Probabilistic thermal runaway early warning method and system based on multi-modal three-dimensional thermal field

Abstract

The invention discloses a probabilistic thermal runaway early warning method and system based on a multi-modal three-dimensional thermal field, and relates to the technical field of energy storage safety monitoring, and the early warning method comprises the specific steps: S100, multi-modal data collection; s200, carrying out time alignment processing; s300, feature fusion extraction is carried out; s400, reconstructing a three-dimensional temperature field; s500, calculating a thermal runaway probability; according to the method, multi-modal data of RGB-D, long-wave infrared and gas sensors are fused, space-time synchronization is realized in combination with a soft Kalman alignment algorithm, the limitation of single physical quantity monitoring is broken through, cross-modal correlation features are extracted by adopting a pyramid cross attention adversarial network, and the multi-modal data of the RGB-D, the long-wave infrared and the gas sensors are subjected to hierarchical linkage response. The generated multi-scale feature tensor can accurately describe the spatio-temporal evolution law of a thermal runaway precursor, after lightweight neural radiation field voxel reconstruction, millimeter-level dynamic updating of a three-dimensional temperature field is realized, the generation and diffusion process of local hot spots in an energy storage cabin can be captured in real time, and a high-fidelity thermal field portrait is provided for early risk identification.

Description

Technical Field

[0001] This invention relates to the field of energy storage safety monitoring technology, specifically to a probabilistic thermal runaway early warning method and system based on multimodal three-dimensional thermal field. Background Technology

[0002] As the global energy structure transitions towards a low-carbon model, lithium-ion battery energy storage systems, with their advantages of high energy density and long cycle life, have become a core supporting technology for grid peak shaving, renewable energy grid integration, and electric vehicles. However, lithium-ion batteries can cause serious safety accidents such as combustion and explosion under thermal runaway conditions. The thermal runaway process is accompanied by multi-physical field coupling phenomena such as rapid temperature rise, gas evolution, and structural deformation. Traditional monitoring methods mainly rely on single-point temperature sensors or gas detectors, which suffer from insufficient spatial coverage and dynamic response lag, making it difficult to accurately capture the generation and diffusion patterns of local hot spots in the early stages of thermal runaway. In addition, existing early warning methods mostly use threshold judgments and lack quantitative analysis of the thermo-mechanical-chemical multi-field coupling mechanism, resulting in high false alarm and false negative rates, which cannot meet the safety requirements of large-scale application of energy storage systems.

