A battery pack thermal runaway early warning system and method based on multi-source information fusion

Abstract

The present application relates to the technical field of lithium ion battery safety monitoring, and particularly relates to a battery pack thermal runaway early warning system and method based on multi-source information fusion, which comprises a hardware architecture and a system architecture, the hardware architecture adopts a layered arrangement and comprises sensing layer hardware, computing and control layer hardware and execution and interaction layer hardware, the sensing layer hardware comprises cell and voltage temperature acquisition units, total current acquisition units, battery pack surface infrared thermal imaging units, battery pack surface visual deformation monitoring units, internal gas sensing units and internal pressure sensing units, the computing and control layer hardware is responsible for the operation of the system and transmits to the execution and interaction layer hardware, and the execution and interaction layer hardware comprises an execution interface and a man-machine interface, and solves the core defects of early warning lag, insufficient information fusion, missing fault positioning capability and limited intelligent degree in the present battery pack thermal runaway early warning.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery safety monitoring technology, specifically to a battery pack thermal runaway early warning system and method based on multi-source information fusion. Background Technology

[0002] Battery thermal runaway is a core issue threatening the safety of electric vehicles and energy storage systems. Currently, mainstream thermal runaway monitoring and early warning solutions mainly follow these paths:

[0003] 1. Passive alarm based on fixed thresholds in BMS: Commercial battery management systems generally monitor voltage, temperature, and current. The underlying technology is to set fixed safety thresholds for each parameter (e.g., voltage > 4.25V, temperature > 60℃), triggering an alarm when these limits are exceeded. This method is simple, but it's a reactive alarm and cannot predict early-stage or latent faults.

[0004] 2. Multi-parameter monitoring based on extended sensors: Some solutions attempt to increase monitoring dimensions, such as integrating gas sensors (monitoring CO) or pressure sensors within the battery pack, hoping to provide early warnings by detecting signs of electrolyte decomposition or casing deformation. However, these parameters are usually viewed independently or simply superimposed logically, lacking in-depth analysis of the inherent correlation and evolution trends of multi-source information.

[0005] 3. Model-Driven State Estimation: This method employs equivalent circuit or electrochemical models, combined with algorithms such as Kalman filtering, to estimate the "non-measurable" states of charge and internal resistance of the battery cell online. Health is assessed by tracking anomalies in these internal states. This method is more sensitive to internal cell degradation, but the model accuracy depends on calibration and it does not effectively integrate with the external physical states of the battery pack (such as heat distribution and deformation).

[0006] 4. Data-driven machine learning prediction: This method uses historical operational data to train classification or regression models to predict failures. It relies on large amounts of high-quality failure data, and the models are like "black boxes" with poor interpretability. In engineering applications, it is difficult to pinpoint specific failure sources and establish operational trust.

[0007] The existing technology has the following shortcomings:

[0008] 1. Severely delayed early warning and short safety window: Whether using a single-parameter threshold method or a simple multi-sensor method, alarm triggering relies on changes in physical quantities reaching a "significantly abnormal" stage. However, thermal runaway is a chain process that begins with internal micro-short circuits, accumulates chemical side reactions, and finally erupts violently. Existing technologies are weak in capturing the initial, weak chemical and thermal anomaly signals, resulting in warnings often being issued when the system is already close to the irreversible stage of thermal runaway, leaving extremely little time for emergency response.

[0009] 2. Severe information silos and limited assessment dimensions: Existing solutions often focus on only one type of information, such as either focusing solely on battery management system (BMS) characteristics or external physical signals (gas, pressure). There is a lack of a unified framework capable of deeply integrating and analyzing multi-dimensional and heterogeneous data, including the internal electrochemical state of the battery cell, the overall thermal distribution of the battery pack, mechanical structural deformation, and the internal gas environment. This prevents the system from comprehensively assessing the overall safety status of the battery pack from a "systems theory" perspective, making it prone to false alarms or omissions due to incomplete information.

[0010] 3. "Early warnings without pinpointing the source," resulting in inefficient maintenance and repair: Most early warning systems can only issue system-level alerts stating "the battery pack has a risk of thermal runaway," failing to answer the crucial question of "which cell(s) the risk originates from." In battery packs containing hundreds or thousands of cells, this necessitates time-consuming and costly comprehensive inspections by maintenance personnel, sometimes requiring the complete disassembly of the battery pack. This is not only extremely inefficient but also carries the risk of triggering secondary risks due to improper handling during the inspection process.

[0011] 4. Insufficient intelligence and over-reliance on experience and fixed rules: Traditional threshold methods and simple fusion methods lack self-learning and adaptive capabilities. Their rules and weights are usually set by engineers based on experience, making it difficult to adapt to complex situations such as battery aging, different operating conditions, and differences in the characteristics of different batches of cells. While some data-driven methods have a certain learning ability, they lack interpretable physical models to support them, the decision-making process is not transparent, there is a lack of confidence in engineering applications, and it is difficult to correlate location information with physical mechanisms.

[0012] In summary, existing technologies for battery pack thermal runaway early warning generally suffer from core defects such as delayed warning, insufficient information fusion, lack of fault location capabilities, and limited intelligence. These shortcomings make it difficult to meet the urgent needs of electric vehicles and energy storage systems for high safety and high reliability. Therefore, there is an urgent need to develop a new generation of battery pack thermal runaway early warning technology that can achieve deep fusion of multi-source information, accurate early warning, accurate fault location, and interpretability. Summary of the Invention

[0013] The purpose of this invention is to provide a battery pack thermal runaway early warning system based on multi-source information fusion, including a hardware architecture and a system architecture. The hardware architecture adopts a layered arrangement, including a sensing layer hardware, a computing and control layer hardware, and an execution and interaction layer hardware. The sensing layer hardware includes: a cell and voltage temperature acquisition unit, a total current acquisition unit, a battery pack surface infrared thermal imaging unit, a battery pack surface visual deformation monitoring unit, an internal gas sensing unit, and an internal pressure sensing unit. The computing and control layer hardware is responsible for the system's calculations and transmits them to the execution and interaction layer hardware. The execution and interaction layer hardware includes an execution interface and a human-machine interface.

[0014] The system architecture includes four modules, namely a multi-source heterogeneous data acquisition module: synchronously acquiring cell-level electrical and thermal data, battery pack-level thermal and deformation data, and battery pack internal environment data;

[0015] Spatiotemporal fusion and feature extraction module: "Construct a four-dimensional spatiotemporal fusion tensor" to align and structure the above asynchronous and heterogeneous data in the temporal and spatial dimensions;

[0016] Risk assessment and precise location module: First, an extended Kalman filter combined with an electrochemical-thermal coupling model is used to estimate the state of charge and DC internal resistance of each cell with high accuracy. Second, a partial order lattice of cell health status is constructed based on the state parameters of all cells. The "high-risk candidate cell set" with the worst health status is automatically screened using the dominance relationship. The phase space grid for thermal runaway risk propagation is innovatively constructed to map the cell state to its three-dimensional physical location. The "joint risk density" of each grid cell and the "vector field" characterizing the risk propagation direction are calculated. Manifold analysis is performed on the vector field to identify the "local attractor region" where the risk flow converges. Thus, the location range is narrowed from the candidate set to one or a few cells that form the risk core in space, achieving precise location.

