: 0.5C, 1C, and 2C, generating a total of 1500 aging trajectory samples. This ensures that the charge heating residual penalty term, after parameter updates, converges to near zero within the minimum number of iterations. The generation and application of the residual gradient flow during the vehicle deployment phase are achieved through a two-stage mechanism: In the offline pre-training phase, a complete residual gradient flow sequence is extracted from the training iterations using the aforementioned method, and this sequence is used as the training input for the parameter update controller, enabling the controller to learn the mapping relationship from gradient flow morphology to parameter increments. In the vehicle-mounted online deployment phase, the battery physical information network operates in inference mode, without performing complete backpropagation. The parameter update controller constructs a lightweight approximate gradient proxy signal based on the difference between the current inference output charge heating residual index value and its historical moving average. This proxy signal replaces the training gradient flow as the controller input, driving the real-time update of 1847 online adaptive parameter subsets. The calculation of the proxy signal involves only forward differential operations, with a single-step calculation time not exceeding 0.8ms, meeting the real-time requirements of vehicle deployment.

[0079] The continuous correction of the electrical characteristic reference drift is achieved by using the error compensation direction indicated by the residual gradient flow to directionally constrain the parameter increment output by the parameter update controller. The electrical characteristic reference is defined as the expected prediction range of the dynamic polarization internal resistance sequence corresponding to normal operating conditions under the current aging state of the battery physical information network. This reference drifts with battery cyclic aging: in a preferred embodiment, after 100 complete charge-discharge cycles, the average total value of the dynamic polarization internal resistance sequence under normal charging conditions (80% state of charge, 25 degrees Celsius) drifts from an initial 2.73 mΩ to 3.04 mΩ, a drift of 0.31 mΩ, representing a drift amplitude of 11.4%; after 300 cycles, it further drifts to 3.61 mΩ, with a cumulative drift of 0.88 mΩ, representing a drift amplitude of 32.2%. If this aging-induced reference drift is not corrected, the network will misjudge the impedance increase caused by normal aging as a precursor to thermal runaway, leading to a significant increase in the false alarm rate. The correction operation specifically involves separating the low-frequency drift component caused by aging in the residual gradient flow (extracted by performing a long-term moving average filter across charge-discharge cycles on the residual gradient flow vector sequence) from the parameter increments output by the parameter update controller. The parameter increments corresponding to this low-frequency drift component are then applied directionally to the mean and variance estimation parameters of the electrical characteristic benchmark, causing the electrical characteristic benchmark to adaptively shift upwards with the degree of aging. The parameter increments corresponding to the high-frequency dynamic components are retained for sensitive detection of thermal runaway precursors. The window length of the long-term moving average filter is set to cover the number of residual gradient flow samples corresponding to 10 consecutive complete charge-discharge cycles. The start and end times of each cycle are detected using the current integral method as the basis for dynamically updating the window boundaries. Under typical operating conditions where a single charge-discharge cycle lasts approximately 45 minutes, the corresponding window covers the gradient flow records of the most recent 27,000 training iterations. The boundary frequency between low-frequency drift components and high-frequency dynamic components is set at 0.002 cycles / cycle (corresponding to a timescale of approximately 500 cycles). Components below this frequency are used for adaptive updates of the baseline mean and variance of electrical characteristics, while components above this frequency are retained for sensitive detection of thermal runaway precursors. In the aforementioned embodiment, after the battery underwent 300 cycles of aging following the correction, the false alarm rate for detecting thermal runaway precursors decreased from 8.7% in the uncorrected state to 1.2%, a reduction of 86.2%, while the false alarm rate for thermal runaway decreased from 3.1% to 0.8%.

