Lithium battery pack thermal runaway early warning system based on multi-modal perception

Abstract

The present application relates to the technical field of lithium battery safety monitoring, and discloses a lithium battery pack thermal runaway early warning system based on multi-modal perception. The system comprises a multi-source perception data acquisition module, a thermal field characteristic tensor construction module, a thermal field reconstruction module and a thermal coupling correlation network generation module. The multi-source perception data acquisition module acquires multi-dimensional heterogeneous data from temperature sensors, voltage and current monitoring units, gas composition detectors and acoustic emission sensors, aligns the data with timestamps and compensates for missing values to generate a standardized data warehouse; the thermal field characteristic tensor construction module extracts temperature gradients, electrochemical responses and other characteristics at multiple scales from the data warehouse, and constructs a tensor in combination with time continuity; the thermal field reconstruction module generates a correlation graph based on the spatiotemporal distribution of the characteristics, and fuses the correlation graph into a three-dimensional thermal field reconstruction atlas of the lithium battery pack; and the thermal coupling correlation network generation module extracts characteristic vector clusters, calculates entropy weight values, and generates a network according to the spatial adjacency relationship of high-entropy nodes. The system realizes multi-dimensional data fusion and can comprehensively depict the thermal runaway evolution law.

Description

Technical Field

[0001] This invention relates to the field of lithium battery safety monitoring technology, specifically to a lithium battery pack thermal runaway early warning system based on multimodal sensing. Background Technology

[0002] Currently, most technical solutions for thermal runaway early warning in lithium battery packs focus on monitoring a single physical quantity. For example, temperature sensors are used to collect real-time temperature data of the battery surface or interior, triggering an early warning when the temperature exceeds a preset threshold. These solutions can only capture abnormal temperature signals in the later stages of thermal runaway, making it difficult to provide effective early warning during the incubation phase (such as internal micro-short circuits or slow electrolyte decomposition). The timeliness of the warnings is severely insufficient, failing to allow enough time for subsequent emergency response. Other solutions use voltage and current monitoring, analyzing sudden changes in voltage and current curves during charging and discharging to determine the battery's health. However, voltage and current signals are easily affected by external factors such as charge / discharge rate and ambient temperature. When minor internal faults occur, voltage and current changes are not significant, leading to high false alarm and false negative rates, making it difficult to meet the reliability requirements for early warning under complex operating conditions.

[0003] In recent years, some studies have attempted to introduce single gas sensors or acoustic sensors for auxiliary monitoring. For example, early warning can be achieved by detecting the concentration of characteristic gases such as CO and HF released in the early stages of lithium battery thermal runaway, or by capturing weak acoustic signals generated by changes in the internal structure of the battery. However, such single-mode extension schemes still have significant limitations: gas diffusion speed is greatly affected by ambient airflow, and local gas concentration monitoring is difficult to reflect the overall fault state of the battery pack; acoustic signals are easily interfered with by external vibrations during propagation, and a single acoustic sensor cannot effectively distinguish fault signals from environmental noise, resulting in limited monitoring accuracy. At the same time, existing technical solutions generally lack the ability to effectively fuse and correlate multi-source monitoring data. The data of each monitoring dimension are independent of each other, making it impossible to construct the intrinsic correlation between multiple physical quantities such as temperature, electrochemistry, gas, and acoustic signatures during the thermal runaway of lithium battery packs. This makes it difficult to comprehensively characterize the evolution law of thermal runaway, resulting in a narrow coverage and weak adaptability of the early warning system, which cannot meet the actual needs of lithium battery pack safety monitoring in different application scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a lithium battery pack thermal runaway early warning system based on multimodal sensing, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a lithium battery pack thermal runaway early warning system based on multimodal sensing, the system comprising:

[0006] The multi-source sensing data acquisition module is used to acquire multi-dimensional heterogeneous monitoring data from temperature sensors, voltage and current monitoring units, gas composition detectors and acoustic emission sensors during the operation of lithium battery packs, and to perform timestamp alignment and missing value compensation processing on the multi-dimensional heterogeneous monitoring data to generate a standardized multi-source sensing data warehouse.

[0007] The thermal field feature tensor construction module is used to extract multi-scale features from the standardized multi-source sensing data warehouse, separate temperature gradient features, electrochemical response features, gas evolution features and acoustic wave features, and construct thermal field feature tensors based on the temporal continuity of each feature dimension.

[0008] The thermal field reconstruction module is used to generate a feature dimension correlation graph based on the spatiotemporal distribution characteristics of each feature dimension in the thermal field feature tensor, and to construct a three-dimensional thermal field reconstruction map of the lithium battery pack by fusing all feature dimension correlation graphs.

[0009] The thermally coupled correlation network generation module is used to extract feature vector clusters of each feature dimension in the thermal field feature tensor, calculate the entropy weight value of each feature vector cluster, filter high-entropy feature nodes in combination with the three-dimensional thermal field reconstruction map of the lithium battery pack, and generate a thermally coupled correlation network based on the spatial adjacency relationship of the high-entropy feature nodes.

[0010] Preferably, the multidimensional heterogeneous monitoring data includes the cell surface temperature matrix, charge-discharge cycle curve, electrolyte decomposition gas concentration spectrum, and battery casing vibration spectrum.

[0011] Preferably, the timestamp alignment and missing value compensation processing specifically includes:

[0012] A sliding window mechanism is used to align the time scales of multi-source data acquisition and establish a unified time reference.