[0017] Dynamic early warning and decision support module: It normalizes and integrates multi-dimensional risk characteristics with adaptive weighting to generate a comprehensive risk coefficient, and divides the warning into three levels: red, yellow and green.

[0018] Further specifying, the cell and voltage / temperature acquisition units employ a high-precision, high-sampling-rate dedicated BMS analog front-end chip with a channel synchronization error of less than 1μs. The positive and negative terminals of each cell are connected to the sampling channel via a Kelvin connection to eliminate the influence of cable resistance. Temperature acquisition uses an NTC thermistor attached to the large surface of the cell and connected to the AFE via a multiplexer to achieve synchronization with voltage sampling. The total current acquisition unit uses a closed-loop current sensor based on the Hall effect principle, with a range covering -500A to +500A, a bandwidth of DC-10kHz, and a linearity better than 0.1%, for accurate measurement of charging and discharging current. The battery pack surface infrared thermal imaging unit... The system includes an uncooled long-wave infrared thermal imager that provides complete coverage of the battery pack's upper surface and main sides; an industrial-grade global shutter CMOS camera positioned at a fixed angle to the battery pack surface, with pre-calibration establishing a mapping between image pixel coordinates and the battery pack's physical coordinates; an electrochemical CO sensor embedded in a key area within the battery pack, connected to an external interface via high-temperature resistant and corrosion-resistant cables; and miniature piezoresistive pressure sensors deployed at the geometric center and four corners of the battery pack's sealed cavity, directly welded to the inner wall of the lower casing to sense internal gas pressure.

[0019] A method for a battery pack thermal runaway early warning and control system based on multi-source information fusion, the specific steps of which are as follows:

[0020] Step S1: Perform high-precision time synchronization multi-source data acquisition and preprocessing;

[0021] Step S2: Construct spatiotemporal fusion tensors and extract advanced features;

[0022] Step S3: Risk assessment and cell-level positioning based on mechanism and data;

[0023] Step S4: Perform dynamic early warning generation and closed-loop decision support.

[0024] Further specifying, the specific method in step S1 is as follows: after the system is powered on, the main processor of the computing control layer sends a global synchronization pulse signal to collect electrothermal data: the BMS AFE synchronously triggers the sampling of all voltage and temperature channels at a fixed period. After sampling, the values ​​are converted by the internal ADC and uploaded in real time via daisy chain or parallel bus. Simultaneously, current sensor data is synchronously acquired via a high-speed ADC. Each data packet is tagged with a microsecond-level timestamp generated by a high-precision hardware clock and encoded with the predefined unique physical location of the battery cell. Image data acquisition: After receiving the synchronization pulse, the infrared thermal imager and industrial camera respectively capture a thermal image and a visible light image. The image data is transmitted through the interface, and the same timestamp is embedded in the image header. Gas and pressure data acquisition: Gas and pressure sensor data are sampled at a frequency of 100Hz through independent ADC modules and also embedded with synchronization timestamps. Data buffering and alignment: All timestamped raw data is sent to a first-in-first-out circular buffer. The main control software uses a reference clock as a reference to perform time alignment processing on all data in the buffer, ensuring that voltage, temperature, image, gas, and pressure data at the same moment can accurately correspond in subsequent fusion analysis.

[0025] Further specifying, the specific steps in step S2 are as follows:

[0026] Step S2.1, Construct the spatiotemporal fusion tensor: Define a sliding time window T w With a sliding step size T, for each cell i within the current window, extract its timestamp-aligned data sequence: voltage sequence V. i(t) Temperature sequence T i(t) and the corresponding total current sequence I (t) Each sequence contains T w For each sampling period of data points, based on the cell location encoding, the four-dimensional spatial topological encoding vector Li is calculated using the Laplace feature mapping algorithm;

[0027] Step S2.2: Organize the above data into a slice of a four-dimensional tensor X. The four dimensions of tensor X are:

[0028] Dimension 1: cell index; Dimension 2: time step; Dimension 3: physical quantity type; and Dimension 4: spatial topology encoding. The sliding window advances by T each time, and the tensor X is dynamically updated, forming the basis for continuous time series analysis.

[0029] Step S2.3: Perform parallel feature extraction, specifically gas feature extraction: For the concentration sequence Cg(t) of each gas sensor within the time window, first perform detrending and normalization processing, then use the 'db4' wavelet basis for 4-level wavelet packet decomposition to obtain 16 sub-bands, and calculate the coefficient energy E of the 7th and 8th sub-bands. high F gas =log10(E high +1);

[0030] Perform pressure feature extraction: extract the original sequence P from each pressure sensor. r(t) A standard Kalman filter is applied for smoothing. The state equation assumes a uniform pressure change. The observation noise covariance is set using sensor calibration data. After filtering, the sequence is calculated over the entire time window T. w The slope of the linear fit within F. press =|k|;

[0031] Surface deformation feature extraction: For consecutive multi-frame visible light images, perspective correction is first performed using calibration parameters. Then, the Lucas-Kanade optical flow method is used to track dozens of predefined feature points (distributed on a flat surface) on the battery pack casing. The T value of each feature point is calculated. w The displacement vector within a time interval is projected onto the normal direction of the shell to obtain the bulge height change sequence. The average slope of the change sequence for all feature points is then calculated to obtain F. deform。

[0032] Further specifying, the specific steps of step S3 are as follows: Step S3.1, perform high-precision estimation of the internal state of the battery cell, and input the current sequence I of battery cell i extracted from the spatiotemporal tensor X. (t) Voltage sequence V i(t) and temperature sequence T i(t) The coupled model of a second-order RC equivalent circuit model and a first-order lumped-parameter thermal model is used as the basis for the state equations of the extended Kalman filter. The state variables of the electrical model are: x e =[SOC, U1, U2] T Where U1 and U2 are polarization voltages, and the thermal model state variables are: x t =[T core ] T The observation equation outputs: terminal voltage Vest = OCV(SOC) + I*R0 + U1 + U2, Surface temperature T surfest For each cell, the EKF's "prediction-update" loop is run at each time step. Through iteration, the filter finally outputs the cell's position within the time window T. w The mean state of charge (SOC) estimated in the best internal estimation i and average DC internal resistance R i The DC internal resistance R0 is updated synchronously with the state estimation through an online parameter identification algorithm;

[0033] Step S3.2: Based on the partial order lattice, perform screening to collect all N cell pairs and define the partial order relation: For cells A and B, if SOC A ≤SOC B And R A ≥R B If A ≤ B, then A ≤ B. Construct a partially ordered set and draw its Hasse diagram to find all minimal elements, forming a high-risk candidate set M. The cells in set M have low SOC and high internal resistance, making them the most vulnerable link in the group.