[0003] Existing lithium battery thermal runaway monitoring technologies have certain shortcomings: On the one hand, single-modal sensors (such as temperature probes and gas sensors) can only acquire local or single physical quantity information, failing to comprehensively reflect the spatiotemporal dynamic evolution of the temperature field and gas concentration field during thermal runaway, thus limiting early risk identification capabilities; on the other hand, traditional data fusion methods often employ simple splicing or weighted averaging, failing to fully exploit the spatiotemporal correlation characteristics between multimodal data, making it difficult to construct accurate characterization models of thermal runaway precursors; simultaneously, early warning mechanisms are based on fixed thresholds, failing to consider the impact of changes in operating conditions such as ambient temperature and charging / discharging states on the thermal runaway threshold, and the response strategy is "open-loop control" (such as directly cutting off the circuit), lacking gradient and adaptive closed-loop intervention capabilities from early warning to disposal. These deficiencies result in a false alarm rate exceeding 15% in practical applications of traditional technologies, and insufficient timeliness in suppressing early thermal runaway, making it difficult to meet the needs of high-safety-level energy storage scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multimodal three-dimensional thermal field probabilistic thermal runaway early warning method and system. By fusing RGB-D images, long-wave infrared images, and gas sensing data, it achieves multi-physics field collaborative perception of thermal runaway precursors. The system employs dynamic thermal field reconstruction and probabilistic risk assessment technology to quantify the probability of thermal runaway occurrence and triggers a gradient-based safety response strategy based on graded thresholds, forming a closed-loop control chain of "perception-assessment-response". This invention overcomes the limitations of traditional single-modal monitoring, improves the accuracy and timeliness of early thermal runaway identification and intervention, and provides a highly reliable, low-false-alarm intelligent solution for lithium battery energy storage safety.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a multimodal three-dimensional thermal field probabilistic thermal runaway early warning method is provided, the specific steps of which are as follows: S100, multimodal data acquisition: acquires RGB-D image sequences, long-wave infrared image sequences, and gas concentration time-series signals in the lithium battery energy storage compartment through RGB-D sensors, long-wave infrared sensors, and gas sensors; S200, Time Alignment Processing: The multimodal data collected in step A is time-aligned using the soft Kalman alignment algorithm to obtain a time-synchronized dataset; S300, Feature Fusion Extraction: Input the time-synchronized dataset obtained in step B into the pyramid cross-attention adversarial network to generate a multi-scale feature tensor; S400, Three-dimensional temperature field reconstruction: The multi-scale feature tensor output from step C is processed using a lightweight neural radiation field voxel reconstruction algorithm to generate the three-dimensional temperature field of the energy storage compartment. S500, thermal runaway probability calculation: A variational Bayesian risk assessment framework is established based on the Arrhenius heat generation model. The thermal potential threshold is derived through the coupling formula of heat conduction and heat generation, and the probability of thermal runaway is calculated by combining the probability inference formula. S600, graded linkage response: The first, second and third thresholds for the probability of thermal runaway are set to 30%, 60% and 80% respectively. When the probability of thermal runaway reaches the corresponding threshold, the power reduction action of the energy storage management system, the emergency mode of the liquid cooling system and the pre-wetting action of the fire protection system are triggered in sequence.

[0006] Furthermore, in S100, the acquisition frame rate of the RGB-D image sequence in the multimodal data acquisition step is 30fps, the acquisition frame rate of the long-wave infrared image sequence is 60fps, the acquisition frequency of the gas concentration time series signal is 10Hz, and the gas concentration time series signal includes H2 and VOC gas concentration signals.

[0007] Furthermore, in step S200, the soft Kalman alignment algorithm formula in the time alignment processing step is: ,in, for State estimate at time 1; This is the state prediction value based on time k-1; Kalman gain; for Multimodal observation data at various times; The observation matrix; The prediction error covariance matrix; To observe the noise covariance matrix.

[0008] Furthermore, in S300, the pyramid cross-attention adversarial network in the feature fusion extraction step can be replaced with a Swin-Transformer fusion network, and the replaced network is optimized using XLA or Fused-QKVINT4 ultra-low bit quantization technology. After optimization, it can still process time-synchronized datasets and generate multi-scale feature tensors.

[0009] Furthermore, in S400, the formula for the lightweight neural radiation field voxel reconstruction algorithm in the three-dimensional temperature field reconstruction step is: ,in, In three-dimensional space coordinates, This represents the temperature value at that coordinate. For the number of voxels, For the first The weighting coefficients of individual elements, For the first density parameters of voxels For the first Characteristic parameters of individual elements This is a function for a multilayer perceptron.

[0010] Furthermore, in the S400 three-dimensional temperature field reconstruction step, the single-frame inference delay of the lightweight neural radiation field voxel reconstruction algorithm does not exceed 50ms, the default voxel granularity is 5cm, and the voxel granularity can be adaptively set in the range of 2cm to 20cm. The network inference supports INT8 or INT4 low-bit quantization.

[0011] Furthermore, in step S500, the formula for deriving the thermal potential threshold coupling between heat conduction and heat generation in the thermal runaway probability calculation step is as follows: ,in, The density of the dielectric material inside the lithium battery energy storage compartment. For the specific heat capacity of the medium, The rate of change of temperature over time. For the thermal conductivity of the medium, For the Laplace operator of temperature, Pre-exponential factor, For activation energy, Let be the ideal gas constant. This refers to the absolute temperature inside the energy storage compartment.

