Multi-modal fusion lithium iron phosphate battery thermal runaway early warning method and system

Abstract

The invention discloses a multi-modal fusion lithium iron phosphate battery thermal runaway early warning method and system, and the method comprises the steps: collecting multi-source heterogeneous data, such as temperature, voltage, gas concentration and shell strain pressure, in real time through deploying a heterogeneous sensor network, and carrying out the noise reduction and time sequence feature extraction through employing a sub-linear time low-rank approximation algorithm of a Hankel matrix; constructing a cross-modal feature association network by applying a secondary time algorithm of a maximum weight sparse subgraph problem, inputting a fused feature vector into a Bayesian network health degree evaluation model for probabilistic reasoning calculation to obtain a battery health degree score and a thermal runaway risk level, generating a graded early warning signal through multi-level early warning threshold comparison, and performing early warning on the battery health degree score and the thermal runaway risk level. And corresponding prevention and control suggestions are matched. The method solves the technical problems that single physical quantity monitoring is difficult to comprehensively reflect the complex change in the battery and the response delay of a centralized processing architecture causes the early warning lag, and achieves the timely capture and accurate early warning of the early weak characteristics of thermal runaway.

Description

Technical Field

[0001] This invention relates to the field of battery safety monitoring technology, specifically to the field of lithium-ion battery thermal runaway early warning, and particularly to a multimodal data fusion method and system for thermal runaway early warning of lithium iron phosphate batteries. Background Technology

[0002] In the current era of rapid development in the new energy industry, lithium iron phosphate batteries have become the mainstream energy storage device in core fields such as electric vehicles and large-scale energy storage systems due to their high safety, long cycle life, and cost advantages. Their operational safety is directly related to the reliable operation of terminal equipment and the safety of personnel and property. However, due to factors such as battery material characteristics, differences in manufacturing processes, extreme operating conditions (such as high temperature, overcharge and over-discharge), and aging degradation, lithium iron phosphate batteries still have the risk of thermal runaway. This most serious safety hazard of lithium-ion batteries, once triggered, will cause the battery temperature to soar exponentially, accompanied by a chain reaction of electrolyte decomposition and violent gas release, which may eventually lead to battery pack fire and explosion, causing serious safety accidents.

[0003] Current mainstream battery safety monitoring technologies can be mainly divided into two categories, both of which have significant technical limitations: The first category is monitoring and simple threshold determination schemes based on a single physical parameter. These schemes monitor a single physical quantity such as battery surface temperature or cell voltage using a single sensor, and determine whether the battery is abnormal based on a preset fixed threshold. For example, they identify thermal anomalies by monitoring temperature surges or detect internal short circuits by voltage jumps. While this type of technology is simple to implement and has low deployment costs, and has seen some application in early, simple scenarios, it fundamentally cannot cover the complex electrochemical and thermodynamic coupling changes during battery thermal runaway. The incubation stage of thermal runaway is often accompanied by the coordinated evolution of weak signals from multiple dimensions, such as temperature, voltage, internal pressure, and the release of characteristic gases. A single parameter is insufficient to capture this complex process, leading to one-sided monitoring and consequently high false alarm or false negative rates, making it impossible to accurately identify the weak abnormal characteristics in the early stages of thermal runaway.

[0004] The second category is an improved scheme that uses multiple similar sensors to collect a single physical quantity. To compensate for the shortcomings of single-parameter monitoring, some technical solutions attempt to deploy multiple similar sensors to collect the same physical quantity, and optimize the judgment results through data aggregation and simple fusion algorithms. However, this type of solution still has core defects: on the one hand, its essence is still to enhance the monitoring of a single physical quantity, failing to overcome the limitation that single-dimensional data cannot reflect the multi-field coupling changes inside the battery, and cannot comprehensively capture the cross-physical field coordinated abnormal signals during the incubation period of thermal runaway; on the other hand, the data processing adopts a traditional centralized computing architecture, and all sensor data must be transmitted to a central server through a communication network for unified processing. This not only requires high communication bandwidth, but also causes system response lag due to data transmission delays and large centralized computing loads, making it difficult to capture the fleeting weak features in the early stage of thermal runaway, ultimately resulting in delayed early warning, missing the best intervention window, and failing to achieve effective risk prediction and prevention.

[0005] In summary, existing technologies generally suffer from two major pain points: first, the monitoring dimensions are limited, making it difficult to comprehensively characterize the complex electrochemical, thermodynamic, and mechanical coupling changes during battery thermal runaway, resulting in insufficient early warning accuracy; second, the data processing architecture has inherent delays, failing to capture weak early signs of thermal runaway in a timely manner, leading to poor early warning timeliness. These problems severely restrict the reliability of lithium iron phosphate battery safety monitoring, necessitating a technical solution that can achieve multi-dimensional signal collaborative sensing, efficient data processing, and accurate early warning to meet the high standards of battery safety monitoring required in the new energy sector. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems that the monitoring of a single physical quantity is insufficient to fully reflect the complex electrochemical and thermodynamic changes inside the battery, resulting in a high false alarm rate, and that the centralized data processing architecture causes system response delays, leading to delayed early warnings and missing the optimal intervention window.

[0007] To achieve the above objectives, the present invention provides a multimodal fusion-based method for early warning of thermal runaway in lithium iron phosphate batteries, comprising: By deploying a heterogeneous sensor network on and inside the battery, temperature data, voltage data, gas concentration data, and casing strain and pressure data are collected in real time to obtain a multi-source heterogeneous dataset. The multi-source heterogeneous dataset is processed for time synchronization and spatial location correspondence, and the sublinear time low-rank approximation algorithm of Hankel matrix is ​​used for noise reduction and temporal feature extraction to obtain preprocessed multimodal feature data. The preprocessed multimodal feature data is subjected to single-modal deep feature extraction, and a cross-modal feature association network is constructed using a quadratic time algorithm for the maximum weight sparse subgraph problem to obtain a fused feature vector. The fused feature vector is input into a Bayesian network health assessment model for probabilistic inference calculation to obtain a battery health score and thermal runaway risk level. Based on the battery health score and the thermal runaway risk level, a graded early warning signal is obtained by comparing multiple early warning thresholds; Based on the tiered early warning signal, the knowledge base of prevention and control strategies is queried, and the intervention measures corresponding to the current state are matched to obtain prevention and control suggestions.

[0008] Furthermore, by deploying a heterogeneous sensor network on and inside the battery, real-time data such as temperature, voltage, gas concentration, and casing strain and pressure are collected to obtain a multi-source heterogeneous dataset, specifically including: Temperature sensor arrays, voltage sensors, gas sensors, and strain pressure sensors are deployed at key locations on the battery surface and inside to obtain a configured heterogeneous sensor network. The heterogeneous sensor network is used to synchronously acquire signals from various sensors to obtain the original multi-source heterogeneous data stream; The original multi-source heterogeneous data stream is subjected to noise assessment and signal quality index calculation, and outliers are filtered out to obtain the multi-source heterogeneous dataset.