[0013] Missing sensor data is compensated by interpolation algorithm based on adjacent cell features to generate a complete spatiotemporal monitoring matrix;

[0014] A standardized multi-source sensing data warehouse is constructed based on the complete spatiotemporal monitoring matrix.

[0015] Preferably, constructing a three-dimensional thermal field reconstruction map of a lithium battery pack specifically includes:

[0016] Extract the propagation direction vector of the temperature gradient feature from the thermal field feature tensor;

[0017] Calculate the covariance degree between electrochemical response characteristics and gas evolution characteristics;

[0018] The weights of feature nodes are determined based on the frequency domain energy distribution of acoustic wave characteristics.

[0019] The three-dimensional thermal field reconstruction map of the lithium battery pack is generated by fusing the conduction direction vector, covariance degree, and feature node weights.

[0020] Preferably, calculating the entropy weight value of each feature vector cluster specifically includes:

[0021] Obtain the environmental interference coefficient for each feature dimension;

[0022] Extract the dispersion index of each feature vector in the feature vector cluster;

[0023] The feature dimension entropy weight value is calculated based on the aforementioned dispersion index and environmental interference coefficient.

[0024] Preferably, generating a thermally coupled correlation network specifically includes:

[0025] Identify high-entropy feature nodes in the three-dimensional thermal field reconstruction map of the lithium battery pack; calculate the thermal radiation influence factor between high-entropy feature nodes; establish a node connection topology based on the thermal radiation influence factor; and generate a thermally coupled correlation network based on the node connection topology.

[0026] Preferably, the system further includes:

[0027] An anomaly propagation chain parsing module is used to identify key heat conduction paths in the thermally coupled network, extract the real-time feature vector cluster change rate of the path nodes, and construct a thermal runaway propagation probability chain based on the change rate sequence.

[0028] The construction of the thermal runaway propagation probability chain specifically includes:

[0029] In the thermally coupled correlation network, key transmission paths are marked; the change in the cluster norm of the feature vector of the path nodes in a continuous time window is extracted.

[0030] An exponential propagation probability function is fitted based on the norm change; the propagation probability function connecting all path nodes generates a thermal runaway propagation probability chain.

[0031] Preferably, the system further includes:

[0032] The dynamic risk quantification module is used to perform time-series slicing on the thermal runaway propagation probability chain, calculate the risk coupling strength of path nodes in each time slice, and aggregate slice risk values ​​to generate a dynamic risk spectrum.

[0033] The generation of the dynamic risk spectrum specifically includes:

[0034] The thermal runaway propagation probability chain is divided into risk slice units with a fixed duration; the mean coupling strength of path nodes in each risk slice unit is calculated; and the mean coupling strength of all risk slice units is aggregated to generate a dynamic risk spectrum.

[0035] Preferably, the system further includes:

[0036] The early warning decision engine module is used to activate graded early warning commands and generate thermal runaway risk location reports based on the peak distribution and duration threshold of the dynamic risk spectrum.

[0037] The generation of the thermal runaway risk location report specifically includes:

[0038] The system analyzes the peak time-series coordinates of the dynamic risk spectrum; maps the peak coordinates to the physical location of the thermally coupled network; marks the distribution area of ​​high-risk node clusters; and generates a thermal runaway risk location report containing risk level and location coordinates.

[0039] Preferably, the activation of the tiered early warning instruction specifically includes:

[0040] A Level 1 warning is activated when the peak value of the dynamic risk spectrum continuously exceeds the first threshold.

[0041] A level-two early warning is activated when the peak value of the dynamic risk spectrum continuously exceeds the second threshold.

[0042] A Level 3 warning is activated when the average coupling strength of three consecutive risk slice units exceeds the critical value.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] By integrating multi-dimensional monitoring data such as temperature, voltage, current, gas composition, and acoustic emission through a multi-source sensing data acquisition module, and performing timestamp alignment and missing value compensation processing, a standardized multi-source sensing data warehouse is generated. This effectively solves the problems of single-dimensional data and incomplete information in traditional single-modal monitoring schemes. Compared with monitoring methods that rely solely on temperature or voltage and current, this module can simultaneously capture thermal, electrochemical, chemical, and acoustic characteristic signals during the operation of lithium battery packs. It comprehensively acquires internal state information of the battery from multiple dimensions. Whether it is the minute local temperature change caused by micro-short circuits inside the battery, the trace characteristic gases released by the early decomposition of the electrolyte, or the weak acoustic signals generated by changes in the electrode material structure, all can be accurately collected and recorded. This enables the effective capture of subtle abnormal signals during the incubation stage of thermal runaway, breaking the limitation of traditional technologies that can only provide early warning in the later stages of thermal runaway, and significantly expanding the monitoring coverage of the early warning system.

[0045] The thermal field feature tensor construction module extracts features at multiple scales, separating temperature gradient, electrochemical response, gas evolution, and acoustic wave characteristics. Based on temporal continuity, it constructs a thermal field feature tensor, transforming monitoring data from each dimension into time-correlated feature vectors rather than isolated discrete data points. This approach clearly presents the evolutionary trends of each feature over time, such as the slow accumulation of temperature gradient features, the gradual changes in electrochemical response features, and the phased appearance patterns of gas evolution and acoustic wave characteristics. This comprehensively depicts the entire feature evolution path of a lithium battery pack from normal operation to fault incubation and thermal runaway, providing rich and continuous feature support for accurate battery status assessment and avoiding the problem of missed fault features caused by data discretization in traditional technologies.