[0034] Step S3.3, construct the thermal runaway risk propagation field: For each cell i, calculate its state deviation vector:

[0035] ΔS i =[ΔSOC i ΔR i , αV i αT i ]. Where ΔSOC i and ΔR i It is the deviation relative to the average value of the cells excluding set M; αV i and αT i The voltage and temperature change rates are calculated from the spacetime tensor. Based on the physical location coordinates of all battery cells, a non-uniform cubic grid is established in the three-dimensional space of the battery pack. The grid resolution is adaptively adjusted according to the cell density to ensure that each grid contains approximately 5-20 cells, balancing spatial resolution and statistical significance. For each grid G, the deviation vector ΔS of all cells within it is considered as the sample points of that grid. A multivariate Gaussian kernel density estimation method is used to calculate the four-dimensional joint risk probability density ρ(G) at the center point of grid G. The bandwidth matrix is ​​selected using cross-validation. For any two adjacent grids Ga and Gb, the risk density gradient direction is calculated. The gradient directions between all adjacent grids constitute a discrete three-dimensional vector field V.

[0036] Step S3.4: Perform precise localization based on manifold analysis. Riemannian manifold analysis is performed on the three-dimensional vector field V, treating the vector field as a smooth manifold defined in three-dimensional space. First, the space is divided into multiple overlapping local neighborhoods. Within each neighborhood, the average direction of all vector directions is calculated, and the standard deviation of the angle between each vector and the average direction is calculated. Neighborhoods with a standard deviation less than a threshold are marked as "high-consistency regions." Next, within these high-consistency regions, regions where vector directions commonly point to a core sub-region are identified. By calculating the divergence and Lyapunov exponent of the vector field, stable "local attractors" are quantitatively identified. The spatial region corresponding to this attractor is the core region where risk is spontaneously converging and amplifying. Finally, the cells contained in all grids covered by this "attractor region" are extracted to obtain set A, and A∩M is calculated. The cells in the intersection are both those with the worst health status in the population and located at the core of risk convergence; therefore, they are ultimately determined to be the cells with the highest current thermal runaway risk, and their precise number list L is output. final .

[0037] Further specifying, the specific steps in step S4 are as follows: Step S4.1, perform multi-feature fusion and risk quantification for L final For each cell j in the list, extract all relevant features: internal state features (SOC). j R j ), gas characteristics F gas Pressure characteristics F press Deformation characteristics F deform For each feature, min-max normalization is performed, mapping it to the interval [0, 1]. The normalization parameters are dynamically updated based on the statistical distribution of long-term historical operating data of the battery pack. An adaptive weighted fusion algorithm is used to calculate the comprehensive risk coefficient K. j :K j =w s *f(SOC j R j )+w g *F gas_norm +w p *F press_norm +w d *F deform_norm

[0038] Where, f(SOC) j R j () is a function that maps SOC and R to a single health score, with initial weights set to w. s =0.5, w g =0.2, w p =0.15, w d=0.15, during system operation, the weights are dynamically fine-tuned based on the contribution of each feature to the accuracy of the early warning over a past period of time;

[0039] Step S4.2: Implement tiered early warning and decision triggering, setting two threshold levels: K high = 0.8, K low = 0.6.

[0040] If max(K) j )≥K high The system triggers a Level 1 (red) alarm, indicating "imminent thermal runaway." The system immediately executes the highest priority response. If K... low ≤max(K j ) <K high This triggers a Level 2 (yellow) alert, classifying it as a "high-risk state," and the system executes preventative measures. If max(K) j )< K low But L final If the system outputs a Level 3 (blue) alert, marked as "Early Anomaly, Requires Attention," the system has a built-in rule library for handling different warning levels and fault modes. For a Level 1 alarm, the rules are: "1. Immediately send the highest priority fault code via the CAN bus to request the vehicle controller to perform an emergency power-off; 2. Activate the audible and visual alarm; 3. Package and upload the battery cell number, risk factor, and real-time data to the cloud monitoring center and the owner's APP; 4. Record all data from the 30 seconds prior to the event for subsequent analysis."

[0041] For a Level 2 alarm, the rules are: "1. Send a power limit command to the entire module where the battery cell is located; 2. Increase the power of the active thermal management system (fan / liquid cooling) of the module to the maximum; 3. Generate a maintenance suggestion work order on the local human-machine interface and the remote monitoring platform, suggesting that the module be checked first during the next shutdown."

[0042] Step S4.3 involves visualizing and outputting the results; all information is dynamically refreshed on the human-computer interaction interface. The 3D model of the battery pack is displayed in the center of the interface, L... final The battery cells in the middle are based on K j The value is highlighted with a gradient from yellow to red.

[0043] The sidebar clearly lists: alarm level, detailed serial number and physical location description of the located battery cell, contribution of each feature, comprehensive risk coefficient, list of triggered handling measures and execution status.

[0044] The system also generates structured early warning reports, which are pushed to the upper-level energy management system or cloud platform through standard protocols to achieve closed-loop operation and maintenance.

[0045] With improvements, the present invention further produces the following beneficial effects:

[0046] 1. Significantly earlier warning: By integrating and analyzing multi-dimensional weak early signals such as slight increase in internal resistance, local micro-heating, and ppb-level gas release, warnings can be issued tens of minutes before traditional physical threshold alarms, gaining valuable response time.

[0047] 2. Comprehensive and accurate assessment: Breaking down information silos, the system cross-verifies risks from multiple dimensions, including electrical, thermal, shape, and gas, greatly reducing the false alarm rate caused by single-point sensor failures or interference.

[0048] 3. Achieve precise cell-level positioning: The output directly points to the specific faulty cell, enabling maintenance personnel to carry out "targeted" handling, reducing troubleshooting time from hours to minutes, and significantly improving the efficiency of safe operation and maintenance.

[0049] 4. High level of system intelligence: It integrates physical mechanisms and data intelligence, making the system decision-making process explainable and adaptable to battery aging and changes in operating conditions.

[0050] 5. Forming a safety closed loop: It integrates a complete chain from "state perception" to "location warning" and then to "decision suggestions", providing proactive and intelligent protection tools for the safe operation of the battery pack.