[0009] Furthermore, the multi-source heterogeneous dataset undergoes time synchronization and spatial location mapping processing, specifically including: The data from different types of sensors in the multi-source heterogeneous dataset are timestamped to obtain a time-synchronized dataset. Spatial mapping processing is performed on the time-synchronized dataset to establish the correspondence between different sensor data and battery spatial location, resulting in a spatiotemporally aligned multi-source heterogeneous dataset. The sublinear time low-rank approximation algorithm of the Hankel matrix is ​​used for noise reduction and temporal feature extraction to obtain preprocessed multimodal feature data, specifically including: Hankel matrix representations were constructed on spatiotemporally aligned multi-source heterogeneous datasets to obtain the matrix form of time-series data. The Hankel matrix is ​​decomposed using a sublinear time low-rank approximation algorithm to obtain denoised structured data. The structured data after noise reduction is subjected to a multi-scale analysis method to extract signal features at different time scales, thereby obtaining the preprocessed multimodal feature data.

[0010] Furthermore, single-modal deep feature extraction is performed on the preprocessed multimodal feature data, including: A thermal feature extraction network is applied to the temperature data in the preprocessed multimodal feature data to obtain deep temperature modal features; An electrical characteristic analysis network is applied to the voltage data in the preprocessed multimodal feature data to obtain deep voltage mode features; A chemical fingerprinting model is applied to the gas concentration data in the preprocessed multimodal feature data to obtain deep gas modal features; By applying a stress analysis model to the strain and pressure data in the preprocessed multimodal feature data, deep features of the strain modes are obtained.

[0011] A quadratic time algorithm for the maximum weight sparse subgraph problem is used to construct a cross-modal feature association network, resulting in a fused feature vector, including: The deep features of the single mode are processed by feature vectorization, and the deep features of temperature mode, voltage mode, gas mode and strain mode are converted into feature vectors of the same dimension to obtain the single mode feature vector set; The correlation metric is calculated between different modal feature vectors in the single-modal feature vector set to obtain the cross-modal correlation matrix; A quadratic time algorithm for the maximum weight sparse subgraph problem is applied to the cross-modal correlation matrix to identify the most informative association structure between different modal features, resulting in a weighted cross-modal feature association graph. Based on the weighted cross-modal feature association graph, the contribution of different features in battery abnormal state identification is calculated by combining historical data, and the feature importance weight vector is obtained. Based on the cross-modal feature association graph and the feature importance weight vector, an adaptive fusion algorithm is used to weight and integrate the multimodal features, and the feature space is optimized by dimensionality reduction to obtain the fused feature vector.

[0012] Furthermore, the fused feature vector is input into the Bayesian network health assessment model for probabilistic inference calculation, including: Based on the fused feature vectors and historical battery operation data, a structural learning algorithm is used to determine the topology of the Bayesian network, resulting in a Bayesian network topology model. Based on the Bayesian network topology model and historical operating data, the conditional probability distribution parameters of each node in the network are calculated using the maximum likelihood estimation method, resulting in a parameterized Bayesian network model. The fused feature vector is used as observational evidence and input into the parameterized Bayesian network model to obtain the input state representation; A probabilistic reasoning algorithm is applied to the input state representation to calculate the posterior probability distribution of the current state of the battery, and the state probability distribution result is obtained. Based on the state probability distribution results, a quantitative calculation is performed using a preset health scoring function to obtain the battery health score and thermal runaway risk level.

[0013] Furthermore, a battery health score and thermal runaway risk level are obtained, specifically including: Based on the posterior probability distribution output by the Bayesian network model, the probability values ​​of each state variable of the battery are extracted to obtain the state probability vector. The state probability vectors are weighted and summed, with the weights determined based on the degree of influence of each state variable on battery health, to obtain a comprehensive health index. The comprehensive health index is normalized and mapped to a preset scoring range to obtain the battery health score. Based on the battery health score and the historical thermal runaway evolution pattern database, the probability of thermal runaway occurring in the current state is calculated to obtain the thermal runaway probability value. Based on the thermal runaway probability value, a risk level is determined by mapping the thermal runaway probability value to a preset risk level range.

[0014] Furthermore, based on the battery health score and the thermal runaway risk level, a graded early warning signal is obtained by comparing multi-level early warning thresholds, specifically including: Based on the battery health score and the thermal runaway risk level, combined with historical early warning data, an adaptive threshold adjustment algorithm is used to set multi-level early warning thresholds to obtain dynamically adjusted multi-level early warning thresholds. The battery health score and the thermal runaway risk level are compared with the multi-level warning threshold to determine the current warning level and obtain the warning level determination result. Based on the warning level determination result, a warning signal containing risk information and confidence level is generated, thus obtaining the graded warning signal; Based on the tiered early warning signals, the prevention and control strategy knowledge base is queried, and the intervention measures corresponding to the current state are matched to obtain prevention and control suggestions, including: Based on the tiered early warning signal, query the prevention and control strategy knowledge base, retrieve intervention measures that match the current early warning level and risk type, and obtain a set of candidate intervention measures; The candidate intervention set is prioritized and its feasibility is assessed. The intervention most suitable for the current situation is selected to obtain the prevention and control recommendations.

[0015] Furthermore, the method also includes: Based on the battery state change data after the implementation of the aforementioned prevention and control recommendations, the early warning accuracy and intervention effectiveness are calculated to obtain the early warning effect evaluation results. Based on the evaluation results of the early warning effect, the multi-level early warning thresholds and intervention strategy priorities are updated using reinforcement learning methods to obtain the system optimization parameters.

[0016] Furthermore, the method also includes: Analyze the data flow characteristics and processing requirements of the multi-source heterogeneous dataset, allocate edge computing resources to various types of sensor data, and obtain a resource allocation scheme; Based on the resource allocation scheme, configure edge computing nodes, establish an edge computing node network, and perform time synchronization, spatial location correspondence processing, and noise reduction processing on the edge computing node network.

[0017] On the other hand, the present invention also provides a multimodal fusion lithium iron phosphate battery thermal runaway early warning system, comprising: The data acquisition module is used to collect temperature data, voltage data, gas concentration data, and shell strain and pressure data in real time through a heterogeneous sensor network deployed on the surface and inside the battery, so as to obtain a multi-source heterogeneous dataset. The data preprocessing module is used to perform time synchronization and spatial location correspondence processing on the multi-source heterogeneous dataset, and to perform noise reduction and temporal feature extraction using the sublinear time low-rank approximation algorithm of the Hankel matrix to obtain preprocessed multimodal feature data. The feature fusion module is used to extract single-modal deep features from the preprocessed multimodal feature data, and to construct a cross-modal feature association network using a quadratic time algorithm for the maximum weight sparse subgraph problem to obtain a fused feature vector. The health assessment module is used to input the fused feature vector into the Bayesian network health assessment model for probabilistic inference calculation to obtain the battery health score and thermal runaway risk level. The early warning generation module is used to obtain graded early warning signals based on the battery health score and the thermal runaway risk level by comparing multiple early warning thresholds; The control strategy module is used to query the prevention and control strategy knowledge base based on the graded early warning signal, match the intervention measures corresponding to the current state, and obtain prevention and control suggestions.

[0018] The beneficial effects of this invention are: (1) This invention deploys a heterogeneous sensor network covering the battery surface and interior to simultaneously collect four core physical quantities: temperature, voltage, gas concentration, and shell strain and pressure, thus constructing a multi-dimensional data sensing system. Compared with traditional single-parameter or similar multi-sensor monitoring schemes, this design can comprehensively capture the synergistic signals of electrochemical-thermodynamic-mechanical coupling changes during the incubation period of thermal runaway, the synchronous occurrence of local temperature rise and trace release of characteristic gases, and the correlation evolution of voltage fluctuations and shell strain, fundamentally solving the problem that a single physical quantity cannot reflect the complex changes inside the battery. The complementarity of multi-modal data can not only effectively filter single-dimensional interference signals such as the influence of ambient temperature fluctuations on surface temperature, but also accurately identify weak abnormal features in the early stage of thermal runaway, reducing the false alarm rate and keeping the false alarm rate at an extremely low level, providing a reliable data foundation for early warning.