[0080] The thermodynamic cross-validation constraint is achieved by comparing the measured values ​​of the deep heat source tensor with the theoretical values ​​predicted by the electrothermal cross-domain residual model element by element, and selecting the estimated dynamic polarization resistance sequence values ​​that satisfy the thermodynamic self-consistency condition. The thermodynamic self-consistency condition is defined as the mean absolute error between the measured and theoretically predicted values ​​of the heat generation rate in the 64 volumetric elements of the deep heat source tensor not exceeding twice the standard deviation of the theoretically predicted value under the current operating condition. For time windows that do not satisfy the thermodynamic self-consistency condition, the corresponding estimated dynamic polarization resistance sequence values ​​are marked as thermodynamically unreliable samples and do not participate in subsequent online evaluation calculations; instead, they are replaced by the estimated dynamic polarization resistance sequence values ​​from the previous time window that satisfy the thermodynamic self-consistency condition. In a preferred embodiment, the proportion of thermodynamically unreliable samples under normal charging conditions is about 2.3%, mainly occurring in the first 30 seconds before the battery temperature stabilizes at the beginning of charging. In the pre-thermal runaway stage, the proportion of thermodynamically unreliable samples drops sharply to 0.4%. This is because the electrothermal coupling relationship tends to be stronger at this stage, and the measured and theoretical values ​​of the deep heat source tensor both show a consistent rapid growth trend, which in turn improves the satisfaction rate of thermodynamic self-consistency conditions. This feature itself also constitutes an auxiliary criterion for identifying thermal runaway states.

[0081] The online evaluation is achieved through multidimensional deviation analysis between the dynamically polarized internal resistance sequence constrained by thermodynamic cross-validation and the dynamically updated electrical characteristic benchmark. The deviation analysis is conducted in three dimensions: first, amplitude deviation, calculated by dividing the difference between the total value of the current dynamically polarized internal resistance sequence and the mean value of the electrical characteristic benchmark by the standard deviation of the electrical characteristic benchmark; second, spectral deviation, by performing a short-time Fourier transform on the dynamically polarized internal resistance sequence to extract the impedance spectrum feature vector in the 5Hz to 500Hz frequency band, and calculating the cosine distance between this vector and the corresponding spectral template of the electrical characteristic benchmark; and third, dynamic evolution deviation, calculated by comparing the mean rate of change of the dynamically polarized internal resistance sequence over 20 consecutive time windows with the mean rate of change of the corresponding electrical characteristic benchmark. The three-dimensional deviation scores are weighted and fused (with weights of 0.5, 0.3, and 0.2 respectively) to output a single evaluation index. The evaluation index sequence over consecutive time windows constitutes an evaluation index sequence reflecting the dynamic evolution of internal interface impedance and the characteristics of thermal runaway evolution. In a preferred embodiment, the mean of the evaluation index sequence under normal charging conditions is 0.083, and the standard deviation is 0.031. During the overcharge-induced thermal runaway precursor stage (internal temperature range of 65°C to 75°C), the mean of the evaluation index sequence increases to 0.412, and the standard deviation increases to 0.087. During the needle-induced thermal runaway precursor stage, the evaluation index sequence exceeds the preset warning threshold of 0.30 8.6 seconds before thermal runaway is triggered, approximately 2.3 seconds earlier than the evaluation scheme using only the dynamic polarization internal resistance sequence without thermodynamic cross-validation constraints, further reducing the false alarm rate from 2.8% to 0.9%. The evaluation index sequence is continuously refreshed with an output period of 100ms, and each output period contains an evaluation history window of 2000 time steps, ensuring that the subsequent thermal runaway identification network can obtain a sufficiently long length of evaluation trend information.

[0082] This embodiment achieves adaptive online correction of electrical feature benchmarks as battery aging progresses by feeding the residual gradient flow back to the parameter update controller of the meta-learning architecture. This improves upon the misjudgment problem caused by the fixed benchmark method neglecting battery aging. The thermodynamic cross-validation constraint mechanism uses the deep heat source tensor to monitor the thermodynamic reliability of the dynamically polarized internal resistance sequence estimate in real time. It removes abnormal estimates that do not meet the energy conservation constraint from the evaluation input, improving the physical reliability of the evaluation results without adding additional sensors. This provides richer, more timely, and more reliable input features for the final determination of the thermal runaway state.

[0083] Furthermore, when a step electrical parameter shock wave is identified in the electrical characteristic reference, the specific implementation process of cutting off the battery charging circuit and outputting a thermal runaway critical blocking signal to converge the battery's thermal runaway risk within a safety threshold includes:

[0084] The mutation detection logic runs continuously within the monitoring window of the electrical characteristic benchmark. When a nonlinear mutation occurs in the evaluation index sequence input to the decision node, and a step electrical parameter shock wave is identified in the electrical characteristic benchmark, a hardware protection interruption command is triggered. According to the hardware protection interruption command, the battery charging circuit is cut off by a physical switching element, and the injection of external power is blocked. Simultaneously, a thermal runaway critical blocking signal is output through the communication interface. Based on the thermal runaway critical blocking signal, the charging and discharging process of the battery is intervened, and the thermal runaway risk of the battery is converged within a safe threshold.