[0046] The thermal field reconstruction module generates a feature dimension correlation graph based on the spatiotemporal distribution characteristics of each feature dimension in the thermal field feature tensor, and integrates these graphs to construct a three-dimensional thermal field reconstruction map. This transforms abstract multi-source feature data into an intuitive spatial distribution map. Through this map, the temperature distribution differences in different regions within the lithium battery pack, the spatial diffusion trend of characteristic gas concentrations, and the propagation path of fault signals within the battery pack can be clearly observed, achieving a visual representation of the overall thermal field state of the battery pack. This visualization reconstruction capability enables personnel to quickly locate the specific area where the fault occurs and understand the scope and speed of fault propagation. Compared to traditional technologies that only provide single data values, this approach is more conducive to a comprehensive understanding of the battery pack's fault status, providing clear spatial information for subsequently developing targeted emergency response strategies.

[0047] The thermally coupled correlation network generation module extracts feature vector clusters, calculates entropy weights, and filters high-entropy feature nodes to construct a thermally coupled correlation network, effectively uncovering the intrinsic correlations between multi-dimensional features. For example, it can accurately identify the coupling relationship between temperature gradient changes and the rate of feature gas evolution, as well as the correlation between abnormal electrochemical responses and the timing of acoustic wave fluctuations, thereby revealing the laws governing the collaborative evolution of multiple physical quantities during the thermal runaway of lithium battery packs. This correlation analysis capability enables the early warning system to make judgments based on the collaborative changes of multiple highly correlated features, rather than relying on threshold breakthroughs of a single feature, significantly reducing the false alarm and false negative rates caused by environmental interference or accidental fluctuations of a single feature. Simultaneously, the filtering of high-entropy feature nodes focuses on feature information that has a key impact on the evolution of thermal runaway, improving the sensitivity of the early warning system to fault signals and ensuring that multi-dimensional collaborative abnormal signals can be captured in the incubation stage of thermal runaway, allowing sufficient time for emergency response and further enhancing the safety and reliability of lithium battery pack operation.

[0048] The system's standardized multi-source sensing data warehouse design enables unified format storage and management of sensor data of different types and specifications, enhancing the system's adaptability to different brands and models of lithium battery packs. It can be widely applied in various scenarios such as electric vehicles, energy storage power stations, and portable electronic devices. Meanwhile, the construction logic of the thermal field feature tensor and thermally coupled correlation network has good scalability. In the future, more dimensions of monitoring data (such as vibration signals and optical signals) can be incorporated based on new research findings, continuously optimizing the performance of the early warning model and adapting to the new demands brought about by the continuous development of lithium battery technology. It possesses long-term technological application value and market promotion potential. Attached Figure Description

[0049] Figure 1 This is a timing diagram of the lithium battery pack thermal runaway early warning system based on multimodal sensing described in this invention.

[0050] Figure 2 A flowchart for timestamp alignment and missing value compensation processing;

[0051] Figure 3 Flowchart for constructing a 3D thermal field reconstruction map of a lithium battery pack;

[0052] Figure 4 A flowchart for constructing the thermal runaway propagation probability chain;

[0053] Figure 5 A flowchart for generating a dynamic risk spectrum. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 This invention provides a lithium battery pack thermal runaway early warning system based on multimodal sensing. The system includes: a multi-source sensing data acquisition module, a thermal field feature tensor construction module, a thermal field reconstruction module, and a thermal coupling correlation network generation module, to achieve comprehensive monitoring and early warning of the thermal runaway process of the lithium battery pack. Specific implementation methods are as follows:

[0056] The multi-source sensing data acquisition module acquires multi-dimensional heterogeneous monitoring data from temperature sensors, voltage and current monitoring units, gas composition detectors, and acoustic emission sensors during the operation of the lithium battery pack. This data includes the cell surface temperature matrix, charge-discharge cycle curves, electrolyte decomposition gas concentration spectra, and battery casing vibration spectra. After timestamp alignment and missing value compensation, this data generates a standardized multi-source sensing data warehouse. The thermal field feature tensor construction module extracts multi-scale features from the standardized multi-source sensing data warehouse, separating temperature gradient features, electrochemical response features, gas evolution features, and acoustic wave features. Based on the temporal continuity of each feature dimension, it constructs a thermal field feature tensor. The thermal field reconstruction module generates a feature dimension correlation graph based on the spatiotemporal distribution characteristics of each feature dimension in the thermal field feature tensor. By fusing all feature dimension correlation graphs, it constructs a three-dimensional thermal field reconstruction map of the lithium battery pack. The thermally coupled correlation network generation module extracts feature vector clusters from each feature dimension of the thermal field feature tensor, calculates the entropy weight value of each feature vector cluster, and, combined with the three-dimensional thermal field reconstruction map of the lithium battery pack, selects high-entropy feature nodes. Based on the spatial adjacency relationships of these high-entropy feature nodes, it generates a thermally coupled correlation network.

[0057] Example 1: See Figure 2 The acquisition and processing of multidimensional heterogeneous monitoring data is illustrated through a specific application scenario. Taking a power battery pack for an electric vehicle as an example, this battery pack consists of 96 square lithium-ion cells arranged in a 6-parallel, 16-string structure. The acquisition of the cell surface temperature matrix utilizes 96 distributed PT1000 platinum resistance temperature sensors. Each sensor records the surface temperature value of its corresponding cell in real time at a sampling frequency of 2Hz. The sensor array is installed at the center of the large surface of the cell, forming a 16×6 two-dimensional temperature distribution matrix, and the data is updated every 5 milliseconds. The charge-discharge cycle curves are acquired by the main control unit of the battery management system. Through Hall current sensors and voltage acquisition circuits, the total voltage, total current, and individual cell voltage fluctuation data of the battery pack are recorded at a frequency of 100Hz, and the number of charge-discharge cycles and instantaneous power changes are simultaneously marked.