[0019] (2) This invention innovatively applies the sublinear time-low-rank approximation algorithm of the Hankel matrix to multimodal time-series data processing. Compared with the time complexity of traditional singular value decomposition, this algorithm reduces computational complexity and improves computational efficiency by at least one order of magnitude through randomization techniques and Fourier transform optimization. This breakthrough enables edge computing devices to process high-dimensional sensor data in real time without relying on centralized servers, avoiding the delays in data transmission and centralized computing. At the same time, the algorithm can completely preserve the key structural information of time-series data while efficiently reducing noise. Combined with multi-scale analysis to extract features at different time scales, it further enhances the ability to capture the composite features of short-term fluctuations and long-term trends in the early stage of thermal runaway, providing high-quality preprocessed data for subsequent feature fusion.

[0020] (3) To address the problems of redundant information interference and difficulty in identifying associated features in multimodal data fusion, this invention introduces a quadratic time algorithm based on the maximum weight sparse subgraph problem to construct a cross-modal feature association network. This algorithm can accurately select the most informative association structure from the correlation matrix of multimodal features, filtering out weak correlations and redundant connections, making the fusion process more targeted. At the same time, it combines SHAP values ​​and attention mechanisms to calculate feature importance weights, achieving dynamic weighted fusion. During the incubation period of thermal runaway, gas concentration and strain pressure features are given higher weights, while during normal operation, temperature and voltage features are emphasized, avoiding the limitations of traditional fixed-weight fusion. The final output low-dimensional, high-information fused feature vector can reduce the computational load of subsequent models and maximize the retention of collaborative anomaly information of multimodal data, improving the identification accuracy of early thermal runaway features.

[0021] (4) This invention abandons the traditional centralized data processing architecture and adopts a layered computing architecture of sensor layer and edge node layer, delegating core processing tasks such as time synchronization, spatial alignment, noise reduction, and preliminary feature extraction to edge nodes close to the sensors. On the one hand, localized processing at edge nodes reduces the volume of original data transmission, lowers the demand for communication bandwidth, and avoids transmission delays; on the other hand, real-time computing at the edge enables end-to-end latency of critical data processing to be controlled at the millisecond level. Compared with the second-level latency of centralized architecture, it can promptly capture the fleeting weak signals in the early stages of thermal runaway, completely solving the industry pain point of delayed early warning and missing the best intervention window, and gaining valuable time for subsequent emergency control. At the same time, edge nodes have local decision-making capabilities and can still perform basic early warning functions even in network interruption scenarios, improving system reliability.

[0022] (5) The Bayesian network health assessment model constructed in this invention breaks through the limitations of traditional fixed threshold or static models, possessing structural self-learning and parameter adaptation capabilities. It dynamically optimizes the network topology through structural learning algorithms, accurately representing the causal relationship between observed variables such as temperature and voltage and implicit variables such as battery health status and thermal runaway risk. By combining maximum likelihood estimation and Bayesian estimation to update the conditional probability distribution parameters, the model can adapt to battery aging, changes in operating conditions, and environmental differences. The health score and five-level risk level output by the model not only achieve quantitative representation of battery status but also output confidence levels through probabilistic inference, providing users with a dual reference of risk level and credible evidence, avoiding the ambiguity of traditional assessments.

[0023] (6) This invention introduces a closed-loop mechanism for early warning effect evaluation and parameter iterative optimization. Based on battery state change data after the implementation of prevention and control recommendations, it calculates core indicators such as early warning accuracy and intervention effectiveness, and locates the root causes of false alarms / missed alarms. It dynamically updates multi-level early warning thresholds and intervention strategy priorities through reinforcement learning algorithms. This closed-loop optimization enables the system to continuously evolve over time, avoiding the problem of gradual performance degradation after traditional systems go online. It ensures high early warning accuracy throughout the entire battery life cycle, extends the effective service life of the system, and reduces later maintenance costs. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the overall process of the multimodal fusion lithium iron phosphate battery thermal runaway early warning method of the present invention; Figure 2This is a schematic diagram illustrating the deployment and data acquisition of the heterogeneous sensor network of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal alignment and preprocessing process of multi-source data in the edge computing architecture of this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] Example 1: like Figure 1 As shown, this embodiment provides a multimodal fusion-based early warning method for thermal runaway of lithium iron phosphate batteries, including the following steps: S1: By deploying a heterogeneous sensor network on and inside the battery, temperature data, voltage data, gas concentration data, and casing strain and pressure data are collected in real time to obtain a multi-source heterogeneous dataset.

[0028] In this step, based on the analysis of the thermal runaway mechanism of lithium iron phosphate batteries, a heterogeneous sensor network is formed by deploying temperature sensor arrays, voltage sensors, gas sensors (detecting gases such as CO, CO2, and H2) and strain / pressure sensors at key locations on the battery surface and inside. The temperature sensor array, employing thermocouples or infrared sensors, is positioned at key locations such as the positive and negative electrodes, separator, and casing of the battery, enabling real-time monitoring of the temperature distribution inside and on the battery surface. Voltage sensors are connected to individual battery cells or modules to monitor voltage change trends. Gas sensors are deployed inside the battery pack to detect characteristic gases generated during thermal runaway. Strain / pressure sensors are installed on the surface of the battery casing to monitor casing deformation and internal pressure changes. A high-precision multi-channel data acquisition system enables synchronous acquisition of signals from various sensors. The sampling frequency is set according to different sensor types: temperature data sampling frequency is 1-10Hz, voltage data sampling frequency is 10-100Hz, gas concentration data sampling frequency is 0.1-1Hz, and strain / pressure data sampling frequency is 1-10Hz. The collected raw data were processed by noise assessment algorithms and signal quality index calculations to remove obvious outliers, resulting in a multi-source heterogeneous dataset with quality labels.

[0029] S2: Perform time synchronization and spatial location correspondence processing on the multi-source heterogeneous dataset, and use the sublinear time low-rank approximation algorithm of Hankel matrix for noise reduction and temporal feature extraction to obtain preprocessed multimodal feature data.

[0030] In this step, the first step is to perform time synchronization processing on the multi-source heterogeneous dataset. Since different types of sensors have different sampling frequencies, a timestamp calibration algorithm is needed to align all sensor data to a unified time reference. Network Time Protocol (NTP) or Precise Time Protocol (PTP) is used to achieve clock synchronization between sensor nodes, achieving millisecond-level time synchronization accuracy. Then, spatial location mapping is performed to establish a mapping relationship between sensor data and battery spatial locations, forming a spatiotemporally aligned multi-source heterogeneous dataset. Next, the Hankel matrix representation of the time-series data is constructed. For a time-series data sequence of length N, a Hankel matrix of dimension (N-L+1)×L is constructed, where L is the window length. A sublinear time low-rank approximation algorithm is applied to decompose the Hankel matrix. This algorithm has a time complexity of O(N log N), which is significantly lower than the O(N log N) time complexity of traditional singular value decomposition. 3 The reduced complexity enables edge computing devices to process high-dimensional sensor data in real time. By retaining the main low-rank components and removing noise and anomalous fluctuations, denoised structured data is obtained. Further multi-scale analysis methods, including wavelet transform and empirical mode decomposition, are applied to extract signal features at different time scales, identify short-term fluctuations and long-term trends, and finally obtain preprocessed multimodal feature data.