[0085] Specifically, the continuous operation of the mutation detection logic is achieved by deploying an online mutation detection module based on a Bayesian change point detection algorithm within the monitoring window of the electrical characteristic benchmark. The length of the monitoring window is set to 200 evaluation index sequence output cycles, corresponding to a time span of 20 seconds. This length is sufficient to cover the typical time span (6 to 14 seconds in actual measurements) during the pre-thermal runaway stage of lithium iron phosphate batteries, where the evaluation index sequence climbs from the baseline level to the mutation point. The Bayesian change point detection algorithm uses the posterior probability distribution of the evaluation index sequence within the monitoring window as the detection basis. In each new output cycle, the latest evaluation index value is appended to the monitoring window, and the posterior probability distribution is updated to calculate the posterior probability of a statistically significant mutation occurring at the current moment. The triggering condition for the mutation detection logic is set to a posterior probability exceeding 0.92, and this condition is met for three consecutive output cycles, i.e., the mutation posterior probability remains above 0.92 for 300 ms. This ensures trigger sensitivity while avoiding false triggering caused by single-point noise. In a preferred embodiment, the mutation detection logic has zero false triggers in 500 cycles of testing under normal charging conditions; in the overcharge-induced thermal runaway experiment, the mutation detection logic first meets the continuous triggering condition 7.4 seconds before thermal runaway is triggered; in the needle-induced thermal runaway experiment, this moment is 8.1 seconds before thermal runaway is triggered. The early triggering time in both scenarios meets the minimum response time margin required for subsequent hardware protection actions.

[0086] The identification of the step electrical parameter shock wave is achieved by performing step feature extraction on the time-domain waveform of the dynamic polarization internal resistance sequence in the electrical characteristic reference simultaneously with the triggering of the sudden change detection logic. The step electrical parameter shock wave is defined as an event in which the amplitude jump of the dynamic polarization internal resistance sequence exceeds three times the standard deviation of the electrical characteristic reference within a single time window (51.2 ms) and the jump direction is positively increasing. This event physically corresponds to the phenomenon of sudden increase in interface impedance caused by local short circuit inside the battery or large-area collapse of the SEI film. The step feature extraction is achieved by calculating the first-order difference of the dynamic polarization internal resistance sequence between adjacent time windows and comparing it with the step judgment threshold. The step judgment threshold is dynamically calculated based on the standard deviation of the electrical characteristic reference under the current operating condition. In the aforementioned embodiment, the step judgment threshold corresponding to the operating condition of 80% state of charge and 25 degrees Celsius is 0.93 mΩ. In the acupuncture-induced thermal runaway experiment, a step electrical parameter shock wave with an amplitude jump of 1.47 mΩ appeared 6.8 seconds before thermal runaway was triggered, which is 1.58 times the corresponding operating condition judgment threshold. The duration was 153.6 ms (3 time windows). The step feature extraction module completed the identification and output the shock wave confirmation flag within the first time window (51.2 ms) after the shock wave appeared. The hardware protection interrupt command is generated immediately when the mutation detection logic trigger flag and the step electrical parameter shock wave confirmation flag are both valid. The logic AND judgment of the two flags is implemented in hardware logic in a dedicated field-programmable gate array chip, and the judgment delay does not exceed 1 microsecond, thereby ensuring that the time delay from the completion of shock wave identification to the generation of the hardware protection interrupt command meets the microsecond-level real-time requirement.

[0087] The charging circuit is shut down by a hardware protection interrupt command driving the physical switching element to turn off. The physical switching element is an insulated-gate bipolar transistor (IGBT) module with a rated current of 250A and a rated voltage of 60V. This module is deployed in series on the positive bus of the battery charging circuit. Its gate drive circuit directly responds to the hardware protection interrupt command from the field-programmable gate array (FPGA) chip. The actual measured shutdown time (from the effective execution of the command to the current dropping below 10% of the rated value) is 8.3 microseconds. After the IGBT module is turned off, the residual inductance energy stored in the charging circuit is released through a parallel freewheeling diode and an energy absorption resistor. The resistance value of the energy absorption resistor is designed to ensure that the peak value of the shutdown overvoltage does not exceed 120% of the rated voltage, i.e., not exceeding 72V, to avoid secondary impact on the battery terminals caused by the shutdown overvoltage. In the aforementioned needle-induced thermal runaway experiment, the total time delay from the completion of step electrical parameter shock wave identification to the charging circuit current dropping to zero was 8.3 microseconds. The margin between the charging circuit cutoff time and the thermal runaway trigger time was 6.8 seconds. During this margin time, the highest surface temperature of the battery rose from 47.3 degrees Celsius to 53.1 degrees Celsius, with a rise rate of 0.88 degrees Celsius per second, which is far below the thermal runaway critical temperature (the thermal runaway trigger threshold of lithium iron phosphate batteries is about 130 to 150 degrees Celsius). This verifies the effective containment effect of the charging circuit cutoff action on the spread of thermal runaway risk.