[0012] Furthermore, in step S500, the probability inference formula in the thermal runaway probability calculation step is as follows: ,in, The thermal potential threshold function, for Three-dimensional temperature field data at time t. for Gas concentration data at any given time. For the posterior probability distribution, For model parameters, This is a multimodal observation dataset.

[0013] Furthermore, in the S600 graded linkage response step, when the probability of thermal runaway reaches the third threshold, the pre-wetting action of the fire protection system can be replaced by the rapid disconnection action of the DC circuit breaker; and when the continuous missing time of any mode sensor data in step A is less than 30s, the missing data is automatically recovered through interpolation and nearest neighbor resampling algorithm to ensure that the data availability is not less than 99.5%.

[0014] On the other hand, based on a multimodal three-dimensional thermal field probabilistic thermal runaway early warning system, the system includes: Sensing subsystem: includes RGB-D sensor, long-wave infrared sensor and gas sensor for detecting H2 and VOC, used to collect multimodal data in lithium battery energy storage compartment; Time alignment module: Configured to execute the soft Kalman alignment algorithm to perform time alignment on the multimodal data collected by the sensing subsystem and output a time-synchronized dataset; Feature fusion module: configured to implement a pyramid cross-attention adversarial network or a Swin-Transformer fusion network to process time-synchronized datasets and generate multi-scale feature tensors; 3D Reconstruction Module: Configured to process multi-scale feature tensors to generate a 3D temperature field for the energy storage compartment based on a lightweight neural radiation field voxel reconstruction algorithm; Probability assessment module: configured to derive the thermal potential threshold and calculate the output probability of thermal runaway based on the Arrhenius heat generation model and variational Bayesian method; Linked execution module: used to trigger corresponding graded linked actions sequentially based on the threshold reached by the probability of thermal runaway; Edge computing unit: It adopts a fanless liquid cooling structure and can operate for a long time in an environment of -45℃ to 55℃, providing computing support for the above modules.

[0015] Compared with existing technologies, this probabilistic thermal runaway early warning method and system based on multimodal three-dimensional thermal field has the following advantages: I. This invention overcomes the limitations of monitoring a single physical quantity by fusing multimodal data from RGB-D, long-wave infrared, and gas sensors, combined with a soft Kalman alignment algorithm to achieve spatiotemporal synchronization. It employs a pyramidal cross-attention adversarial network to extract cross-modal correlation features, and the generated multi-scale feature tensor can accurately characterize the spatiotemporal evolution of precursors to thermal runaway. After being reconstructed by lightweight neural radiation field voxels, the three-dimensional temperature field achieves millimeter-level dynamic updates, enabling real-time capture of the generation and diffusion process of local hot spots within the energy storage chamber, providing a high-fidelity thermal field profile for early risk identification.

[0016] Second, this invention derives the dynamic thermal potential threshold through the coupling relationship between heat conduction and heat generation, quantifies the probability of thermal runaway by combining gas concentration and temperature field data, and automatically triggers gradient intervention strategies such as power reduction, liquid cooling enhancement, and fire prevention based on three probability thresholds of 30%, 60%, and 80% to form a closed-loop control of the entire chain of "sensing-assessment-disposal". The system has a built-in automatic data loss recovery function and, together with a wide temperature range edge computing unit, improves the early warning robustness and response time under complex operating conditions.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a schematic diagram of the overall process of the thermal runaway early warning method; Figure 2 This is a schematic diagram of probability assessment and linkage response logic. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Thermal runaway early warning scenario for large-scale industrial and commercial lithium battery energy storage power stations I. Multimodal Data Acquisition Inside the energy storage compartment of a large-scale industrial and commercial lithium battery energy storage power station, RGB-D sensors, long-wave infrared sensors, and gas sensors are installed according to a pre-determined deployment plan. The RGB-D sensors continuously capture RGB-D image sequences of the lithium battery packs and their surrounding environment within the energy storage compartment, clearly recording visual information such as the appearance and location of the equipment. The long-wave infrared sensors focus on collecting long-wave infrared image sequences within the energy storage compartment, accurately capturing the thermal radiation in different areas and reflecting temperature differences at different locations. The gas sensors monitor the concentration changes of H2 and VOC gases within the energy storage compartment in real time, generating gas concentration time-series signals. This comprehensive acquisition of multi-dimensional data within the energy storage compartment provides fundamental data support for subsequent early warning analysis, such as... Figure 1 As shown.