[0031] S3: Perform single-modal deep feature extraction on the preprocessed multimodal feature data, and construct a cross-modal feature association network using a quadratic time algorithm for the maximum weight sparse subgraph problem to obtain a fused feature vector.

[0032] In this step, specialized feature extraction algorithms are applied to the preprocessed multimodal feature data. A thermal feature extraction network based on a convolutional neural network (CNN) is applied to the temperature data, capable of identifying the spatial distribution patterns and temporal evolution characteristics of the temperature field. An electrical property analysis network based on a long short-term memory (LSTM) network is applied to the voltage data to capture the temporal dependencies of voltage changes. A chemical fingerprinting model is applied to the gas concentration data to identify the chemical characteristics of thermal runaway by analyzing the concentration ratios and trends of different gas components. A stress analysis model based on a physical constraint neural network is applied to the strain and pressure data to analyze the correlation between shell deformation and internal pressure. The extracted single-modal deep features are then vectorized into feature vectors of a unified dimension, forming a single-modal feature vector set. Correlation metrics between different modal feature vectors are calculated using mutual information, Pearson correlation coefficient, and other indicators to obtain a cross-modal correlation matrix. A quadratic time algorithm for the maximum weight sparse subgraph problem is applied, with a time complexity of O(n^2). 2(n) represents the number of features. By identifying the sparse subgraph structure with the largest weight in the correlation matrix, the most informative association patterns among different modal features are found, resulting in a weighted cross-modal feature association graph. Combining historical data, the contribution of different features to battery anomaly identification is calculated, and the importance of features is evaluated using SHAP values ​​or attention mechanisms, yielding a feature importance weight vector. Based on the cross-modal feature association graph and the feature importance weight vector, an adaptive fusion algorithm is used to weight and integrate multimodal features. The fusion algorithm dynamically adjusts the weights of each modal feature according to the current battery state, and performs dimensionality reduction processing through principal component analysis (PCA) or an autoencoder to optimize the feature space, ultimately obtaining a low-dimensional, high-information fused feature vector.

[0033] S4: Input the fused feature vector into the Bayesian network health assessment model for probabilistic inference calculation to obtain the battery health score and thermal runaway risk level.

[0034] In this step, the topology of the Bayesian network is first determined using a structure learning algorithm based on fused feature vectors and historical battery operation data. Structure learning algorithms include constraint-based methods (such as the PC algorithm) and scoring-based methods (such as the K2 algorithm and hill-climbing algorithm). By analyzing the conditional independence relationships between variables, a network structure reflecting the causal relationships between battery state variables is established, resulting in a Bayesian network topology model. This model includes observation nodes such as battery temperature, voltage, gas concentration, and strain pressure, as well as implicit nodes such as battery health status and thermal runaway risk. Using maximum likelihood estimation or Bayesian estimation methods, the conditional probability distribution parameters of each node in the network are calculated based on historical operation data, including conditional probability tables for discrete variables and conditional probability density functions for continuous variables, resulting in a parameterized Bayesian network model. The fused feature vectors are then input into the parameterized Bayesian network model as observational evidence, and probabilistic inference algorithms (such as variable elimination, belief propagation, or Markov chain Monte Carlo methods) are applied to calculate the posterior probability distribution of the battery's current state. The probability values ​​of each battery state variable are extracted to form a state probability vector. A weighted summation of the state probability vectors is performed, with weights determined based on the impact of each state variable on battery health, to obtain a comprehensive health index. This comprehensive health index is then normalized and mapped to a preset scoring range (e.g., 0-100 points) to obtain a battery health score. Based on the battery health score and a historical thermal runaway evolution pattern database, a risk mapping algorithm is used to calculate the probability of thermal runaway occurring in the current state, yielding a thermal runaway probability value. This thermal runaway probability value is then mapped to a preset risk level range, such as normal (probability <1%), low risk (1%-5%), medium risk (5%-20%), high risk (20%-50%), and extremely high risk (>50%), to obtain the thermal runaway risk level.

[0035] S5: Based on the battery health score and the thermal runaway risk level, a graded warning signal is obtained by comparing multi-level warning thresholds.

[0036] In this step, based on the battery health score and thermal runaway risk level, combined with historical warning data analysis, an adaptive threshold adjustment algorithm is used to set multi-level warning thresholds. The adaptive threshold adjustment algorithm dynamically adjusts the warning thresholds according to the battery's usage history, aging state, and current operating conditions to ensure a balance between the sensitivity and specificity of the warnings. For example, for severely aged batteries, the warning threshold is appropriately lowered to improve sensitivity; for new batteries, the threshold is appropriately raised to reduce false alarms. This results in dynamically adjusted multi-level warning thresholds, including attention threshold, warning threshold, emergency threshold, and danger threshold. The battery health score and thermal runaway risk level are compared with the multi-level warning thresholds to determine the current warning level. If the health score is below the attention threshold or the risk level is low risk, an attention-level warning is triggered; if the health score is below the warning threshold or the risk level is medium risk, a warning-level warning is triggered; if the health score is below the emergency threshold or the risk level is high risk, an emergency-level warning is triggered; if the health score is below the danger threshold or the risk level is extremely high risk, a danger-level warning is triggered. Warning signals containing risk information, confidence level, trigger time, and relevant feature values ​​are generated, resulting in graded warning signals.

[0037] S6: Based on the tiered early warning signal, query the prevention and control strategy knowledge base, match the intervention measures corresponding to the current state, and obtain prevention and control suggestions.

[0038] In this step, the prevention and control strategy knowledge base contains intervention measures corresponding to different warning levels and risk types, such as load adjustment, active cooling, current limiting, and emergency power failure. Based on the hierarchical warning signal query knowledge base, intervention measures matching the current warning level and risk type are retrieved, resulting in a candidate intervention measure set. For example, for a warning level, it is recommended to increase monitoring frequency; for a warning level, it is recommended to reduce charging and discharging power and activate active cooling; for an emergency level, it is recommended to limit current and enhance cooling; and for a danger level, it is recommended to immediately disconnect power and activate emergency response. The candidate intervention measure set is prioritized and its feasibility is assessed, considering factors such as the effectiveness of the intervention measures, implementation costs, and impact on system performance. The most suitable intervention measure for the current state is selected, resulting in prevention and control recommendations. These recommendations are output in structured data format, including the intervention measure type, execution parameters, expected effects, and execution priority, for reference and implementation by the battery management system or operators.

[0039] This embodiment collects multimodal data by deploying a heterogeneous sensor network, uses the sublinear time low-rank approximation algorithm of Hankel matrix for efficient noise reduction, applies a quadratic time algorithm for the maximum weight sparse subgraph problem to construct a cross-modal feature association network, and combines a Bayesian network health assessment model for probabilistic inference, thereby achieving accurate early warning and timely intervention for thermal runaway of lithium iron phosphate batteries. This solves the technical problems of limited accuracy of single physical quantity monitoring and response delay of centralized processing.