[0088] The output of the thermal runaway critical interruption signal is achieved by simultaneously sending a standardized message to the upper-level battery management system and vehicle controller via the communication interface at the same time as the charging circuit cutoff action is triggered. The communication interface uses a controller area network (CLAN) bus conforming to the ISO 11898 standard, with a baud rate set to 1 Mbit / s and a message priority identifier set to the highest priority (identifier value 0x000), ensuring that the thermal runaway critical interruption signal message always obtains the highest transmission priority in bus contention. The measured delay from the triggering of the charging circuit cutoff action to the completion of the first message transmission by the communication interface is 120 microseconds. The thermal runaway critical interruption signal message contains seven data fields: trigger time timestamp, current charge heating residual index value, step electrical parameter shock wave amplitude, shock wave duration, current estimated state of charge value, current highest surface temperature value, and suggested intervention measure code. The total message length is 8 bytes, conforming to the standard CLAN data frame format. Upon receiving a thermal runaway critical interruption signal, the upper-level battery management system executes a tiered response based on the recommended intervention measure code: when the recommended intervention measure code is Level 1, the battery management system only suspends charging and initiates forced cooling; when the code is Level 2, it simultaneously disconnects the discharge circuit and triggers the emergency cooling program of the vehicle thermal management system; when the code is Level 3, it disconnects all charging and discharging circuits while issuing an audible and visual alarm to the driver and pushing a warning notification to the cloud monitoring platform. In a preferred embodiment, the thermal runaway critical interruption signal triggers a Level 2 response. The vehicle thermal management system initiates emergency cooling within 1.2 seconds of receiving the message. The highest internal temperature of the battery drops from 68.4 degrees Celsius to 51.7 degrees Celsius after the emergency cooling intervention, with a temperature drop rate of 7.8 degrees Celsius per minute. Ultimately, the risk of thermal runaway completely converges to within the safety threshold (surface temperature below 45 degrees Celsius) within 240 seconds after triggering, and no thermal runaway accident occurs.

[0089] This embodiment improves upon the inherent trade-off between sensitivity and false alarm rate in single-threshold judgment methods by jointly using two complementary detection mechanisms—Bayesian variable point detection and step electrical parameter shock wave identification—in the form of hardware logic AND gates. By physically cutting off the external energy supply to prevent thermal runaway, differentiated response measures can be taken based on the level of thermal runaway risk, avoiding unnecessary driving interruptions caused by a uniform hard power cut, and ensuring the safe operation of the new energy vehicle's power battery system throughout its entire lifecycle.

[0090] Example 2:

[0091] This embodiment fully deploys the aforementioned lithium battery thermal runaway state identification method based on physical information neural networks in a pure electric passenger vehicle equipped with a ternary lithium-ion power battery pack. The battery pack consists of 192 prismatic cells with a rated capacity of 50Ah and a nominal voltage of 3.65V, arranged in a 96-series-2-parallel configuration, with a total rated energy of 35kWh. The battery pack underwent systematic verification under four operating conditions in actual road operation and DC fast charging scenarios: normal charging and discharging, low-temperature charging, high-rate discharging, and externally induced internal short circuit.

[0092] During the wideband fluctuating current acquisition and multi-mode signal synchronous acquisition stage, the bandwidth current transformer measures a frequency band from 0.01Hz to 15kHz, with a sampling rate set to 25kSa / s, and is deployed in series in the positive main circuit of the battery pack. Under 1C DC fast charging conditions, the captured fluctuating current amplitude is approximately 0.4% to 2.6% of the base charging current, with frequency components mainly concentrated in the 8Hz to 600Hz range. The voltage sensing channel uses a 26-bit high-precision analog-to-digital converter with a resolution of 0.08μV and a common-mode rejection ratio of no less than 105dB. The 22 Pt1000 temperature sensing nodes have a measurement accuracy of ±0.08 degrees Celsius, are acquired at a rate of 6Sa / s, and are distributed in a regular array on the surface of the battery pack casing with a maximum spacing of no more than 45mm. The three signals are synchronously imported into the buffer queue after baseline calibration and sixth-order Butterworth anti-aliasing filtering, with a clock synchronization error of no more than 80 nanoseconds for each channel.