[0058] The concentration spectrum of electrolyte decomposition gases was acquired using four TGS8100 electrochemical gas sensors positioned on the top of the battery pack. These sensors detected changes in the concentrations of CO, CO2, H2, and hydrocarbons within the chamber, with a sampling period of 500 milliseconds. Each sensor outputs an 8-dimensional gas concentration vector, which was then fused into the overall gas concentration spectrum of the battery pack using a spatial weighting algorithm. The vibration spectrum of the battery casing was acquired using 12 resonant acoustic emission sensors, uniformly mounted on the surface of the lower casing of the battery pack, capturing mechanical vibration signals at a sampling rate of 200 kHz. The original time-domain waveform was subjected to a 1024-point Hanning window Fourier transform to generate a 128-dimensional spectral energy distribution vector in the 50-200 kHz frequency band.

[0059] The timestamp alignment process employs a dynamic sliding window mechanism, setting a base time window of 500 milliseconds. This window contains 10 sampling points from the temperature matrix, 100 sampling points from the charge-discharge curve, 1 sampling point from the gas spectrum, and 100 sampling points from the vibration spectrum. All data points are resampled to a unified time axis using a cubic spline interpolation algorithm, with alignment accuracy controlled within ±1 millisecond. When a temperature sensor malfunctions, causing data loss, spatial bilinear interpolation compensation is used: data from the four adjacent sensors (east, west, south, and north) are taken, and the missing values ​​are calculated using Euclidean distance weighting. For example, when the sensor at position (3, 5) fails, the temperature values ​​from points (3, 4), (3, 6), (2, 5), and (4, 5) are taken, and weighted by coefficients of 0.3, 0.3, 0.2, and 0.2 respectively, based on the inverse square of the distance, for weighted fusion.

[0060] Gas sensor data compensation employs a time-series regression model. When a gas sensor fails, the historical data from the previous 30 seconds is used to establish an ARIMA time-series model, which is then combined with the trend of adjacent sensor data to predict missing values. Vibration spectrum missing data is handled using a frequency domain migration method: the spectrum of the nearest spatially located normal sensor is selected, and the frequency response characteristics are adjusted according to the shell stiffness transfer function to generate compensated data. All compensated data is labeled with a data source identifier.

[0061] The construction of the complete spatiotemporal monitoring matrix comprises a four-dimensional data structure: time dimension (0 to T seconds), spatial dimension (16×6 cell locations), feature dimension (temperature / voltage / current / gas concentration / vibration energy), and frequency domain dimension (128 frequency points). Standardization processing involves three steps: dimensional unification converts temperature to Kelvin, voltage and current to per-unit values, gas concentration to ppm, and vibration energy to decibels; dynamic normalization uses the moving average method, calculating the mean and standard deviation of each feature dimension in a 10-second window, and performing z-score standardization; format reconstruction encapsulates heterogeneous data into time-series tensors, with each time slice containing a 96×4 monitoring matrix (temperature / voltage / gas / vibration) and a 16×128 frequency domain matrix. The final standardized multi-source sensing data warehouse adopts a hierarchical storage structure: the raw layer retains unprocessed data, the compensation layer stores interpolated data, the standard layer stores normalized data, and the metadata layer records sensor parameters and compensation logs.

[0062] Under fast charging conditions, this implementation captured typical data characteristics. When the charging current suddenly increased to 2C, the temperature matrix showed an increase in the temperature gradient of the 7th string of cells, with a temperature difference of 4.2K between adjacent cells. The gas sensor detected that the CO concentration increased from 8ppm to 35ppm within 120 seconds. The vibration spectrum showed an energy peak at 172kHz, with an amplitude increase of 12dB compared to the normal state. The timestamp alignment mechanism successfully aligned the moment of temperature change, the inflection point of gas concentration, and the moment of vibration peak to the same time axis with an error of less than 2 milliseconds. When the 5th gas sensor temporarily failed, the CO concentration value predicted by the ARIMA model deviated from the actual recovered data by less than 3ppm. The standardization process eliminated the ±0.5K system error of the temperature sensor and the ±2dB baseline drift of the vibration sensor, forming a data basis that can be compared across modes.

[0063] Example 2: See Figure 3 The construction of the thermal field feature tensor is based on a standardized multi-source sensing data warehouse. Taking a lithium battery cluster of an energy storage power station as an example, the battery cluster consists of 288 cylindrical cells of size 21700 arranged in a 12-parallel 24-string matrix. The extraction of temperature gradient features adopts multi-scale sliding window processing: in the spatial dimension, a 3×3 neighborhood window is constructed with each cell as the center, and the temperature difference between adjacent cells is calculated to form a 288-dimensional temperature gradient vector; in the temporal dimension, three time scales of 10 seconds, 30 seconds, and 60 seconds are set to calculate the temperature change rate of each cell respectively, and finally a temperature gradient feature set containing 864 feature points is generated.