[0040] Example 2: This embodiment is a further refinement of step S1 in embodiment 1, such as... Figure 2 As shown, a heterogeneous sensor network deployed on and inside the battery is used to collect temperature data, voltage data, gas concentration data, and casing strain and pressure data in real time, resulting in a multi-source heterogeneous dataset, including: S1.1: Deploy temperature sensor arrays, voltage sensors, gas sensors, and strain pressure sensors at key locations on the battery surface and inside to obtain a configured heterogeneous sensor network.

[0041] In this step, based on the structural characteristics and thermal runaway mechanism analysis of lithium iron phosphate batteries, the sensor deployment locations are determined. The temperature sensor array includes at least eight temperature measurement points, deployed at the positive and negative electrode tabs, the center of the battery, the four corners of the battery, and the surface of the outer casing, forming a three-dimensional temperature field monitoring network. K-type thermocouples or Pt100 platinum resistance temperature sensors are used, with a temperature measurement range of -40℃ to 200℃ and an accuracy of ±0.5℃. Voltage sensors are connected to the positive and negative terminals of the individual battery cells, employing high-precision voltage acquisition chips, with a measurement range of 0-5V and an accuracy of ±1mV. Gas sensors are deployed inside the battery pack, including CO, CO2, and H2 sensors, using electrochemical or semiconductor gas sensors, with a detection range of 0-1000ppm and a response time of <30 seconds. Strain and pressure sensors, using thin-film pressure sensors or fiber optic strain sensors, are attached to key locations on the battery casing, with a measurement range of 0-10MPa and an accuracy of ±0.1%. All sensors are connected to the edge computing node via a data acquisition bus (such as CAN bus, Modbus or Ethernet) to form a configured heterogeneous sensor network.

[0042] S1.2: Synchronously acquire various sensor signals through the heterogeneous sensor network to obtain the original multi-source heterogeneous data stream.

[0043] In this step, a high-precision multi-channel data acquisition system is designed, including a signal conditioning circuit, an analog-to-digital converter (ADC) module, and a data buffer unit. The signal conditioning circuit amplifies, filters, and isolates the analog signals output from the sensors to improve the signal-to-noise ratio. The ADC module uses a high-precision ADC chip with a resolution of 16 bits or higher to convert analog signals to digital signals. The data buffer unit uses a FIFO buffer or a circular buffer to temporarily store the acquired data and prevent data loss. Sampling frequencies are set according to different sensor types: temperature data is sampled at 1-10Hz to capture slow temperature changes; voltage data at 10-100Hz to detect rapid voltage fluctuations; gas concentration data at 0.1-1Hz to adapt to the timescale of gas diffusion; and strain and pressure data at 1-10Hz to monitor the dynamic process of shell deformation. Synchronous acquisition of various sensor signals is achieved through a hardware clock or software timer to ensure the time correspondence of different data types, resulting in a raw, multi-source heterogeneous data stream. The raw data stream is stored in time-series format, with each data point containing information such as a timestamp, sensor ID, data type, and measured value.

[0044] S1.3: Perform noise assessment and signal quality index calculation on the original multi-source heterogeneous data stream, filter out outliers, and obtain the multi-source heterogeneous dataset.

[0045] In this step, the original multi-source heterogeneous data stream undergoes quality assessment and preliminary processing. Noise assessment algorithms include statistical analysis and signal processing methods. Statistical analysis methods calculate statistical indicators such as mean, variance, skewness, and kurtosis to identify outliers deviating from the normal distribution. Signal processing methods use Fourier transform or wavelet transform to analyze the frequency domain characteristics of the data, identifying high-frequency noise and periodic interference. Signal quality indicators include signal-to-noise ratio (SNR), effective bits per second (ENOB), and data integrity. Data with an SNR below a preset threshold (e.g., SNR < 20 dB) is marked as low-quality data. Data exhibiting significant jumps, exceeding physical limits, or remaining unchanged for extended periods are identified as outliers. Median filtering, Kalman filtering, or moving average filtering are used for initial noise suppression. For outliers, different processing strategies are employed based on their severity: isolated outliers are replaced using interpolation or the average of preceding and following data; continuous outlier segments are marked as invalid data and excluded in subsequent processing. After noise assessment and outlier removal, a multi-source heterogeneous dataset with quality labels is obtained. Each data point is assigned a quality label (such as excellent, good, average, poor) to provide a quality reference for subsequent processing.

[0046] This embodiment ensures the accuracy and reliability of multi-source heterogeneous datasets by scientifically deploying a heterogeneous sensor network, designing a high-precision synchronous acquisition system, and conducting rigorous data quality assessments, laying a solid foundation for subsequent data processing and feature extraction.

[0047] Example 3: This embodiment is a refinement of step S2 in embodiment 1, performing time synchronization and spatial location mapping processing on the multi-source heterogeneous dataset, such as... Figure 3 As shown, it includes: S2.1: Timestamp calibration is performed on the data from different types of sensors in the multi-source heterogeneous dataset to obtain a time-synchronized dataset.

[0048] In this step, due to the different sampling frequencies of different types of sensors and the potential drift of the local clocks of each sensor node, time synchronization is required. Network Time Protocol (NTP) or Precision Time Protocol (PTP) is used to synchronize the clocks between sensor nodes. NTP synchronizes the clocks of each node periodically via network communication, achieving millisecond-level synchronization accuracy; PTP uses hardware timestamps and bidirectional delay measurements, achieving microsecond-level synchronization accuracy. For sensors with low sampling frequencies (such as gas sensors), linear interpolation or spline interpolation methods are used to resample the data to a unified time reference. For sensors with high sampling frequencies (such as voltage sensors), downsampling or sliding window averaging methods are used to reduce the data rate to match other sensors. A timestamp calibration algorithm calculates the time deviation between each sensor node and the master clock, correcting the timestamp of each data point to ensure all data points are aligned on a unified time axis. After timestamp calibration, a time-synchronized dataset is obtained, with all sensor data having a consistent time reference and the time synchronization error controlled within 10 milliseconds.

[0049] S2.2: Perform spatial mapping processing on the time-synchronized dataset to establish the correspondence between different sensor data and battery spatial location, and obtain a spatiotemporally aligned multi-source heterogeneous dataset.

[0050] In this step, a mapping relationship is established between sensor data and the spatial location of the battery. Based on the sensor deployment plan, the three-dimensional spatial coordinates (x, y, z) of each sensor are recorded, and a coordinate system is established with the battery's geometric center as the origin. For the temperature sensor array, a spatial distribution model of the temperature field is established, using Kriging interpolation or radial basis function interpolation methods to reconstruct the temperature field distribution of the entire battery based on data from discrete temperature measurement points. For voltage sensors, a correlation model between voltage and the battery's internal potential distribution is established, considering the battery's electrochemical characteristics and geometric structure. For gas sensors, a gas concentration diffusion model is established, and the location and intensity of the gas source are calculated based on the sensor location and the airflow characteristics inside the battery pack. For strain and pressure sensors, a finite element model of shell stress and strain is established, and the stress distribution of the entire shell is inverted based on strain data from discrete measurement points. Through spatial mapping processing, data from different types of sensors are correlated to a unified spatial coordinate system, forming a spatiotemporally aligned multi-source heterogeneous dataset. This dataset not only contains time-series information but also spatial distribution information, comprehensively reflecting the spatiotemporal evolution of the battery's internal state.