[0093] During the multidimensional electrical measurement data stream generation stage, the temperature signal was upsampled to 25 kSa / s using Akima spline interpolation, with a root mean square error of 0.002 degrees Celsius for the reconstruction of the original sampling points. The time-aligned 24-channel time series was reconstructed using a 1024-point sliding window, extracting 276 cross-correlation delay features between channels. The joint state matrix had dimensions of 1024 rows by 300 columns. After max-min normalization and tensor quantization encapsulation, a four-dimensional multidimensional electrical measurement data stream was output, with a measured end-to-end processing delay of 19.8 ms.

[0094] In the dynamic polarization internal resistance sequence and deep heat source tensor stripping stage, the battery physical information network adopts a third-order equivalent circuit structure for the NMC811 system, including ohmic internal resistance, high-frequency SEI film polarization pairs, mid-frequency charge transfer polarization pairs, and low-frequency Warburg diffusion elements, with an output parameter dimension of 7. The heat transfer constraint discretizes the battery pack interior into 216 volumetric units, and the deep heat source tensor dimension is 216. The network was optimized to an optimal validation loss of 0.0071 on a training set of 18,500 time windows. Under normal fast charging conditions with a state of charge of 90% and a temperature of 25 degrees Celsius, the total value of the dynamic polarization internal resistance sequence stabilizes between 2.75 mΩ and 3.21 mΩ. Five seconds after external extrusion induces an internal short circuit, this value surges to 5.87 mΩ to 8.34 mΩ, an increase of 113.4% to 159.8%, with a deviation from the reference impedance spectrum of only 3.7%. The deep heat source tensor extreme value unit achieves an accuracy of 93.2% in identifying internal short-circuit hotspot locations.

[0095] In the construction and quantification stages of the electrothermal cross-domain residual model, the electrothermal cross-domain residual model, based on Gaussian process regression, was fitted on 10,500 time window samples under normal operating conditions. The average prediction error of the validation set was 2.1 kW per cubic meter, with a relative error of 4.3%. The high-dimensional spatial projection selected the first 10 principal components (explaining a cumulative variance of 91.4%) to compress the 216-dimensional difference vector into a scalar charge-heating residual index. Under normal fast charging conditions, the mean value of this index was 0.078. The mean value rose to 0.57 in the pre-short circuit stage induced by external compression. The index exceeded the warning threshold of 0.35 10.2 seconds before the internal short circuit was triggered. The false alarm rate for 347 cycles under normal operating conditions was 0.19%.

[0096] In the residual gradient flow extraction and online correction stage of the electrical feature benchmark, the weight of the charge heating residual penalty term in the target optimization equation is set to 0.35. The automatic differentiation engine fully preserves the gradient of the penalty term and backpropagates it to ensure that the underlying feature extraction parameters are effectively driven by the electrothermal conservation constraint. The L2 norm of the residual gradient flow sequence monotonically decreases from 0.0447 at the beginning of training to 0.0068 at convergence, a decrease of 84.8%. The F1 score of the validation set before and after the phase transition point jumps from 0.853 to 0.916. The parameter update controller performs online correction of the electrical feature benchmark based on the low-frequency components of the residual gradient flow, and compensates in real time for the drift of the dynamic polarization internal resistance benchmark after 347 cycles of aging (cumulative 0.76mΩ, drift amplitude 25.5%). The false alarm rate of thermal runaway precursors decreases from 9.1% in the uncorrected state to 1.4%, a decrease of 84.6%.