[0051] 6. Based on the aforementioned multi-dimensional fusion analysis results, the system constructs a dynamic risk level assessment model. This model does not use a fixed threshold; instead, it adaptively adjusts the risk weight coefficients by combining the battery's current state of charge, temperature distribution uniformity, historical cycle data, and real-time operating load. When the comprehensive risk index exceeds the preset dynamic threshold, the system automatically triggers a tiered early warning mechanism. The warning levels are divided into three levels from low to high: yellow alert, orange warning, and red emergency, corresponding to different response times and handling intensities.

[0052] Upon triggering the warning, the location module initiates a spatial tracing algorithm. This algorithm integrates gas diffusion models, temperature field inversion calculations, and voltage anomaly propagation path analysis to pinpoint the fault source to a specific module number or even the coordinates of a single battery cell. The location results are presented in a 3D visualization on the monitoring interface, annotating key parameters such as the real-time temperature, voltage deviation, and gas concentration of the faulty battery cell, and overlaying the historical performance degradation curve of the cell, providing maintenance personnel with intuitive decision-making support.

[0053] Subsequently, the decision support engine automatically generates suggested solutions based on the warning level and fault characteristics. For the yellow alert level, the system recommends increasing monitoring frequency and recording abnormal trends; for the orange alert level, the system pushes control commands to reduce power or initiate active balancing; for the red emergency level, the system directly links with the battery management system to execute emergency protection measures such as cutting off the main circuit and activating fire suppression devices, while simultaneously pushing a complete report to the operation and maintenance management platform containing fault location information, preliminary cause analysis, and recommended maintenance strategies.

[0054] Furthermore, the system establishes a full lifecycle archive for each early warning event, incorporating the triggering conditions, fusion analysis process, location results, actual handling measures, and subsequent verification data into a knowledge base. Through continuous accumulation and machine learning, the system continuously optimizes its risk feature extraction model and decision rules, achieving iterative improvements in early warning accuracy. Operations and maintenance personnel can also manually verify and correct system decisions based on actual feedback, forming an intelligent, human-machine collaborative security management closed loop. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the overall system architecture and data flow in this invention;

[0056] Figure 2 This is a flowchart illustrating the principle of multi-source data fusion and feature extraction in this invention.

[0057] Figure 3 This is a flowchart illustrating the risk assessment and location module of the present invention; Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0059] Example:

[0060] like Figures 1-3 As shown, a battery pack thermal runaway early warning system based on multi-source information fusion includes a hardware architecture and a system architecture. The hardware architecture adopts a layered layout, including a sensing layer hardware, a computing and control layer hardware, and an execution and interaction layer hardware. The sensing layer hardware includes: a cell and voltage temperature acquisition unit, a total current acquisition unit, a battery pack surface infrared thermal imaging unit, a battery pack surface visual deformation monitoring unit, an internal gas sensing unit, and an internal pressure sensing unit. The computing and control layer hardware is responsible for the system's calculations and transmits them to the execution and interaction layer hardware. The execution and interaction layer hardware includes an execution interface and a human-machine interface.

[0061] The system architecture includes four modules, namely the multi-source heterogeneous data acquisition module: synchronously acquiring cell-level electrical and thermal data, battery pack-level thermal and deformation data, and battery pack internal environment data;

[0062] Spatiotemporal fusion and feature extraction module: "Construct a four-dimensional spatiotemporal fusion tensor" to align and structure the above asynchronous and heterogeneous data in the temporal and spatial dimensions;

[0063] Risk assessment and precise location module: First, an extended Kalman filter combined with an electrochemical-thermal coupling model is used to estimate the state of charge and DC internal resistance of each cell with high accuracy. Second, a partial order lattice of cell health status is constructed based on the state parameters of all cells. The "high-risk candidate cell set" with the worst health status is automatically screened using the dominance relationship. The phase space grid for thermal runaway risk propagation is innovatively constructed to map the cell state to its three-dimensional physical location. The "joint risk density" of each grid cell and the "vector field" characterizing the risk propagation direction are calculated. Manifold analysis is performed on the vector field to identify the "local attractor region" where the risk flow converges. Thus, the location range is narrowed from the candidate set to one or a few cells that form the risk core in space, achieving precise location.

[0064] Dynamic early warning and decision support module: It normalizes and integrates multi-dimensional risk characteristics with adaptive weighting to generate a comprehensive risk coefficient, and divides the warning into three levels: red, yellow and green.

[0065] The battery cell and voltage and temperature acquisition unit adopt a dedicated BMS analog front-end chip with high precision (error ±1mV), high sampling rate (up to 1kHz), and channel synchronization error of less than 1μs. The positive and negative terminals of each battery cell are connected to the sampling channel through Kelvin connection to eliminate the influence of cable resistance. Temperature acquisition uses an NTC thermistor attached to the large surface of the battery cell and is connected to the AFE through a multiplexer to achieve synchronization with voltage sampling.

[0066] The total current acquisition unit uses a closed-loop current sensor based on the Hall principle, with a range of -500A to +500A, a bandwidth of DC-10kHz, and a linearity better than 0.1%, for accurate measurement of charging and discharging current.

[0067] The surface infrared thermal imaging unit of the battery pack is equipped with an uncooled long-wave infrared thermal imager with a spatial resolution of not less than 320×240 pixels, thermal sensitivity (NETD) < 50mK, frame rate of not less than 10Hz, and field of view that completely covers the upper surface and main sides of the battery pack. It is mounted on a stable bracket above or to the side front of the battery pack.

[0068] The battery pack surface visual deformation monitoring unit is equipped with an industrial-grade global shutter CMOS camera with a resolution of no less than 2 million pixels and a vibration-stabilized lens. The camera is at a fixed angle to the surface of the battery pack. Through pre-calibration, a mapping relationship between image pixel coordinates and physical coordinates of the battery pack is established.

[0069] Internal gas sensing unit: An electrochemical CO sensor is pre-embedded in key areas inside the battery pack (such as module gaps, inside the pressure relief valve, and exhaust pipes). The range is 0-1000ppm, the resolution is 1ppm, and the response time T90 < 30s. The sensor is connected to the external interface through a high-temperature resistant and corrosion-resistant cable.

[0070] Internal pressure sensing unit: Miniature piezoresistive pressure sensors are deployed at the geometric center and four corners of the sealed cavity of the battery pack, with a range of 0-200 kPa (absolute pressure), an accuracy of ±0.1%FS, and a sampling rate of 100 Hz. The sensors are directly welded to the inner wall of the lower housing of the battery pack to sense the internal gas pressure;

[0071] Gas sensors can be replaced or added to models sensitive to H2 and VOCs. The specific selection depends on the characteristics of the battery chemistry system. For example, H2 sensors are preferred for ternary lithium batteries to capture early gas evolution characteristics, while VOC sensors are emphasized for lithium iron phosphate batteries to monitor electrolyte decomposition products. Pressure monitoring can adopt a distributed fiber optic sensing scheme, with optical fibers arranged in a serpentine pattern along the inner wall of the battery pack. Fiber Bragg gratings or distributed acoustic wave sensing technology are used to achieve continuous spatial pressure field measurement with a spatial resolution of 0.1m. This can accurately locate areas of abnormal pressure and avoid the risk of missed detection by single-point sensors.