[0051] This embodiment achieves spatiotemporal alignment of multi-source heterogeneous data through precise time synchronization and spatial mapping processing, providing a high-quality data foundation for subsequent feature extraction and fusion analysis.

[0052] Example 4: This embodiment is a further refinement of step S2 in embodiment 1. It employs the sublinear time low-rank approximation algorithm of the Hankel matrix for noise reduction and temporal feature extraction, resulting in preprocessed multimodal feature data, including: S2.3: Construct a Hankel matrix representation for the spatiotemporally aligned multi-source heterogeneous dataset to obtain the matrix form of the time-series data.

[0053] In this step, a Hankel matrix is ​​constructed for each temporal-series data sequence in a spatiotemporally aligned multi-source heterogeneous dataset. For a temporal data sequence of length N {x1, x2, ..., x...} n We select a window length L (usually L=N / 2) and construct a Hankel matrix H of dimension (N-L+1)×L, where the matrix element H(i,j)=x(i+j-1). The Hankel matrix has a special structural property: the anti-diagonal elements are equal, which effectively characterizes the autocorrelation and periodic patterns of time-series data. For multimodal data, we construct corresponding Hankel matrices for temperature, voltage, gas concentration, and strain-pressure data respectively, obtaining a matrix representation of the time-series data. The rank of the Hankel matrix reflects the complexity of the time-series data: low-rank matrices correspond to simple periodic or trend signals, while high-rank matrices correspond to complex non-stationary signals or noise.

[0054] S2.4: The Hankel matrix is ​​decomposed using a sublinear time low-rank approximation algorithm to obtain denoised structured data.

[0055] In this step, a sublinear time low-rank approximation algorithm is applied to decompose the Hankel matrix. The traditional Singular Value Decomposition (SVD) algorithm has a time complexity of O(N^2). 3 For large-scale data processing, traditional methods are inefficient. The sublinear time-based low-rank approximation algorithm employs randomization techniques and Fast Fourier Transform (FFT), reducing the time complexity to O(N log N), significantly improving computational efficiency. The specific algorithm flow includes: first, utilizing the special structure of the Hankel matrix, the matrix-vector product is quickly calculated using FFT; second, a low-dimensional subspace is constructed using a random projection method to capture the main information of the matrix; finally, SVD decomposition is performed in the low-dimensional subspace to obtain approximate singular values ​​and singular vectors. By retaining the components corresponding to the top k largest singular values ​​(k is much smaller than the matrix rank), the low-rank approximation matrix is ​​reconstructed, achieving noise reduction. The low-rank approximation matrix retains the main structural information and long-term trends of the time-series data while removing high-frequency noise and random fluctuations. The low-rank approximation matrix is ​​converted back to time-series data form, and the denoised structured data is obtained by averaging the anti-diagonal elements of the Hankel matrix. The signal-to-noise ratio of the denoised data is significantly improved, more clearly reflecting the true changes in battery state.

[0056] S2.5: Apply multi-scale analysis to the denoised structured data to extract signal features at different time scales, and obtain the preprocessed multimodal feature data.

[0057] In this step, multi-scale analysis methods are applied to the denoised structured data to extract signal features at different time scales. Multi-scale analysis methods include wavelet transform and empirical mode decomposition (EMD). Wavelet transform uses a multi-resolution analysis framework to decompose the signal into wavelet coefficients at different scales. Low-frequency coefficients reflect the long-term trend of the signal, while high-frequency coefficients reflect short-term fluctuations. Appropriate wavelet basis functions (such as Daubechies wavelets or Symlet wavelets) are selected, and discrete wavelet transform is performed on the denoised data to obtain multi-level wavelet decomposition coefficients. Empirical mode decomposition adaptively decomposes the signal into several intrinsic mode functions (IMFs), each IMF representing an oscillation mode within a specific frequency range. IMF components are iteratively extracted through a screening process, decomposing from high frequency to low frequency, finally obtaining the residual trend term. For temperature data, long-term trends (reflecting the overall battery temperature rise) and short-term fluctuations (reflecting local hot spots) are extracted; for voltage data, charge / discharge cycle characteristics and small voltage drops are extracted; for gas concentration data, the start time and cumulative amount of gas release are extracted; and for strain and pressure data, the rate and amplitude of casing deformation are extracted. The extracted multi-scale features are organized into feature vectors, including statistical features (mean, variance, peak value), frequency domain features (dominant frequency, spectral energy), and time domain features (rise time, rate of change), resulting in preprocessed multimodal feature data. This feature data not only contains the original measurements but also deep-seated temporal pattern information, providing rich input for subsequent feature fusion and state assessment.

[0058] This embodiment innovatively applies the sublinear time low-rank approximation algorithm of the Hankel matrix to the processing of battery multimodal time series data, which greatly improves the computational efficiency, enabling edge computing devices to process high-dimensional sensor data in real time. At the same time, it extracts features at different time scales through multi-scale analysis, thereby improving the detection capability of weak abnormal signals.

[0059] Example 5: This embodiment is a refinement of step S3 in embodiment 1, performing single-modal deep feature extraction on the preprocessed multimodal feature data, including: S3.1: Apply a thermal feature extraction network to the temperature data in the preprocessed multimodal feature data to obtain deep temperature modal features.

[0060] In this step, a dedicated thermal feature extraction network is designed to process the temperature data. This network is based on a one-dimensional convolutional neural network (1D-CNN) architecture, containing multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers use convolutional kernels of different sizes (e.g., 3, 5, 7) to extract local patterns and global trends from the temperature time-series data. The pooling layers use max pooling or average pooling to reduce feature dimensionality and enhance feature robustness. The fully connected layers map the extracted features to a high-dimensional feature space. The network input consists of preprocessed temperature time-series data and spatial distribution data, and the output is a deep feature vector of temperature modes, including features such as temperature rise rate, temperature gradient, temperature uniformity, and local hotspot intensity. The network is trained using historical data to learn temperature patterns under normal and abnormal conditions, enabling it to automatically identify early signs of thermal runaway, such as abnormal local temperature increases and uneven temperature distribution.

[0061] S3.2: Apply an electrical characteristic analysis network to the voltage data in the preprocessed multimodal feature data to obtain deep voltage mode features.

[0062] In this step, an electrical characteristic analysis network is designed to process voltage data. This network is based on a Long Short-Term Memory (LSTM) architecture, capable of capturing long-term dependencies and temporal patterns in voltage changes. The LSTM network includes input gates, forget gates, and output gates, selectively retaining and forgetting historical information through a gating mechanism. The network input is preprocessed voltage time-series data, including voltage values, voltage change rates, and voltage fluctuation amplitudes. The network output is a deep feature vector of voltage modes, including voltage plateau characteristics, voltage drop depth, internal resistance changes, and charge / discharge efficiency. By analyzing the morphological changes of the voltage curve, the network can identify voltage characteristics of abnormal states such as internal short circuits, electrolyte decomposition, and diaphragm damage, providing electrochemical-level information for thermal runaway early warning.

[0063] S3.3: Apply a chemical fingerprinting model to the gas concentration data in the preprocessed multimodal feature data to obtain deep gas modal features.