[0022] II. Time Alignment Processing Due to differences in the acquisition mechanisms and response speeds of different sensors, the acquired multimodal data may be asynchronous in the time dimension, which can affect the accuracy of subsequent data processing and analysis. Therefore, the time alignment module is activated, and the soft Kalman alignment algorithm is invoked to perform time alignment processing on the RGB-D image sequence, long-wave infrared image sequence, and gas concentration time series signal acquired in the first step. The formula for the soft Kalman alignment algorithm is as follows: ,in, for State estimate at time 1; This is the state prediction value based on time k-1; Kalman gain; for Multimodal observation data at various times; The observation matrix; The prediction error covariance matrix; To observe the noise covariance matrix, this algorithm estimates and predicts the state of each modality's data, continuously corrects errors by combining observation data, and ultimately adjusts multimodal data collected at different times to a unified time reference, forming a time-synchronized dataset. This ensures that subsequent data processing is based on a time-consistent data foundation.

[0023] III. Feature Fusion Extraction The time-synchronized dataset, after time alignment processing, is input into the feature fusion module. This module, by default, uses a pyramid cross-attention adversarial network to perform feature fusion and extraction. The pyramid cross-attention adversarial network can mine and analyze data features at different scales, achieving effective interaction and correlation between features of different modalities through a cross-attention mechanism. Simultaneously, adversarial training enhances the discriminative power and robustness of the features, ultimately generating a multi-scale feature tensor that comprehensively reflects the state inside the energy storage chamber, providing high-quality feature input for three-dimensional temperature field reconstruction.

[0024] IV. Three-dimensional temperature field reconstruction After receiving the multi-scale feature tensor, the 3D reconstruction module uses the lightweight neural radiation field voxel reconstruction algorithm to process it. The formula for the lightweight neural radiation field voxel reconstruction algorithm is as follows: ,in, In three-dimensional space coordinates, This represents the temperature value at that coordinate. For the number of voxels, For the first The weighting coefficients of individual elements, For the first density parameters of voxels For the first Characteristic parameters of individual elements This is a multilayer perceptron function. Based on the principle of neural radiation fields and combined with voxelization, the algorithm efficiently transforms multi-scale feature tensors into temperature distribution information in three-dimensional space. By dividing the space inside the energy storage compartment into voxels and calculating the temperature value corresponding to each voxel, a complete three-dimensional temperature field of the energy storage compartment is gradually constructed. This three-dimensional temperature field can intuitively and accurately present the temperature conditions at various locations within the energy storage compartment, providing an intuitive spatial temperature distribution basis for judging whether there is a risk of thermal runaway.

[0025] V. Calculation of Thermal Runaway Probability The probability assessment module constructs a variational Bayesian risk assessment framework based on the Arrhenius heat generation model, and uses this framework to calculate the probability of thermal runaway. First, it utilizes the coupling formula between heat conduction and heat generation, which is: ,in, The density of the dielectric material inside the lithium battery energy storage compartment. For the specific heat capacity of the medium, The rate of change of temperature over time. For the thermal conductivity of the medium, For the Laplace operator of temperature, Pre-exponential factor, For activation energy, Let be the ideal gas constant. Given the absolute temperature inside the energy storage chamber, and combining temperature data from the three-dimensional temperature field with the relevant characteristics of the medium within the chamber, a thermal potential threshold is derived. This threshold serves as a crucial indicator for assessing the risk of thermal runaway, reflecting the critical condition for the system to transition from normal heat dissipation to thermal runaway. Next, the probabilistic inference formula is applied, substituting the three-dimensional temperature field data (i.e., the three-dimensional temperature field data at time t) and the gas concentration data (i.e., the gas concentration data at time t). The probabilistic inference formula is as follows: ,in, The thermal potential threshold function, for Three-dimensional temperature field data at time t. for Gas concentration data at any given time. For the posterior probability distribution, For model parameters, The system uses a multimodal observation dataset and constructs a posterior probability distribution based on the dataset. It then calculates the probability of thermal runaway at the current moment through integral operations, thereby quantifying the likelihood of thermal runaway occurring within the energy storage compartment.