[0072] The computing and control layer hardware can be selected based on the following application scenarios:

[0073] Automotive scenarios: Use high-performance multi-core microcontrollers or SoCs (system-on-a-chip) that meet automotive-grade standards (such as AEC-Q100), with a main frequency of no less than 800MHz, integrating hardware floating-point units and AI acceleration cores, and running a real-time operating system;

[0074] Energy storage scenario: Industrial-grade embedded computers or industrial control computers are used, equipped with high-performance CPUs and dedicated graphics cards, to run complex manifold analysis algorithms;

[0075] This hardware layer will communicate the computations to the sensing and execution layers via CAN FD, Ethernet, or high-speed SPI interfaces.

[0076] The execution and interaction layer hardware includes an execution interface and a human-machine interface. The execution interface provides digital I / O and analog outputs for directly controlling the battery pack's thermal management system (such as fans and liquid cooling pump valves) and power relays (to execute power limiting or disconnect commands). The human-machine interface is equipped with a local touch screen to display warning information and also has a 4G / 5G or Ethernet interface to upload all alarm data and location results to the cloud monitoring platform.

[0077] A method for a battery pack thermal runaway early warning and control system based on multi-source information fusion, the specific steps of which are as follows:

[0078] Step S1 involves high-precision time synchronization and multi-source data acquisition and preprocessing. After the system is powered on, the main processor of the computing control layer sends a global synchronization pulse signal to acquire electrothermal data: the BMS AFE synchronously triggers sampling of all voltage and temperature channels at a fixed period (e.g., 10ms). The sampled values ​​are converted by the internal ADC and uploaded in real time via daisy chain or parallel bus. At the same time, current sensor data is synchronously acquired through a high-speed ADC. Each data packet is stamped with a microsecond-level timestamp generated by a high-precision hardware clock and bound to a predefined unique physical location code of the cell (format: M01_R02_C03, indicating the 2nd row and 3rd column of the 1st module).

[0079] Image data acquisition: After receiving the synchronization pulse, the infrared thermal imager and the industrial camera respectively capture a thermal image and a visible light image. The image data is transmitted through GigE Vision or USB 3.0 interface, and the same timestamp is embedded in the image header.

[0080] Gas and pressure data acquisition: Gas and pressure sensor data are sampled at a frequency of 100Hz through a separate ADC module, and a synchronization timestamp is also embedded.

[0081] Data buffering and alignment: All timestamped raw data is sent to a first-in-first-out circular buffer. The main control software uses a reference clock (the absolute time after the system is powered on) as a reference to perform time alignment processing on all data in the buffer, ensuring that voltage, temperature, image, gas, and pressure data at the same moment can be accurately matched in subsequent fusion analysis.

[0082] Step S2: Construct spatiotemporal fusion tensors and extract advanced features;

[0083] Step S2.1, Construct the spatiotemporal fusion tensor: Define a sliding time window T w (30S) and sliding step size T (5S), for each cell i in the current window, extract its timestamp-aligned data sequence: voltage sequence V i(t) Temperature sequence T i(t) and the corresponding total current sequence I (t) Each sequence contains T w For each sampling period of data points, based on the cell location encoding, the four-dimensional spatial topological encoding vector Li is calculated using the Laplace eigenmap algorithm. This method constructs an adjacency graph using the physical coordinates of all cells. By solving the generalized eigenvalue problem, the high-dimensional spatial relationship of the cells is reduced to a low-dimensional manifold. This encoding can characterize the potential paths of thermal conduction and electrical connection between cells.

[0084] Step S2.2: Organize the above data into a slice of a four-dimensional tensor X. The four dimensions of tensor X are: Dimension 1 (cell index): i = 1, 2, ..., N (N is the total number of cells).

[0085] Dimension 2 (Time Step): t = t1, t2, ..., tM (M is the number of time points within the window);

[0086] Dimension 3 (Type of physical quantity): p∈{voltage, temperature, current};

[0087] Dimension 4 (spatial topological coding): l = 1, 2, 3, 4 (corresponding to the 4 components of Li);

[0088] As the sliding window advances by T each time, the tensor X is dynamically updated, forming the basis for continuous time series analysis.

[0089] Step S2.3: Perform parallel feature extraction, gas feature extraction: extract the concentration sequence C of each gas sensor within the time window. g(t) First, detrending and standardization are performed, then 4-level wavelet packet decomposition is performed using the 'db4' wavelet basis to obtain 16 sub-bands. The coefficient energy E of the 7th and 8th sub-bands is calculated. high F gas =log10(E high +1), this feature is sensitive to sudden gas release;

[0090] Perform pressure feature extraction: extract the original sequence P from each pressure sensor. r(t) A standard Kalman filter is applied for smoothing. The state equation assumes a uniform pressure change. The observation noise covariance is set using sensor calibration data. After filtering, the sequence is calculated over the entire time window T. w The linear fitting slope k, F within the range press =|k|, this feature reflects the drasticness of pressure changes;

[0091] Surface deformation feature extraction: For consecutive multi-frame visible light images, perspective correction is first performed using calibration parameters. Then, the Lucas-Kanade optical flow method is used to track dozens of predefined feature points (distributed on a flat surface) on the battery pack casing, and the T value of each feature point is calculated. w The displacement vector within a time interval is projected onto the normal direction of the shell to obtain the bulge height change sequence. The average slope of the change sequence for all feature points is then calculated to obtain F. deform ;

[0092] Step S3: Risk assessment and cell-level positioning based on mechanism and data;

[0093] Step S3.1: Perform high-precision estimation of the internal state of the battery cell, and input the current sequence I of battery cell i extracted from the spatiotemporal tensor X. (t) (As input u), voltage sequence V i(t) and temperature sequence T i(t) (As observed value z), a coupled model of a second-order RC equivalent circuit model and a first-order lumped-parameter thermal model is used as the basis for the state equations of the extended Kalman filter. The state variables of the electrical model are: x e =[SOC, U1, U2] T Where U1 and U2 are polarization voltages, and the thermal model state variables are: x t =[T core ] T The observation equation outputs: terminal voltage V est = OCV(SOC) + I*R0 + U1 + U2, Surface temperature T surfest (From T through the heat conduction equation) core (Calculation) For each cell, the EKF "prediction-update" loop is run at each time step. Through iteration, the filter finally outputs the cell's position within the time window T. w The mean state of charge (SOC) estimated in the best internal estimation i and average DC internal resistance R i The DC internal resistance R0 is updated synchronously with the state estimation through an online parameter identification algorithm;

[0094] Step S3.2: Based on the partial order lattice, perform screening to collect all N cell pairs and define the partial order relation: For cells A and B, if SOC A ≤SOC B And R A ≥R B If A ≤ B, then A ≤ B. Construct a partially ordered set and draw its Hasse diagram to find all minimal elements, forming a high-risk candidate set M. The cells in set M have low SOC and high internal resistance, making them the most vulnerable link in the group.