[0064] In this step, a chemical fingerprinting model is designed to process gas concentration data. This model is based on a multi-task learning framework, simultaneously analyzing the concentration changes of multiple gases such as CO, CO2, and H2. The model employs an attention mechanism to automatically identify the importance of different gas components in the thermal runaway process. The model input is preprocessed time-series data of multiple gas concentrations, and the output is a deep feature vector of gas modes, including features such as gas release rate, gas component ratio, gas accumulation, and gas release pattern. By analyzing the gas fingerprint features, the model can identify different types of thermal runaway reactions, such as electrolyte decomposition (producing CO2 and hydrocarbon gases), negative electrode SEI film decomposition (producing CO and H2), and positive electrode material decomposition (producing O2), providing chemical-level evidence for the diagnosis of thermal runaway mechanisms.

[0065] S3.4: Apply a stress analysis model to the strain and pressure data in the preprocessed multimodal feature data to obtain deep features of the strain modes.

[0066] In this step, a stress analysis model is designed to process the strain and pressure data. This model is based on a Physically Informed Neural Network (PINN), combining principles of materials mechanics with a data-driven approach. The model incorporates physical equations relating stress and strain into its loss function, ensuring that the predictions conform to physical laws. The model inputs are preprocessed time-series and spatial distribution data of strain and pressure, and the output is a deep feature vector of strain modes, including features such as shell deformation rate, stress concentration location, internal pressure changes, and shell stiffness changes. By analyzing the shell strain modes, the model can identify physical changes such as internal gas generation, electrode expansion, and separator contraction, providing information for the mechanical characterization of thermal runaway. The stress analysis model can also predict the risk of shell rupture, providing early warning for safety protection.

[0067] This embodiment designs a dedicated feature extraction network and model to perform deep feature extraction based on the characteristics of different modal data, fully mining the state information contained in each modal data, and laying the foundation for subsequent cross-modal feature fusion.

[0068] Example 6: This embodiment is a further refinement of step S3 in embodiment 1. It employs a quadratic time algorithm for the maximum weight sparse subgraph problem to construct a cross-modal feature association network, obtaining a fused feature vector, including: S3.5: Perform feature vectorization processing on the single-mode deep features, converting the temperature mode deep features, voltage mode deep features, gas mode deep features, and strain mode deep features into feature vectors of a unified dimension, thus obtaining a single-mode feature vector set.

[0069] In this step, the deep features extracted from each mode are standardized and vectorized. Since the dimensions and numerical ranges of different modal features vary, normalization is required to map all features to a unified numerical range (e.g., [0, 1] or [-1, 1]). Z-score normalization or Min-Max normalization is used to eliminate the influence of different dimensions. For features with inconsistent dimensions, feature expansion or compression methods are used to convert all modal features to a unified dimension d (e.g., d=128 or d=256). Feature expansion uses zero-padding or repeat-padding methods, while feature compression uses principal component analysis (PCA) or autoencoder methods. After processing, a single-modal feature vector set is obtained, including temperature feature vector V_T∈R^d, voltage feature vector V_V∈R^d, gas feature vector V_G∈R^d, and strain feature vector V_S∈R^d. All feature vectors have the same dimension and numerical range.

[0070] S3.6: Calculate the correlation metric between different modal feature vectors in the single-modal feature vector set to obtain the cross-modal correlation matrix.

[0071] In this step, correlation metrics between feature vectors of different modalities are calculated. Correlation metrics include Pearson correlation coefficient, Spearman rank correlation coefficient, mutual information, and cosine similarity. Pearson correlation coefficient measures linear correlation and is suitable for normally distributed data; Spearman rank correlation coefficient measures monotonic correlation and is robust to outliers; mutual information measures nonlinear correlation and can capture complex dependencies; cosine similarity measures the similarity of vector directions and is suitable for high-dimensional sparse data. For each pair of modal feature vectors (V_i, V_j), a correlation metric C(i,j) is calculated, and a cross-modal correlation matrix C∈R^(m×m) is constructed, where m is the number of modalities (m=4 in this embodiment). The correlation matrix is ​​a symmetric matrix, with diagonal elements equal to 1 (autocorrelation), and off-diagonal elements reflecting the association strength between different modalities. To improve computational efficiency, a sliding window method is used to calculate the correlation within different time windows, capturing the time-varying characteristics of the correlation.

[0072] S3.7: Apply a quadratic time algorithm for the maximum weight sparse subgraph problem to the cross-modal correlation matrix to identify the most informative correlation structure between different modal features, and obtain a weighted cross-modal feature correlation graph.

[0073] In this step, the cross-modal correlation matrix is ​​treated as a weighted complete graph, where nodes represent different modal features and edge weights represent correlation strengths. The maximum weight sparse subgraph problem aims to select k edges from the complete graph such that the sum of the weights of the selected edges is maximized, while satisfying the sparsity constraint (k is much smaller than the total number of edges). This problem is NP-hard, but it can be solved in quadratic time O(m) using greedy or approximate algorithms.2 The algorithm solves the problem within a given timeframe. The specific algorithm flow is as follows: First, sort all edges by weight from largest to smallest; second, sequentially add the edge with the largest weight to the subgraph until a preset number of edges k is reached or other termination conditions are met; finally, check the connectivity and sparsity of the subgraph to ensure a reasonable subgraph structure. This algorithm identifies the most informative association structures between different modal features, filters out redundant and weakly correlated connections, and obtains a weighted cross-modal feature association graph. This association graph reveals cooperative anomaly patterns in multimodal data, such as a strong correlation between temperature rise and gas release, and a strong correlation between voltage drop and strain increase. These cooperative patterns are important characteristics of thermal runaway.

[0074] S3.8: Based on the weighted cross-modal feature association graph, and combined with historical data, calculate the contribution of different features in battery abnormal state identification to obtain the feature importance weight vector.

[0075] In this step, based on weighted cross-modal feature association graphs and historical battery operation data, the contribution of different features to anomaly state identification is calculated. The SHAP (Shapley Additive exPlanations) method or an attention mechanism is used to evaluate feature importance. The SHAP value, based on the Shapley value concept in game theory, quantifies feature importance by calculating the marginal contribution of each feature to the prediction result. For each feature, its average marginal contribution across all possible feature combinations is calculated to obtain the SHAP value. The larger the SHAP value, the more important the feature. The attention mechanism automatically identifies the features most relevant to the task by learning the feature weight coefficients. An attention network is trained, with multimodal features as input and anomaly state prediction as output; the attention weights in the network reflect feature importance. Combining normal and anomaly state samples from historical data, the discriminative ability of different features in anomaly detection is statistically analyzed, and the information gain or Gini coefficient of the features is calculated. By combining SHAP values, attention weights, and statistical indicators, we obtain a feature importance weight vector W∈R^d, where each element W_i represents the importance weight of the i-th feature. The weight values ​​are normalized so that ∑W_i=1.

[0076] S3.9: Based on the cross-modal feature association graph and the feature importance weight vector, an adaptive fusion algorithm is used to weight and integrate the multimodal features, and the feature space is optimized by dimensionality reduction to obtain the fused feature vector.