[0064] The extraction of electrochemical response characteristics focuses on the dynamic characteristics of voltage and current. By analyzing charge-discharge cycle curves, three key feature groups are extracted: instantaneous impedance characteristics are calculated by the ratio of voltage change to current change within a 1-second window to obtain the DC internal resistance sequence of 288 cells; relaxation time characteristics are obtained by fitting the exponential decay time constant to the voltage recovery curve after constant current charging; and rate response characteristics record the percentage voltage drop at different discharge rates of 2C, 1C, and 0.5C. Gas evolution characteristics are processed using spectral clustering. Principal component analysis is performed on the 32-dimensional concentration vectors output by four gas sensors, and the first eight principal components are extracted as gas evolution feature vectors, with each principal component corresponding to the release pattern of a specific gas combination.

[0065] The acoustic wave characteristics were analyzed using a joint time-frequency analysis method. Vibration signals collected by 12 acoustic emission sensors were first decomposed into 32 sub-frequency bands through wavelet packet transform, and the energy entropy value of each sub-frequency band was calculated. Subsequently, burst pulse characteristics in the time-domain waveform were extracted, including three parameters: pulse amplitude, rise time, and duration. Finally, spatial coherence analysis was used to calculate the cross-correlation coefficients of adjacent sensor signals to identify the vibration propagation mode. All feature dimensions were synchronized with a time step of 100 milliseconds to construct a thermal field feature tensor (space × feature × time) with dimensions of 288 × 128 × 60.

[0066] The generation process of the 3D thermal field reconstruction map involves multi-feature fusion. The conduction direction vector of the temperature gradient feature is obtained by calculating the heat flux density field. The negative gradient direction of the temperature field is solved at each cell location, and the heat flux conduction intensity is quantified. The covariance analysis of electrochemical response features and gas evolution features adopts a dynamic time warping algorithm to calculate the optimal path matching degree between the voltage fluctuation sequence and the gas concentration sequence. Feature pairs with a matching degree greater than 0.85 are marked as strongly correlated. The frequency domain energy weight allocation of acoustic wave features is based on the frequency band energy proportion. When the energy of the 172kHz frequency band accounts for 38% of the total energy, the weight of the feature node corresponding to this frequency band is set to 0.38.

[0067] The entropy weight calculation for the feature vector cluster incorporates an environmental interference compensation mechanism. The environmental interference coefficient is obtained through temperature and humidity sensors installed inside the battery cluster. When the ambient temperature changes by more than 5°C or the relative humidity changes by more than 20%, the interference coefficient of the corresponding feature dimension is automatically increased by 0.2. The dispersion index is calculated using a sliding window method, calculating the standard deviation and kurtosis of the feature values ​​within a 60-second window. The dispersion calculation for the voltage feature includes differentiation processing for different charging and discharging conditions. The final calculation of the entropy weight value combines the dispersion index and the environmental interference coefficient. When the dispersion index of the gas concentration feature is 0.35 and the environmental interference coefficient is 0.1, the entropy weight value for that feature dimension is 0.45.

[0068] The thermally coupled network was generated based on the selection of high-entropy feature nodes. An entropy weight threshold of 0.4 was set, selecting 126 high-entropy feature nodes from 288 cell locations. These nodes are mainly distributed at the edge of the battery cluster and at the cooling duct outlet. The calculation of the thermal radiation influence factor considered spatial distance and heat transfer efficiency; the influence factor between two adjacent cell nodes is inversely proportional to the square of the distance and directly proportional to the temperature gradient. The node connection topology was established using the Delaunay triangulation algorithm, connecting the 126 high-entropy nodes to form 352 edges, each labeled with the thermal radiation influence factor value. When the influence factor exceeds 0.75, the connection lines between nodes are thickened to indicate a strong coupling relationship. The final generated thermally coupled network contains 126 nodes and 352 edges; the node size represents the entropy weight value, and the edge thickness represents the intensity of the thermal radiation influence.

[0069] In actual operation of the battery cluster, this implementation exhibited characteristic correlation properties. When a cell experienced a slight internal short circuit, its temperature gradient characteristic value increased 2.8 times within 10 seconds. Electrochemical response characteristics showed that the voltage drop of this cell was 12% faster than that of adjacent cells, and the gas sensor detected an abnormal increase in CO concentration. Acoustic wave characteristics showed an energy peak in the 165kHz frequency band, and entropy weight calculation showed that the feature dimension weight of this cell increased to 0.52. The thermal coupling correlation network successfully identified three high-impact factor connection edges centered on the faulty cell, with impact factors reaching 0.82, 0.79, and 0.76, respectively, accurately reflecting the propagation path of thermal runaway risk.

[0070] Example 3: See Figure 4 The anomaly propagation chain analysis module operates based on a thermally coupled correlation network, which contains 126 nodes and 352 edges. Each node corresponds to a high-entropy cell in the battery cluster, and the edges represent the thermal radiation influence relationships between nodes. A modified Dijkstra algorithm is used to identify critical heat conduction paths, employing the thermal radiation influence factor as the edge weight. The algorithm starts from the node with the highest entropy weight and searches for the optimal path to other nodes. A path length constraint is set, retaining only paths containing 3-6 nodes, ultimately identifying 7 critical conduction paths. The longest path contains 6 nodes, with a total path weight of 4.35, indicating a higher probability of thermal runaway risk propagating along this path.

[0071] The calculation of the rate of change of the real-time feature vector clusters of path nodes employs a multivariate differential method. Each node's feature vector contains 32 feature values ​​across four dimensions: temperature gradient, electrochemical response, gas evolution, and acoustic wave fluctuation. A 5-second sliding window is used for the rate of change calculation, and the change in the feature vector norm within the window is measured using Euclidean distance.