[0095] Step S3.3, construct the thermal runaway risk propagation field: For each cell i, calculate its state deviation vector:

[0096] ΔS i =[ΔSOC i ΔR i , αV i αT i ]. Where ΔSOC i and ΔR i It is the deviation relative to the average value of the cells excluding set M; αV i and αT iThe voltage and temperature change rates are calculated from the spacetime tensor. Based on the physical location coordinates of all battery cells, a non-uniform cubic grid is established in the three-dimensional space of the battery pack. The grid resolution is adaptively adjusted according to the cell density to ensure that each grid contains approximately 5-20 cells, balancing spatial resolution and statistical significance. For each grid G, the deviation vector ΔS of all cells within it is considered as the sample point of that grid. A multivariate Gaussian kernel density estimation method is used to calculate the four-dimensional joint risk probability density ρ(G) at the center point of grid G. The bandwidth matrix is ​​selected using cross-validation. For any two adjacent grids Ga and Gb, the risk density gradient direction is calculated. The gradient directions between all adjacent grids constitute a discrete three-dimensional vector field V, which intuitively shows the "flow" trend of risk in space.

[0097] Step S3.4: Perform precise localization based on population analysis. Riemannian manifold analysis is performed on the three-dimensional vector field V, treating the vector field as a smooth manifold defined in three-dimensional space. First, local consistency analysis is performed, dividing the space into multiple overlapping local neighborhoods. Within each neighborhood, the average direction of all vector directions is calculated, and the standard deviation of the angle between each vector and the average direction is calculated. Neighborhoods with a standard deviation less than a threshold are marked as "high consistency regions." Second, attractor identification is performed. Within these high consistency regions, regions where vector directions commonly point to a core sub-region are identified. By calculating the divergence of the vector field and the Lyapunov exponent, stable "local attractors" are quantitatively identified. The spatial region corresponding to this attractor is the core region where risk is spontaneously converging and amplifying. Finally, cell localization is performed, extracting all cells contained in all grids covered by the "attractor region" to obtain set A. The intersection of A and M is calculated. The cells in the intersection are both the ones with the worst health status in the population and located at the core of risk convergence, thus ultimately being determined as the cells with the highest current thermal runaway risk. Their precise number list L is output. final ;

[0098] Step S4: Dynamic early warning generation and closed-loop decision support are performed.

[0099] Step S4.1 involves multi-feature fusion and risk quantification for L. final For each cell j in the list, extract all relevant features: internal state features (SOC). j R j ), gas characteristics F gas (Take the most recent gas sensor value), pressure characteristic F press Deformation characteristics F deform For each feature, min-max normalization is performed, mapping it to the interval [0, 1]. The normalization parameters are dynamically updated based on the statistical distribution of long-term historical operating data of the battery pack. An adaptive weighted fusion algorithm is used to calculate the comprehensive risk coefficient K. jK j =w s *f(SOC j R j )+w g *F gas_norm +w p *F press_norm +w d *F deform_norm

[0100] Where, f(SOC) j R j () is a function that maps SOC and R to a single health score, with initial weights set to w. s =0.5, w g =0.2, w p =0.15, w d =0.15, during system operation, the weights are dynamically fine-tuned based on the contribution of each feature to the accuracy of the early warning over a past period of time;

[0101] Step S4.2: Implement tiered early warning and decision triggering, setting two threshold levels: K high = 0.8, K low = 0.6;

[0102] If max(K) j )≥K high The system triggers a Level 1 (red) alarm, indicating "imminent thermal runaway." The system immediately executes the highest priority response. If K... low ≤max(K j ) <K high If a Level 2 (yellow) alert is triggered, indicating a "high-risk state," the system will execute preventative measures; if max(K) j )< K low But L final If the system outputs a Level 3 (blue) alert, marked as "Early Anomaly, Requires Attention," the system has a built-in rule library for handling different warning levels and fault modes. For a Level 1 alarm, the rules are: "1. Immediately send the highest priority fault code via the CAN bus to request the vehicle controller to perform an emergency power-off; 2. Activate the audible and visual alarm; 3. Package and upload the battery cell number, risk factor, and real-time data to the cloud monitoring center and the owner's APP; 4. Record all data from the 30 seconds prior to the event for subsequent analysis."

[0103] For a Level 2 alarm, the rules are: "1. Send a power limit command to the entire module where the battery cell is located; 2. Increase the power of the active thermal management system (fan / liquid cooling) of the module to the maximum; 3. Generate a maintenance suggestion work order on the local human-machine interface and the remote monitoring platform, suggesting that the module be checked first during the next shutdown."

[0104] Step S4.3 involves visualizing and outputting the results; all information is dynamically refreshed on the human-computer interaction interface. The 3D model of the battery pack is displayed in the center of the interface, L... final The battery cells in the middle are based on K j The value is highlighted with a gradient from yellow to red;

[0105] The sidebar clearly lists: alarm level, detailed serial number of the located battery cell (e.g., M01_R05_C08) and physical location description, contribution of each feature, comprehensive risk coefficient, list of triggered handling measures and execution status.

[0106] The system also generates structured early warning reports, which are pushed to the upper-level energy management system or cloud platform through standard protocols (such as DDS and MQTT) to achieve closed-loop operation and maintenance.

[0107] A graph neural network model can be used, with battery cells as nodes and connections as edges, to perform end-to-end risk prediction instead of a fusion architecture. Extended Kalman filtering can be replaced by unscented Kalman filtering or particle filtering. Manifold analysis in risk localization can be replaced by topological data analysis methods. A cloud-edge-device collaborative architecture can be adopted to balance real-time performance and computational complexity to replace the deployment mode. Early warning information can be combined with augmented reality devices to directly highlight fault points on physical devices; or it can be directly connected to the control system to trigger automatic adjustment to replace the output mode.