[0097] During the identification of step electrical parameter shock waves and the output of critical blocking signals for thermal runaway, the Bayesian change point detection module has a monitoring window of 25 seconds. The triggering condition is that the posterior probability of the mutation exceeds 0.91 and is continuously satisfied for 400ms. In 347 normal operating condition tests, the number of false triggers was zero. In the external extrusion-induced internal short circuit experiment, the dynamic polarization internal resistance sequence exhibits a step electrical parameter shock wave with an amplitude jump of 1.83mΩ 7.1 seconds before the internal short circuit is triggered (1.73 times the judgment threshold of 1.06mΩ). The field-programmable gate array chip hardware logic AND gate completes the joint decision within 1 microsecond. The insulated gate bipolar transistor module has a turn-off time of 7.6 microseconds. The total response delay from abnormal detection to physical disconnection of the charging circuit does not exceed 30ms. The thermal runaway critical blocking signal is sent to the vehicle controller within 43 microseconds via the controller area network bus (1 Mbit / s, priority identifier 0x000). After triggering the secondary response, the highest internal temperature of the battery pack drops from 71.8 degrees Celsius to 54.3 degrees Celsius, and the risk of thermal runaway is completely reduced to within the safety threshold within 215 seconds.

[0098] This embodiment achieves reliable detection of precursory thermal runaway states. The residual gradient flow-driven online reference correction mechanism reduces the false alarm rate after cyclic aging and improves the problem of frequent false alarms caused by reference drift due to battery aging. The combination of equivalent circuit model and heat transfer discretization improves the accuracy of identifying internal short-circuit hotspot locations in the deep heat source tensor, providing a reliable basis for accurate positioning of thermal runaway risks and effectively improving the high-precision, low-latency, and high-reliability thermal runaway state identification capability in new energy vehicles.

[0099] It should be clarified that the embodiments described above are merely exemplary and are intended to aid in understanding the present invention, not to limit it. Those skilled in the art can make various changes and modifications after grasping the core ideas of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the thermal runaway state of a lithium battery based on a physical information neural network, characterized in that, include: The system acquires the spontaneously superimposed wideband fluctuating current in the charging circuit of new energy vehicle batteries, and simultaneously collects the voltage response trajectory and surface temperature signal at the battery terminal. The broadband fluctuating current and voltage response trajectories are reconstructed by time-domain interleaving with the surface temperature signal to generate a multi-dimensional electrical measurement data stream. The multidimensional electrical measurement data stream is injected into the battery physical information network, and the dynamic polarization internal resistance sequence and deep heat source tensor characterizing the current battery polarization characteristics are extracted; an electrothermal cross-domain residual model is constructed, and boundary tests are performed based on the dynamic polarization internal resistance sequence and deep heat source tensor to quantify the charge heating residual reflecting the thermoelectric coupling imbalance state. The charge heating residual is forced to collapse to zero as a convergence boundary, and the residual gradient flow is extracted. Based on the residual gradient flow, the electrical characteristic benchmark of the battery physical information network is continuously corrected, and the dynamic polarization internal resistance sequence after being constrained by the deep heat source tensor is evaluated online. When a step electrical parameter shock wave is identified in the electrical characteristic reference, the battery charging circuit is cut off and a thermal runaway critical blocking signal is output, thus converging the risk of thermal runaway of the battery within a safe threshold.

2. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 1, characterized in that, The specific implementation process of acquiring the spontaneously superimposed broadband fluctuating current in the charging circuit of a new energy vehicle battery, and simultaneously collecting the voltage response trajectory and surface temperature signal at the battery terminal includes: A wideband current transformer is used to continuously monitor the battery charging circuit and capture the broadband fluctuating current that spontaneously superimposes during the charge transfer process. A voltage sensing channel sharing a global clock with the broadband fluctuating current is established to continuously sample the potential difference between battery terminals and generate a voltage response trajectory. Based on the temperature measurement node attached to the battery casing, the surface temperature signal diffused from the battery's internal heat conduction to the external surface is captured. The broadband fluctuating current, voltage response trajectory, and surface temperature signal are synchronously imported into a buffer queue for baseline calibration and anti-aliasing filtering.

3. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 2, characterized in that, The specific implementation process of reconstructing the broadband fluctuating current and voltage response trajectories with the surface temperature signal in the time domain to generate a multidimensional electrical measurement data stream includes: The timestamps of the baseline-calibrated broadband fluctuating current, voltage response trajectories, and surface temperature signals are extracted. Based on the global clock, the broadband fluctuating current, voltage response trajectories, and surface temperature signals are uniformly interpolated to the same time resolution. Following a feature fusion mechanism, the time-aligned broadband fluctuating current, voltage response trajectories, and surface temperature signals are concatenated into a multi-channel time series. A time-domain interleaving reconstruction operation is performed on the multi-channel time series to establish a joint state matrix representing the time-delay correlation between variables. The features of each dimension of the joint state matrix are normalized and tensorized for encapsulation to generate a multi-dimensional electrical measurement data stream.

4. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 1, characterized in that, The specific implementation process of injecting the multidimensional electrical measurement data stream into the battery physical information network and extracting the dynamic polarization internal resistance sequence and deep heat source tensor characterizing the current battery polarization features includes: A battery physical information network is established that integrates prior knowledge of electrothermal coupling physics with a deep feature extraction architecture. The multidimensional electrical measurement data stream is forward-injected into the hidden layer of the battery physical information network via the input layer. The battery physical information network is driven to perform online calculations using the internally embedded equivalent circuit physical laws to decouple the dynamic polarization internal resistance sequence containing ohmic impedance and polarization capacitance from the multidimensional electrical measurement data stream. The implicit solution operator of the heat transfer equations is triggered synchronously to map the residual time series features into a deep heat source tensor reflecting the internal heat generation rate distribution.

5. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 1, characterized in that, The electrothermal cross-domain residual model is constructed, and boundary tests are performed based on the dynamic polarization internal resistance sequence and the deep heat source tensor. The specific implementation process of quantifying the charge-heating residual reflecting the thermoelectric coupling imbalance state includes: Using the dynamic polarization internal resistance sequence and the deep heat source tensor as mutually verifying independent and dependent variables, an electrothermal cross-domain residual model characterizing the mismatch between Joule heat dissipation and actual temperature rise is constructed. Thermoelectric conversion boundary conditions characterizing safe operating conditions are set, and the dynamic polarization internal resistance sequence and the deep heat source tensor are input into these boundary conditions for boundary testing. The difference between the excessive ohmic heat generation caused by abnormal polarization and the theoretical heat transfer distribution in the deep heat source tensor is calculated. This difference is then transformed into an error index through high-dimensional spatial projection, quantifying the charge-heating residual reflecting the internal critical state.

6. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 1, characterized in that, The specific implementation process of forcing the charge-generating residual to collapse to zero as a convergence boundary and extracting the residual gradient flow includes: The charge heating residual is introduced into the loss function of the battery physical information network, and a target optimization equation with the charge heating residual as the penalty term is constructed. The automatic differentiation engine is started to perform backpropagation calculation on the target optimization equation, and the charge heating residual is forced to collapse to zero as the convergence boundary. During the iterative optimization process of network weight update, the direction vector representing the state evolution trend on the loss plane is continuously collected. The residual gradient flow containing nonlinear mapping relationship is extracted along the trajectory direction of the fastest descent of the penalty term.

7. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 1, characterized in that, The specific implementation process for online evaluation of the dynamic polarization internal resistance sequence after deep heat source tensor constraint, based on the electrical characteristic benchmark of the residual gradient flow continuously correcting battery physical information network, includes: The extracted residual gradient flow is fed back to the parameter update controller of the battery physical information network; the error compensation direction indicated by the residual gradient flow is used to continuously correct the electrical characteristic reference drift caused by the increase in impedance due to battery aging; the dynamic polarization internal resistance sequence after thermodynamic cross-validation constraint by the deep heat source tensor is obtained, and the degree of deviation of the dynamic polarization internal resistance sequence from the normal polarization trajectory is evaluated online in combination with the dynamically updated electrical characteristic reference; an evaluation index sequence reflecting the dynamic evolution of internal interface impedance and thermal runaway evolution characteristics is output.

8. The method for identifying the thermal runaway state of a lithium battery based on a physical information neural network according to claim 7, characterized in that, When a step electrical parameter shock wave is identified in the electrical characteristic reference, the specific implementation process of cutting off the battery charging circuit and outputting a thermal runaway critical blocking signal to converge the battery's thermal runaway risk within a safety threshold includes: The mutation detection logic runs continuously within the monitoring window of the electrical characteristic benchmark. When a nonlinear mutation occurs in the evaluation index sequence input to the decision node, and a step electrical parameter shock wave is identified in the electrical characteristic benchmark, a hardware protection interruption command is triggered. According to the hardware protection interruption command, the battery charging circuit is cut off by a physical switching element, and the injection of external power is blocked. Simultaneously, a thermal runaway critical blocking signal is output through the communication interface. Based on the thermal runaway critical blocking signal, the charging and discharging process of the battery is intervened, and the thermal runaway risk of the battery is converged within a safe threshold.

Images