[0072] ,

[0073] in: Represents a node At any moment The rate of change of the eigenvector norm, It is a node At any moment eigenvectors, It's a 5-second time interval. This represents the Euclidean norm of the vector. The calculation process uses the central difference method, and the rate of change sequence is updated every 5 seconds.

[0074] For the seven identified key propagation paths, each containing 4-6 nodes, the system establishes a rate-of-change time series for each node, sampling at 5-second intervals, and continuously tracking 60 time points (5 minutes). Cross-correlation analysis is performed on the rate-of-change sequences of adjacent nodes along the path; a propagation relationship is determined when the time-delay correlation coefficient exceeds 0.8. The exponential propagation probability function is fitted using a nonlinear least squares method. The rate-of-change sequence for each node is first standardized to eliminate dimensional differences. Then, the Levenberg-Marquardt algorithm is used to fit the exponential function.

[0075] ,

[0076] in: It is a node At any moment The probability of propagation, It is the node's basic risk coefficient (calculated based on entropy weight). It is a propagation sensitivity factor (determined based on the node's position in the path). These are environmental adjustment parameters (adjusted according to changes in ambient temperature). The fitting process is iterated until the root mean square error is less than 0.05.

[0077] The thermal runaway propagation probability chain is generated by applying the propagation probability functions to all nodes along the path. For a path containing K nodes, K propagation probability function sequences are generated, each containing 120 time points (10 minutes of data). The propagation delay between nodes is determined through cross-correlation analysis. and nodes The maximum cross-correlation coefficient of the rate of change sequence appears in the time delay At that time, node The propagation probability function is shifted to the right. Each time unit. The final generated propagation probability chain is a multi-dimensional time series containing probability propagation data for 7 paths, with each path's data dimension being path length × number of time points.

[0078] The system employs a probability threshold mechanism. When the propagation probability of a node exceeds 0.7 for 10 consecutive time points, the node is considered to be in a pre-thermal runaway state. Propagation paths are prioritized based on their weighted probability values, which are calculated by considering path length, node entropy weight, and the integral of the propagation probability. The probability chain data is updated every 30 seconds, retaining data from 60 historical time points for trend analysis during the update process.

[0079] During battery cluster operation, the module successfully captured an abnormal propagation pattern. When the temperature change rate of a certain edge node suddenly increased, its propagation probability rose from 0.3 to 0.78 within 20 seconds, followed by an increase in probability at adjacent nodes after an 8-second delay. The system identified this propagation path as containing 5 nodes, with the entire propagation taking 45 seconds to complete. The probability chain showed that the third node had the highest propagation probability peak of 0.86, indicating that this location was a critical node for thermal runaway propagation. By analyzing the propagation time differences of different paths, the system was able to predict multiple possible propagation paths of thermal runaway and their temporal sequence.

[0080] Example 4: See Figure 5 The dynamic risk quantification module performs time-series analysis based on the thermal runaway propagation probability chain. This probability chain contains data from seven propagation paths, each with 4-6 nodes. Each node has propagation probability values ​​at 120 time points, with a time resolution of 5 seconds. Time-series slicing employs a fixed-duration segmentation strategy, setting each risk slice unit to 10 seconds. Each unit contains probability data from two consecutive time points. The entire probability chain is divided into 60 risk slice units, covering a 10-minute analysis period.

[0081] The risk coupling strength of path nodes within each risk slice unit is calculated using a multi-factor weighted method. First, the propagation probability values ​​of all nodes within the slice unit are identified, and probability values ​​below 0.2 are eliminated to reduce noise interference. For each node, the average probability at two time points within that slice unit is calculated as the base strength value. Then, a node weight coefficient is introduced, determined based on the node type: 0.35 for temperature-related nodes, 0.30 for electrochemical nodes, 0.25 for gas nodes, and 0.10 for acoustic nodes. The final risk coupling strength of a node is the product of the base strength value and the node weight.

[0082] The coupling relationship between nodes is adjusted through spatial proximity. When two nodes have a connecting edge in the thermally coupled network and the distance is less than a set threshold, the risk coupling strength of these two nodes will mutually enhance each other. The enhancement magnitude is determined based on the thermal radiation influence factor of the connecting edge; for every 0.1 increase in the influence factor, the strength value of adjacent nodes increases by 5%. After calculating the risk coupling strength of all nodes within each slice unit, the arithmetic mean of these strength values ​​is taken as the overall coupling strength of that slice unit. Table 1 below shows the calculation process and results of the coupling strength of risk slice units in different time periods.

[0083] Table 1: Example of risk slice unit coupling strength calculation:

[0084]

[0085] The dynamic risk spectrum is generated by aggregating the coupling strength values ​​of all risk slice units. The 60 slice units are arranged chronologically to form a time-intensity sequence. This sequence is smoothed using a moving average filter to eliminate short-term fluctuations, with a window width of 3 slice units (30 seconds). The smoothed data constitutes the dynamic risk spectrum, with the horizontal axis representing time and the vertical axis representing the risk coupling strength value, ranging from 0 to 1.

[0086] Feature extraction of the risk spectrum includes peak detection and trend analysis. Peak detection uses the local extremum method to identify points in the risk spectrum that have risen continuously before falling, requiring the peak height to be at least 0.3 times the height of the adjacent trough. Trend analysis calculates the slope of the risk spectrum through linear fitting; a positive slope indicates an upward trend in risk, and a negative slope indicates a downward trend. The integral value of the risk spectrum is used to assess the overall risk level, calculating the integral value of risk intensity over time.