[0108] The foregoing has provided a detailed description of a battery pack thermal runaway early warning system and method based on multi-source information fusion provided by the present invention. The specific embodiments are described only to aid in understanding the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its scope, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A battery pack thermal runaway early warning system based on multi-source information fusion, characterized in that: The system includes a hardware architecture and a system architecture. The hardware architecture adopts a layered layout, including a sensing layer hardware, a computing and control layer hardware, and an execution and interaction layer hardware. The sensing layer hardware includes: a cell and voltage temperature acquisition unit, a total current acquisition unit, a battery pack surface infrared thermal imaging unit, a battery pack surface visual deformation monitoring unit, an internal gas sensing unit, and an internal pressure sensing unit. The computing and control layer hardware is responsible for the system's calculations and transmits them to the execution and interaction layer hardware. The execution and interaction layer hardware includes an execution interface and a human-machine interface. The system architecture includes four modules, namely a multi-source heterogeneous data acquisition module: synchronously acquiring cell-level electrical and thermal data, battery pack-level thermal and deformation data, and battery pack internal environment data; Spatiotemporal fusion and feature extraction module: "Construct a four-dimensional spatiotemporal fusion tensor" to align and structure the above asynchronous and heterogeneous data in the temporal and spatial dimensions; Risk assessment and precise location module: First, an extended Kalman filter combined with an electrochemical-thermal coupling model is used to estimate the state of charge and DC internal resistance of each cell with high accuracy. Second, a partial order lattice of cell health status is constructed based on the state parameters of all cells. The "high-risk candidate cell set" with the worst health status is automatically screened using the dominance relationship. The phase space grid for thermal runaway risk propagation is innovatively constructed to map the cell state to its three-dimensional physical location. The "joint risk density" of each grid cell and the "vector field" characterizing the risk propagation direction are calculated. Manifold analysis is performed on the vector field to identify the "local attractor region" where the risk flow converges. This narrows the location range from the candidate set to one or a few cells that form the risk core in space, thus achieving precise location. Dynamic early warning and decision support module: It normalizes and integrates multi-dimensional risk characteristics with adaptive weighting to generate a comprehensive risk coefficient, and divides the warning into three levels: red, yellow and green.

2. The battery pack thermal runaway early warning and control system based on multi-source information fusion according to claim 1, characterized in that: The cell and voltage / temperature acquisition units employ a high-precision, high-sampling-rate dedicated BMS analog front-end chip with a channel synchronization error of less than 1μs. The positive and negative terminals of each cell are connected to the sampling channel via a Kelvin connection to eliminate the influence of cable resistance. Temperature acquisition uses an NTC thermistor attached to the large surface of the cell and connected to the AFE via a multiplexer to achieve synchronization with voltage sampling. The total current acquisition unit uses a closed-loop current sensor based on the Hall effect principle, with a range covering -500A to +500A, a bandwidth of DC-10kHz, and a linearity better than 0.1%, for accurate measurement of charging and discharging current. The battery pack surface infrared thermal imaging unit is deployed... An uncooled long-wave infrared thermal imager provides a complete field of view covering the upper surface and main sides of the battery pack. The battery pack surface visual deformation monitoring unit deploys an industrial-grade global shutter CMOS camera, positioned at a fixed angle to the battery pack surface. Pre-calibration establishes a mapping relationship between image pixel coordinates and the physical coordinates of the battery pack. The internal gas sensing unit embeds electrochemical CO sensors in key areas within the battery pack, connected to an external interface via high-temperature resistant and corrosion-resistant cables. Miniature piezoresistive pressure sensors are deployed at the geometric center and four corners of the battery pack's sealed cavity, directly welded to the inner wall of the lower casing to sense internal gas pressure.

3. The method for a battery pack thermal runaway early warning and control system based on multi-source information fusion according to claim 1, characterized in that: The specific steps are as follows: Step S1: Perform high-precision time synchronization multi-source data acquisition and preprocessing; Step S2: Construct spatiotemporal fusion tensors and extract advanced features; Step S3: Risk assessment and cell-level positioning based on mechanism and data; Step S4 involves generating dynamic early warnings and providing closed-loop decision support.

4. The method for a battery pack thermal runaway early warning and control system based on multi-source information fusion according to claim 3, Its features are: The specific method in step S1 is as follows: After the system is powered on, the main processor of the computing control layer sends a global synchronization pulse signal to collect electrothermal data: the BMS AFE synchronously triggers the sampling of all voltage and temperature channels at a fixed period. After the sampled values ​​are converted by the internal ADC, they are uploaded in real time through a daisy chain or parallel bus. Simultaneously, current sensor data is synchronously acquired via a high-speed ADC. Each data packet is stamped with a microsecond-level timestamp generated by a high-precision hardware clock and bound to a predefined unique physical location code of the battery cell. Image data acquisition is performed: after receiving the synchronization pulse, the infrared thermal imager and the industrial camera respectively capture a thermal image and a visible light image. The image data is transmitted through the interface, and the same timestamp is embedded in the image header. Gas and pressure data acquisition is performed: gas and pressure sensor data are sampled at a frequency of 100Hz through independent ADC modules and also embedded with synchronization timestamps. Data buffering and alignment are performed: all raw data with timestamps are sent to a first-in-first-out circular buffer. The main control software uses a reference clock as a reference to perform time alignment processing on all data in the buffer, ensuring that voltage, temperature, image, gas, and pressure data at the same moment can accurately correspond in subsequent fusion analysis.

5. The method for a battery pack thermal runaway early warning and control system based on multi-source information fusion according to claim 3, characterized in that: The specific steps in step S2 are as follows: Step S2.1, Construct the spatiotemporal fusion tensor: Define a sliding time window T w With a sliding step size T, for each cell i within the current window, extract its timestamp-aligned data sequence: voltage sequence V. i(t) Temperature sequence T i(t) and the corresponding total current sequence I (t) Each sequence contains T w For each sampling period of data points, based on the cell location encoding, the four-dimensional spatial topological encoding vector Li is calculated using the Laplace feature mapping algorithm; Step S2.2: Organize the above data into a slice of a four-dimensional tensor X. The four dimensions of tensor X are: Dimension 1: cell index; Dimension 2: time step; Dimension 3: physical quantity type; and Dimension 4: spatial topology encoding. The sliding window advances by T each time, and the tensor X is dynamically updated, forming the basis for continuous time series analysis. Step S2.3: Perform parallel feature extraction, specifically gas feature extraction: extract the concentration sequence C of each gas sensor within the time window. g(t) First, detrending and standardization are performed, then 4-level wavelet packet decomposition is performed using the 'db4' wavelet basis to obtain 16 sub-bands. The coefficient energy E of the 7th and 8th sub-bands is calculated. high F gas =log10(E high +1); Perform pressure feature extraction: extract the original sequence P from each pressure sensor. r(t) A standard Kalman filter is applied for smoothing. The state equation assumes a uniform pressure change. The observation noise covariance is set using sensor calibration data. After filtering, the sequence is calculated over the entire time window T. w The linear fitting slope k, F within the range press =|k|; Surface deformation feature extraction: For consecutive multi-frame visible light images, perspective correction is first performed using calibration parameters. Then, the Lucas-Kanade optical flow method is used to track dozens of predefined feature points on the battery pack casing, and the T value of each feature point is calculated. w The displacement vector within a time interval is projected onto the normal direction of the shell to obtain the bulge height change sequence. The average slope of the change sequence for all feature points is then calculated to obtain F. deform .