[0026] VI. Tiered and coordinated response Based on preset thermal runaway probability thresholds—a first threshold of 30%, a second threshold of 60%, and a third threshold of 80%—the calculated thermal runaway probability is monitored and assessed in real time. When the thermal runaway probability reaches 30% (the first threshold), the linkage execution module sends a command to the energy storage management system, triggering a power reduction action to decrease the operating power of the energy storage system, reduce system heat generation, and delay the development of thermal runaway risk. When the thermal runaway probability further increases to 60% (the second threshold), the liquid cooling system's emergency mode is triggered, operating at maximum heat dissipation capacity to enhance heat dissipation within the energy storage compartment and strive to control the temperature rise. If the thermal runaway probability continues to rise to 80% (the third threshold), the fire suppression system's pre-wetting action is immediately triggered to pre-wet relevant areas within the energy storage compartment, preparing for potential fires and minimizing losses from severe accidents caused by thermal runaway.

[0027] VII. Handling of Special Circumstances Throughout the early warning process, the system monitors the operational status and data transmission of each sensor in real time. If any modal sensor data is continuously missing for less than 30 seconds, the system automatically initiates interpolation and nearest neighbor resampling algorithms to recover the missing data. By analyzing the patterns of effective data before and after the missing data, appropriate interpolation methods are used to supplement the missing data points. Simultaneously, the nearest neighbor resampling method is combined to further optimize the data recovery effect, ensuring data availability is no less than 99.5%. This guarantees that the early warning system can continue to operate normally even with brief data loss, avoiding early warning interruptions or misjudgments due to data loss.

[0028] In summary, in the scenario of thermal runaway early warning for large-scale industrial and commercial lithium battery energy storage power stations, this invention comprehensively collects data within the energy storage compartment using multimodal sensors, achieves time synchronization through a soft Kalman alignment algorithm, extracts and fuses features using a pyramidal cross-attention adversarial network, and constructs a three-dimensional temperature field using a lightweight neural radiation field voxel reconstruction algorithm. Based on the Arrhenius heat generation model and the variational Bayesian risk assessment framework, the probability of thermal runaway is calculated using the heat conduction and heat generation coupling formula and the probability inference formula. Corresponding linkage responses are triggered based on a three-level threshold. Simultaneously, interpolation and nearest neighbor resampling algorithms are used to address temporary data gaps, ensuring the accuracy and continuity of thermal runaway early warning for large-scale energy storage power stations and effectively reducing accident risks.

[0029] Example 2: Early warning scenario of thermal runaway in the power battery storage compartment of a new energy vehicle I. Multimodal Data Acquisition Inside the power battery storage compartment of new energy vehicles, RGB-D sensors, long-wave infrared sensors, and gas sensors are strategically placed. The RGB-D sensors acquire real-time RGB-D image sequences of the power battery pack within the storage compartment, clearly displaying information such as the battery pack's appearance, structure, and connection status, and promptly detecting any abnormalities such as bulging or damage to the battery's appearance. The long-wave infrared sensors continuously acquire long-wave infrared image sequences within the storage compartment, accurately capturing the thermal distribution of the battery pack and surrounding components, and quickly identifying localized high-temperature areas. The gas sensors detect real-time changes in the concentration of H2 and VOC gases within the storage compartment, generating gas concentration time-series signals to promptly detect potential leaks or other issues within the battery, comprehensively collecting multimodal data from within the power battery storage compartment.