[0087] The system sets risk level thresholds, dividing the dynamic risk spectrum into three regions: low-risk (intensity value 0-0.2), medium-risk (intensity value 0.2-0.5), and high-risk (intensity value 0.5-1.0). A risk warning mechanism is triggered when three consecutive slices of the risk spectrum are in the high-risk region. Risk spectrum data is updated every 10 seconds, retaining historical data from the most recent 60 minutes for trend analysis. During battery system operation, this module clearly displays the risk evolution process. A typical event shows that the risk spectrum slowly rises from the low-risk region, first entering the medium-risk region in slice 23, reaching a peak intensity of 0.68 in slice 45, and then gradually decreasing. The entire risk event lasts approximately 8 minutes, with a risk spectrum integral value of 12.3. By analyzing the shape characteristics of the risk spectrum, different types of thermal runaway precursor modes can be distinguished, such as rapidly rising, plateauing, and fluctuating types.

[0088] Example 5: The early warning decision engine module triggers an early warning mechanism based on real-time analysis of a dynamic risk spectrum. This risk spectrum contains coupling strength values ​​of 60 time slice units with a time resolution of 10 seconds and a total duration of 10 minutes. Peak detection uses an adaptive threshold method, setting the initial threshold to 0.6 times the maximum value of the risk spectrum. When three consecutive data points exceed the threshold and the intermediate point is a local maximum, it is determined to be a valid peak. The peak time sequence coordinate record includes the start time, peak time, and end time, with time accuracy controlled at the 10-millisecond level. The duration threshold is set to three levels: Level 1 warning threshold is 30 seconds, Level 2 is 60 seconds, and Level 3 is 90 seconds. When a risk spectrum peak is detected to last for more than 30 seconds, the system activates a Level 1 warning command, sending warning code 0x01 to the battery management system via the CAN bus, and simultaneously illuminating the yellow indicator light on the dashboard.

[0089] The risk positioning process is achieved through coordinate mapping. The system maintains a mapping table between the physical locations of battery clusters and the nodes of the thermal coupling correlation network. This table records the correspondence between the installation position coordinates (X, Y, Z) of 288 battery cells and 126 network nodes. After parsing the peak timing coordinates of the risk spectrum, the high-risk node clusters at that moment are retrieved according to the timestamp. Cluster identification uses the density clustering algorithm. Taking the nodes with a risk value exceeding 0.5 as the center, adjacent nodes within a spatial distance of 15 cm are searched to form a node cluster with a diameter not exceeding 30 cm. Each cluster is marked with a three-dimensional space bounding box, and the coordinates of the center point of the bounding box are used as the risk position reference.

[0090] The generation of the thermal runaway risk positioning report includes the encapsulation of structured data. The report header records the event ID, trigger time, duration, and the highest risk level. The main body part contains three data blocks: The risk distribution map block stores the three-dimensional coordinate point set of high-risk clusters, in the format of an N×4 matrix (node ID, X, Y, Z); The level evaluation block records the risk value distribution of the nodes within the cluster, and counts the number of nodes in four intervals of 0.2 - 0.4, 0.4 - 0.6, 0.6 - 0.8, and 0.8 - 1.0; The disposal suggestion block pre-sets a disposal plan code library according to the risk level. The first-level warning corresponds to code A01 (suggest reducing the load operation), the second level corresponds to B02 (starting forced air cooling), and the third level corresponds to C03 (performing emergency isolation). The report output adopts a dual-channel mechanism. The text report is encapsulated in JSON format, containing 12 fields such as timestamp, risk level, and position coordinates, and is uploaded to the cloud platform through the MQTT protocol. The visualization report generates a heat map overlaid on a three-dimensional model, and renders the risk area on the surface of the battery cluster digital twin: yellow represents the first-level risk area, orange represents the second level, and red represents the third level. The rendering engine dynamically adjusts the transparency of the color blocks according to the risk value. For every 0.1 increase in the risk value, the transparency increases by 30%. The positioning report is updated every 5 seconds. The historical reports are stored in a circular buffer according to the time series, retaining the data of the most recent 24 hours.

[0091] The implementation of hierarchical warning includes linkage control. When the first-level warning is activated, the system reduces the charging current to 0.5C and increases the rotational speed of the cooling fan by 20%. After the second-level warning is triggered, the charging process is immediately terminated, the discharge current is limited within 0.3C, and at the same time, the standby coolant circulation pump is started. The third-level warning executes a hard cut-off instruction, disconnects the main contactor, and activates the fire retardant injection system. The injection target area is automatically calibrated according to the spatial coordinates of the risk positioning report. All warning events record operation logs, including the instruction sending time, the response status of the executing equipment, and the system parameter change curve.