6. The method for a battery pack thermal runaway early warning and control system based on multi-source information fusion according to claim 3, characterized in that: The specific steps of step S3 are as follows: Step S3.1: Perform high-precision estimation of the internal state of the battery cell, and input the current sequence I of battery cell i extracted from the spatiotemporal tensor X. (t) (As input u), voltage sequence V i(t) and temperature sequence T i(t) (As observed value z), a coupled model of a second-order RC equivalent circuit model and a first-order lumped-parameter thermal model is used as the basis for the state equations of the extended Kalman filter. The state variables of the electrical model are: x e =[SOC, U1, U2] T Where U1 and U2 are polarization voltages, and the thermal model state variables are: x t =[T core ] T The observation equation outputs: terminal voltage V est = OCV(SOC) + I*R0 + U1 + U2, Surface temperature T surfest For each cell, the EKF's "prediction-update" loop is run at each time step. Through iteration, the filter finally outputs the cell's position within the time window T. w The mean state of charge (SOC) estimated in the best internal estimation i and average DC internal resistance R i The DC internal resistance R0 is updated synchronously with the state estimation through an online parameter identification algorithm; Step S3.2: Based on the partial order lattice, perform screening to collect all N cell pairs and define the partial order relation: For cells A and B, if SOC A ≤SOC B And R A ≥R B If A ≤ B, then A ≤ B. Construct a partially ordered set and draw its Hasse diagram to find all minimal elements, forming a high-risk candidate set M. The cells in set M have low SOC and high internal resistance, making them the most vulnerable link in the group. Step S3.3, construct the thermal runaway risk propagation field: For each cell i, calculate its state deviation vector: ΔS i =[ΔSOC i ΔR i , αV i αT i ]. Where ΔSOC i and ΔR i It is the deviation relative to the average value of the cells excluding set M; αV i and αT i The voltage and temperature change rates are calculated from the spacetime tensor. Based on the physical location coordinates of all battery cells, a non-uniform cubic grid is established in the three-dimensional space of the battery pack. The grid resolution is adaptively adjusted according to the cell density to ensure that each grid contains approximately 5-20 cells, balancing spatial resolution and statistical significance. For each grid G, the deviation vector ΔS of all cells within it is considered as the sample points of that grid. A multivariate Gaussian kernel density estimation method is used to calculate the four-dimensional joint risk probability density ρ(G) at the center point of grid G. The bandwidth matrix is ​​selected using cross-validation. For any two adjacent grids Ga and Gb, the risk density gradient direction is calculated. The gradient directions between all adjacent grids constitute a discrete three-dimensional vector field V. Step S3.4: Perform precise localization based on manifold analysis. Riemannian manifold analysis is performed on the three-dimensional vector field V, treating the vector field as a smooth manifold defined in three-dimensional space. First, the space is divided into multiple overlapping local neighborhoods. Within each neighborhood, the average direction of all vector directions is calculated, and the standard deviation of the angle between each vector and the average direction is calculated. Neighborhoods with a standard deviation less than a threshold are marked as "high-consistency regions." Next, within these high-consistency regions, regions where vector directions commonly point to a core sub-region are identified. By calculating the divergence and Lyapunov exponent of the vector field, stable "local attractors" are quantitatively identified. The spatial region corresponding to this attractor is the core region where risk is spontaneously converging and amplifying. Finally, the battery cells contained in all grids covered by this "attractor region" are extracted to obtain set A, and A∩M is calculated. The battery cells in the intersection are both the ones with the worst health status in the population and located at the core of risk convergence, thus ultimately being determined as the battery cells with the highest current thermal runaway risk. Their precise number list L is output. final .

7. The method for a battery pack thermal runaway early warning and control system based on multi-source information fusion according to claim 3, characterized in that: The specific steps in step S4 are as follows: Step S4.1 involves multi-feature fusion and risk quantification for L. final For each cell j in the list, extract all relevant features: internal state features (SOC). j R j ), gas characteristics F gas Pressure characteristics F press Deformation characteristics F deform For each feature, min-max normalization is performed, mapping it to the interval [0, 1]. The normalization parameters are dynamically updated based on the statistical distribution of long-term historical operating data of the battery pack. An adaptive weighted fusion algorithm is used to calculate the comprehensive risk coefficient K. j :K j =w s *f(SOC j R j )+w g *F gas_norm +w p *F press_norm +w d *F deform_norm Where, f(SOC) j R j () is a function that maps SOC and R to a single health score, with initial weights set to w. s =0.5, w g =0.2, w p =0.15, w d =0.15, during system operation, the weights are dynamically fine-tuned based on the contribution of each feature to the accuracy of the early warning over a past period of time; Step S4.2: Implement tiered early warning and decision triggering, setting two threshold levels: K high = 0.8, K low = 0.6; If max(K) j )≥K high The system triggers a Level 1 (red) alarm, indicating "imminent thermal runaway." The system immediately executes the highest priority handling procedures. If K... low ≤max(K j ) <K high If a level 2 (yellow) alert is triggered, indicating a "high-risk state," the system will execute preventative measures; if max(K) j )< K low But L final If the system outputs a level 3 alert, marked as "Early Anomaly, Requires Attention," the system has a built-in rule library for handling different warning levels and fault modes. For a level 1 alarm, the rules are: "1. Immediately send the highest priority fault code via the CAN bus to request the vehicle controller to perform an emergency power-off; 2. Activate the audible and visual alarm; 3. Package and upload the battery cell number, risk factor, and real-time data to the cloud monitoring center and the owner's APP; 4. Record all data from the 30 seconds prior to the event for subsequent analysis." For a Level 2 alarm, the rules are: "1. Send a power limit command to the entire module where the battery cell is located; 2. Increase the power of the active thermal management system (fan / liquid cooling) of the module to the maximum; 3. Generate a maintenance suggestion work order on the local human-machine interface and the remote monitoring platform, suggesting that the module be checked first during the next shutdown." Step S4.3 involves visualizing and outputting the results; all information is dynamically refreshed on the human-computer interaction interface. The 3D model of the battery pack is displayed in the center of the interface, L... final The battery cells in the middle are based on K j The value is highlighted with a gradient from yellow to red; The sidebar clearly lists: alarm level, detailed serial number and physical location description of the located battery cell, contribution of each feature, comprehensive risk coefficient, list of triggered handling measures and execution status.

8. The system also generates structured early warning reports, which are pushed to the upper-level energy management system or cloud platform through standard protocols to achieve closed-loop operation and maintenance.

Images