[0030] II. Time Alignment Processing Considering the potential vibrations and bumps that may occur during the operation of new energy vehicles, which could affect the timing of sensor acquisition and lead to asynchronous data from different modes, a time alignment module is activated. This module uses a soft Kalman alignment algorithm to time-align the acquired RGB-D image sequences, long-wave infrared image sequences, and gas concentration time-series signals. The soft Kalman alignment algorithm formula is as follows: This algorithm can accurately correct the timestamps of each modality data based on the dynamic characteristics of the sensor and the changing patterns of the collected data, eliminate time deviations between different sensors, and integrate multimodal data into a time-synchronized dataset, providing time-unified data assurance for subsequent feature fusion and analysis.

[0031] III. Feature Fusion Extraction Given the high real-time performance and computational efficiency requirements of early warning systems in new energy vehicles, the feature fusion module can employ an optimized Swin-Transformer fusion network, optimized using XLA or Fused-QKVINT4 ultra-low bit quantization techniques. The optimized Swin-Transformer fusion network effectively reduces computational complexity and resource consumption while maintaining feature extraction accuracy, enabling rapid feature fusion and extraction from time-synchronized datasets. This network deeply fuses features from different modalities, fully exploring the correlations between them to generate multi-scale feature tensors that accurately reflect the state of the power battery storage compartment, meeting the rapid response requirements of early warning systems during dynamic driving of new energy vehicles.

[0032] IV. Three-dimensional temperature field reconstruction After receiving the multi-scale feature tensor, the 3D reconstruction module runs the lightweight neural radiation field voxel reconstruction algorithm to reconstruct the 3D temperature field. The formula for the lightweight neural radiation field voxel reconstruction algorithm is as follows: This algorithm, tailored to the spatial structure characteristics of power battery storage compartments in new energy vehicles, rationally sets the voxel granularity (adaptively adjustable within the range of 2cm to 20cm to accommodate different battery layouts), efficiently converting multi-scale feature tensors into three-dimensional temperature field data. The reconstructed three-dimensional temperature field clearly displays the temperature distribution of each battery cell, battery module, and surrounding area within the power battery storage compartment, accurately capturing the location and extent of abnormal battery heating, and providing detailed spatial temperature distribution data for assessing thermal runaway risks.

[0033] V. Calculation of Thermal Runaway Probability The probability assessment module continues to use the Arrhenius heat generation model to construct a variational Bayesian risk assessment framework for calculating the probability of thermal runaway. First, considering the characteristics of the medium within the energy storage compartment of a new energy vehicle's power battery, the coupling formula between heat conduction and heat generation is applied. The coupling formula between heat conduction and heat generation is as follows: Based on the temperature data in the three-dimensional temperature field and the heat generation of the battery, a thermal potential threshold suitable for this scenario is derived, clearly defining the critical standard for determining whether the power battery faces the risk of thermal runaway. Subsequently, a probabilistic inference formula is applied, which is: Substituting the three-dimensional temperature field data and gas concentration data at time t, and constructing a posterior probability distribution based on the multimodal observation dataset, the probability of thermal runaway within the current power battery energy storage compartment is accurately calculated through integral operations. This provides a quantitative risk basis for subsequent graded linkage response, such as... Figure 2 As shown.

[0034] VI. Tiered and coordinated response Based on three preset thresholds for thermal runaway probability, the calculated thermal runaway probability is monitored and its level is judged in real time. When the thermal runaway probability reaches 30% (first threshold), the linkage execution module sends a command to the battery management system of the new energy vehicle, triggering the energy storage management system to reduce power output, reduce the heat generated by the battery, and slow down the development of thermal runaway risk. When the thermal runaway probability rises to 60% (second threshold), the liquid cooling system emergency mode is immediately activated, and the liquid cooling system operates at maximum efficiency to enhance heat dissipation from the power battery and control the rise in battery temperature as much as possible. If the thermal runaway probability exceeds 80% (third threshold), considering the safety and emergency handling needs during the operation of the new energy vehicle, the pre-wetting action of the fire protection system can be replaced by the rapid disconnection action of the DC circuit breaker to quickly cut off the power supply circuit of the power battery, prevent the fire from spreading and the accident from escalating, and ensure the safety of the vehicle and personnel.