[0092] In actual operation of the energy storage power station, this module successfully handled multiple risk events. In one instance, a localized overheating event in a battery cluster showed a peak in the dynamic risk spectrum lasting 45 seconds at the 32-minute mark, with an intensity value of 0.58. The system activated a level-two warning, and the location report identified three high-risk node clusters within battery module CL-7. The visualization interface showed an orange warning color in the southeast corner of this module, and temperature monitoring showed the temperature at the center of this area reached 48.6℃. The system automatically executed contingency plan B02, cutting off the charging circuit and activating dual fans for forced cooling. Ten minutes later, the risk spectrum intensity value dropped to 0.21, and the warning was lifted. The entire process generated an event report EV-20250826-07, containing 12 sets of location coordinates, 8 operation records, and cooling system operating parameter curves.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lithium battery pack thermal runaway early warning system based on multimodal sensing, characterized in that, The system includes: The multi-source sensing data acquisition module is used to acquire multi-dimensional heterogeneous monitoring data from temperature sensors, voltage and current monitoring units, gas composition detectors and acoustic emission sensors during the operation of lithium battery packs, and to perform timestamp alignment and missing value compensation processing on the multi-dimensional heterogeneous monitoring data to generate a standardized multi-source sensing data warehouse. The thermal field feature tensor construction module is used to extract multi-scale features from the standardized multi-source sensing data warehouse, separate temperature gradient features, electrochemical response features, gas evolution features and acoustic wave features, and construct thermal field feature tensors based on the temporal continuity of each feature dimension. The thermal field reconstruction module is used to generate a feature dimension correlation graph based on the spatiotemporal distribution characteristics of each feature dimension in the thermal field feature tensor, and to construct a three-dimensional thermal field reconstruction map of the lithium battery pack by fusing all feature dimension correlation graphs. The thermally coupled correlation network generation module is used to extract feature vector clusters of each feature dimension in the thermal field feature tensor, calculate the entropy weight value of each feature vector cluster, filter high-entropy feature nodes in combination with the three-dimensional thermal field reconstruction map of the lithium battery pack, and generate a thermally coupled correlation network based on the spatial adjacency relationship of the high-entropy feature nodes. Also includes: An anomaly propagation chain parsing module is used to identify key heat conduction paths in the thermally coupled network, extract the real-time feature vector cluster change rate of the path nodes, and construct a thermal runaway propagation probability chain based on the change rate sequence. The construction of the thermal runaway propagation probability chain specifically includes: In the thermally coupled correlation network, key transmission paths are marked; the change in the cluster norm of the feature vector of the path nodes in a continuous time window is extracted. An exponential propagation probability function is fitted based on the norm change; a thermal runaway propagation probability chain is generated by connecting the propagation probability functions of all path nodes. Also includes: The dynamic risk quantification module is used to perform time-series slicing on the thermal runaway propagation probability chain, calculate the risk coupling strength of path nodes in each time slice, and aggregate slice risk values ​​to generate a dynamic risk spectrum. The generation of the dynamic risk spectrum specifically includes: The thermal runaway propagation probability chain is divided into risk slice units with a fixed duration; the average coupling strength of path nodes within each risk slice unit is calculated; and the average coupling strength of all risk slice units is aggregated to generate a dynamic risk spectrum. Also includes: The early warning decision engine module is used to activate graded early warning commands and generate thermal runaway risk location reports based on the peak distribution and duration threshold of the dynamic risk spectrum. The generation of the thermal runaway risk location report specifically includes: The system analyzes the peak time-series coordinates of the dynamic risk spectrum; maps the peak coordinates to the physical location of the thermally coupled network; marks the distribution area of ​​high-risk node clusters; and generates a thermal runaway risk location report containing risk level and location coordinates.

2. The lithium battery pack thermal runaway early warning system based on multimodal sensing as described in claim 1, characterized in that, The multidimensional heterogeneous monitoring data includes the cell surface temperature matrix, charge-discharge cycle curves, electrolyte decomposition gas concentration spectrum, and battery casing vibration spectrum.

3. The lithium battery pack thermal runaway early warning system based on multimodal sensing as described in claim 1, characterized in that, The timestamp alignment and missing value compensation process specifically includes: A sliding window mechanism is used to align the time scales of multi-source data acquisition and establish a unified time reference. Missing sensor data is compensated by interpolation algorithm based on adjacent cell features to generate a complete spatiotemporal monitoring matrix; A standardized multi-source sensing data warehouse is constructed based on the complete spatiotemporal monitoring matrix.

4. The lithium battery pack thermal runaway early warning system based on multimodal sensing as described in claim 1, characterized in that, The construction of a three-dimensional thermal field reconstruction map of a lithium battery pack specifically includes: Extract the propagation direction vector of the temperature gradient feature from the thermal field feature tensor; Calculate the covariance degree between electrochemical response characteristics and gas evolution characteristics; The weights of feature nodes are determined based on the frequency domain energy distribution of acoustic wave characteristics. The three-dimensional thermal field reconstruction map of the lithium battery pack is generated by fusing the conduction direction vector, covariance degree, and feature node weights.

5. The lithium battery pack thermal runaway early warning system based on multimodal sensing as described in claim 1, characterized in that, The calculation of the entropy weight values ​​for each feature vector cluster specifically includes: Obtain the environmental interference coefficient for each feature dimension; Extract the dispersion index of each feature vector in the feature vector cluster; The feature dimension entropy weight value is calculated based on the aforementioned dispersion index and environmental interference coefficient.

6. The lithium battery pack thermal runaway early warning system based on multimodal sensing as described in claim 1, characterized in that, The generation of thermally coupled correlation networks specifically includes: Identify high-entropy feature nodes in the three-dimensional thermal field reconstruction map of the lithium battery pack; calculate the thermal radiation influence factor between high-entropy feature nodes; establish a node connection topology based on the thermal radiation influence factor; and generate a thermally coupled correlation network based on the node connection topology.

7. The lithium battery pack thermal runaway early warning system based on multimodal sensing as described in claim 1, characterized in that, The specific instructions for activating tiered early warning systems include: A Level 1 warning is activated when the peak value of the dynamic risk spectrum continuously exceeds the first threshold. A secondary early warning is activated when the peak value of the dynamic risk spectrum continuously exceeds the second threshold. A Level 3 warning is activated when the average coupling strength of three consecutive risk slice units exceeds the critical value.

Images