[0035] VII. Handling of Special Circumstances During the operation of new energy vehicles, sensor data from any mode may be temporarily lost due to sensor malfunctions, signal interference, or other reasons. When the system detects that the continuous loss of any mode's sensor data is less than 30 seconds, it automatically initiates interpolation and nearest neighbor resampling algorithms to recover the missing data. By analyzing the valid data before and after the missing data, interpolation methods are used to fill in the missing data points, and the data quality is further optimized using nearest neighbor resampling technology to ensure that data availability is no less than 99.5%. This avoids the impact of temporary data loss on the normal operation of the early warning system, ensuring the continuity and accuracy of thermal runaway early warning, and providing reliable protection for the safe operation of new energy vehicles. Simultaneously, the edge computing unit in the system adopts a fanless liquid-cooled heat dissipation structure, enabling long-term stable operation within the harsh environmental temperature range of -45℃ to 55℃ that new energy vehicles may encounter during operation. This provides continuous and stable computing support for the normal operation of the aforementioned modules, ensuring that the entire early warning system can reliably perform its early warning function in various complex environments.

[0036] In summary, in the scenario of thermal runaway early warning for the power battery storage compartment of new energy vehicles, this invention first deploys multimodal sensors to collect data, and then processes the data using a soft Kalman alignment algorithm to achieve time synchronization. An optimized Swin-Transformer fusion network is used to extract features, and a three-dimensional temperature field is generated using a lightweight neural radiation field voxel reconstruction algorithm. The thermal runaway probability is then calculated based on the Arrhenius heat generation model and variational Bayesian methods, triggering a coordinated response according to a three-level threshold, and supporting the replacement of response actions in emergencies. Edge computing units ensure operation in harsh environments, and interpolation and nearest neighbor resampling algorithms ensure data availability, providing a reliable thermal runaway early warning guarantee for the safety of new energy vehicle power batteries.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A probabilistic thermal runaway early warning method based on multimodal three-dimensional thermal field, characterized in that, The specific steps of this early warning method are as follows: S100, multimodal data acquisition: acquires RGB-D image sequences, long-wave infrared image sequences, and gas concentration time-series signals in the lithium battery energy storage compartment through RGB-D sensors, long-wave infrared sensors, and gas sensors; S200, Time Alignment Processing: The multimodal data collected in step A is time-aligned using the soft Kalman alignment algorithm to obtain a time-synchronized dataset; S300, Feature Fusion Extraction: Input the time-synchronized dataset obtained in step B into the pyramid cross-attention adversarial network to generate a multi-scale feature tensor; S400, Three-dimensional temperature field reconstruction: The multi-scale feature tensor output from step C is processed using a lightweight neural radiation field voxel reconstruction algorithm to generate the three-dimensional temperature field of the energy storage compartment. S500, thermal runaway probability calculation: A variational Bayesian risk assessment framework is established based on the Arrhenius heat generation model. The thermal potential threshold is derived through the coupling formula of heat conduction and heat generation, and the probability of thermal runaway is calculated by combining the probability inference formula. S600, graded linkage response: The first, second and third thresholds for the probability of thermal runaway are set to 30%, 60% and 80% respectively. When the probability of thermal runaway reaches the corresponding threshold, the power reduction action of the energy storage management system, the emergency mode of the liquid cooling system and the pre-wetting action of the fire protection system are triggered in sequence.

2. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, In step S100, the acquisition frame rate of the RGB-D image sequence in the multimodal data acquisition step is 30fps, the acquisition frame rate of the long-wave infrared image sequence is 60fps, the acquisition frequency of the gas concentration time series signal is 10Hz, and the gas concentration time series signal includes H2 and VOC gas concentration signals.

3. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, In step S200, the soft Kalman alignment algorithm formula in the time alignment process is: ,in, for State estimate at time 1; This is the state prediction value based on time k-1; Kalman gain; for Multimodal observation data at various times; The observation matrix; The prediction error covariance matrix; To observe the noise covariance matrix.

4. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, In S300, the pyramid cross-attention adversarial network in the feature fusion extraction step can be replaced with a Swin-Transformer fusion network. The replaced network is optimized using XLA or Fused-QKVINT4 ultra-low bit quantization technology. After optimization, it can still process time-synchronized datasets and generate multi-scale feature tensors.

5. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, The formula for the lightweight neural radiation field voxel reconstruction algorithm in the S400 three-dimensional temperature field reconstruction step is as follows: ,in, In three-dimensional space coordinates, This represents the temperature value at that coordinate. For the number of voxels, For the first The weighting coefficients of individual elements, For the first density parameters of voxels For the first Characteristic parameters of individual elements This is a function for a multilayer perceptron.

6. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, In the S400 three-dimensional temperature field reconstruction step, the single-frame inference delay of the lightweight neural radiation field voxel reconstruction algorithm does not exceed 50ms, the default voxel granularity is 5cm, and the voxel granularity can be adaptively set in the range of 2cm to 20cm. The network inference supports INT8 or INT4 low-bit quantization.

7. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, In S500, the formula for the coupling of heat conduction and heat generation used to derive the thermal potential threshold in the thermal runaway probability calculation step is as follows: ,in, The density of the dielectric material inside the lithium battery energy storage compartment. For the specific heat capacity of the medium, The rate of change of temperature over time. For the thermal conductivity of the medium, For the Laplace operator of temperature, Pre-exponential factor, For activation energy, Let be the ideal gas constant. This refers to the absolute temperature inside the energy storage compartment.

8. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, The probability inference formula in the thermal runaway probability calculation step of S500 is as follows: ,in, The thermal potential threshold function, for Three-dimensional temperature field data at time t. for Gas concentration data at any given time. For the posterior probability distribution, For model parameters, This is a multimodal observation dataset.

9. The method for early warning of probabilistic thermal runaway based on multimodal three-dimensional thermal field according to claim 1, characterized in that, In the S600 graded linkage response step, when the probability of thermal runaway reaches the third threshold, the pre-wetting action of the fire protection system can be replaced by the rapid disconnection action of the DC circuit breaker; and when the continuous missing time of any mode sensor data in step A is less than 30s, the missing data is automatically recovered through interpolation and nearest neighbor resampling algorithm to ensure that the data availability is not less than 99.5%.

10. A multimodal three-dimensional thermal field probabilistic thermal runaway early warning system, characterized in that, This system is applicable to the multimodal three-dimensional thermal field probabilistic thermal runaway early warning method according to any one of claims 1-9, characterized in that the system comprises: Sensing subsystem: includes RGB-D sensor, long-wave infrared sensor and gas sensor for detecting H2 and VOC, used to collect multimodal data in lithium battery energy storage compartment; Time alignment module: Configured to execute the soft Kalman alignment algorithm to perform time alignment on the multimodal data collected by the sensing subsystem and output a time-synchronized dataset; Feature fusion module: configured to implement a pyramid cross-attention adversarial network or a Swin-Transformer fusion network to process time-synchronized datasets and generate multi-scale feature tensors; 3D Reconstruction Module: Configured to process multi-scale feature tensors to generate a 3D temperature field for the energy storage compartment based on a lightweight neural radiation field voxel reconstruction algorithm; Probability assessment module: configured to derive the thermal potential threshold and calculate the output probability of thermal runaway based on the Arrhenius heat generation model and variational Bayesian method; Linked execution module: used to trigger corresponding graded linked actions sequentially based on the threshold reached by the probability of thermal runaway; Edge computing unit: It adopts a fanless liquid cooling structure and can operate for a long time in an environment of -45℃ to 55℃, providing computing support for the above modules.

Images