[0077] In this step, based on the cross-modal feature correlation graph and the feature importance weight vector, an adaptive fusion algorithm is used to integrate multi-modal features. The adaptive fusion algorithm dynamically adjusts the fusion weights of each modal feature according to the current battery state. The specific method is as follows: First, according to the structure of the cross-modal feature correlation graph, the basic fusion weights of each modal feature are determined, and a larger weight is assigned to the modal with a higher correlation degree. Second, according to the feature importance weight vector, the weights of each feature dimension are adjusted, and a larger weight is assigned to the important features. Finally, according to the current working condition (such as charging, discharging, standing still) and health state (such as new battery, aging battery) of the battery, the fusion weights are dynamically adjusted. For example, the weight of the voltage feature is increased in the charging state, and the weight of the temperature feature is increased in the high-temperature environment. The fusion algorithm uses the weighted sum or weighted splicing method to integrate the multi-modal feature vectors into a high-dimensional fusion feature vector F∈R^(m×d). To reduce the computational complexity and avoid the curse of dimensionality, the fusion feature vector is processed for dimensionality reduction. The dimensionality reduction methods include principal component analysis (PCA), linear discriminant analysis (LDA), or autoencoder. PCA transforms the correlated features into uncorrelated principal components through orthogonal transformation, and retains the first k principal components with the largest variance; LDA maximizes the between-class distance and minimizes the within-class distance to find the optimal projection direction; the autoencoder learns the low-dimensional representation of the data through a neural network and retains the main information of the data. After dimensionality reduction processing, a low-dimensional and high-information fusion feature vector F'∈R^k is obtained, where k<<m×d (such as k = 64 or k = 128). The fusion feature vector synthesizes the complementary information of multi-modal data, eliminates redundancy and noise, and provides high-quality input for subsequent health assessment.

[0078] In this embodiment, a quadratic time algorithm for the maximum weight sparse subgraph problem is innovatively introduced into the construction of the cross-modal feature correlation network. By efficiently identifying the most informative correlation structure between different modal features, the problem of redundant information interference in traditional fusion methods is solved, enabling the system to accurately capture the collaborative abnormal patterns in multi-modal data.

[0079] Embodiment 7: This embodiment further refines step S4 in Embodiment 1. The fusion feature vector is input into the Bayesian network health assessment model for probabilistic inference calculation, including: S4.1: Based on the fusion feature vector and historical battery operation data, a structure learning algorithm is used to determine the topological structure of the Bayesian network, and a Bayesian network topological structure model is obtained.

[0080] In this step, based on fused feature vectors and historical battery operation data, a structure learning algorithm is used to determine the topology of the Bayesian network. Structure learning algorithms are divided into two categories: constraint-based methods and scoring-based methods. Constraint-based methods (such as PC algorithm and IC algorithm) identify causal relationships between variables through conditional independence tests, constructing a directed acyclic graph (DAG). The specific process is as follows: First, starting from a completely undirected graph, independence tests are performed on each pair of variables; second, edges between independent variables are deleted based on the test results; finally, the direction of the remaining edges is determined to avoid forming loops. Scoring-based methods (such as K2 algorithm, hill-climbing algorithm, and genetic algorithm) search for the optimal network structure by optimizing scoring functions (such as BIC score, AIC score, and MDL score). The scoring function balances the model's fit and complexity, avoiding overfitting. The hill-climbing algorithm starts from the initial network structure and iteratively searches for the network structure that maximizes the scoring function by adding, deleting, or reversing edges. This embodiment uses a hybrid approach: first, the PC algorithm is used to determine the network's skeleton structure, and then the hill-climbing algorithm is used to optimize the direction and weight of the edges. A Bayesian network comprises multiple nodes: observation nodes (representing various dimensions of the fused feature vector, temperature, voltage, gas concentration, strain pressure, etc.) and hidden nodes (representing battery health status, SOC, SOH, internal resistance, thermal runaway risk, etc.). Directed edges between nodes represent causal or conditional dependencies; for example, increased temperature leads to increased gas release, and a voltage drop leads to increased internal resistance. Through structure learning, a Bayesian network topology model reflecting the causal relationships between battery state variables is obtained. This model can characterize the complex dynamic behavior of the battery system.

[0081] S4.2: Based on the Bayesian network topology model and historical operating data, the conditional probability distribution parameters of each node in the network are calculated using the maximum likelihood estimation method to obtain a parameterized Bayesian network model.

[0082] In this step, based on the established Bayesian network topology and historical data, the conditional probability distribution parameters of each node in the network are calculated. For discrete variable nodes, the conditional probability distribution is represented by a conditional probability table (CPT), which records the probability of each state of the current node given the value of its parent node. The maximum likelihood estimation (MLE) method is used to estimate the probability values ​​in the CPT by statistically analyzing the frequency of each state combination based on historical data. For continuous variable nodes, the conditional probability distribution is typically assumed to be a Gaussian or Gaussian mixture distribution, with parameters including the mean and covariance matrix. The expectation-maximization (EM) algorithm or variational Bayesian method is used to estimate the distribution parameters. For missing or incomplete data, data imputation or the EM algorithm is used. To avoid overfitting and zero-probability problems, a Bayesian estimation method is used, introducing a prior distribution (such as a Dirichlet prior or Gaussian prior) to regularize the parameter estimation. Through parameter learning, a parameterized Bayesian network model is obtained, which fully describes the conditional probability distribution of each node and enables probabilistic inference and prediction. Parametric models also include hyperparameters, such as parameters of the prior distribution, learning rate, regularization coefficient, etc., which are optimized through cross-validation or grid search methods.

[0083] S4.3: Input the fused feature vector as observation evidence into the parameterized Bayesian network model to obtain the input state representation.

[0084] In this step, the fused feature vector obtained through real-time acquisition and processing is used as the input parameterized Bayesian network model for observational evidence. Each dimension of the fused feature vector corresponds to an observation node in the Bayesian network, and feature values ​​are assigned to the corresponding nodes to form the evidence set E = {e1, e2, ..., e...}. n} where eᵢ represents the observed value of the i-th observation node. For continuous observations, the value is directly assigned to the corresponding node; for discrete observations, continuous values ​​are converted into discrete states according to a preset discretization rule (such as equal-width binning or equal-frequency binning). After the evidence is input, the states of the observed nodes in the Bayesian network are determined, while the states of the hidden nodes are unknown and need to be calculated through probabilistic inference. The input state representation is obtained, which is the state configuration of the Bayesian network under the given observation evidence E, providing input for subsequent probabilistic inference.

[0085] S4.4: Apply a probabilistic reasoning algorithm to the input state representation to calculate the posterior probability distribution of the current state of the battery, and obtain the state probability distribution result.

[0086] In this step, probabilistic inference algorithms are applied to the input state representation to calculate the posterior probability distribution of hidden nodes. Probabilistic inference algorithms are divided into exact inference and approximate inference. Exact inference algorithms include variable elimination, clustered tree propagation, and joint tree algorithms, suitable for simple network structures and a small number of nodes. Variable elimination calculates marginal probability distributions step-by-step by eliminating non-query variables; its time complexity depends on the elimination order. Clustered tree propagation transforms the Bayesian network into a clustered tree structure and calculates marginal probabilities through message passing; it is suitable for tree-like or near-tree-like networks. For complex networks, exact inference is computationally intensive, so approximate inference algorithms are used, including Markov chain Monte Carlo (MCMC) methods, variational inference, and belief propagation algorithms. MCMC methods approximate the posterior distribution through random sampling, including Gibbs sampling and the Metropolis-Hastings algorithm; they have slow convergence speed but high accuracy. Variational inference transforms probabilistic inference into an optimization problem, finding the optimal approximation of the posterior distribution by minimizing the KL divergence; it is computationally efficient but may have biases. This embodiment employs the belief propagation algorithm, which iteratively propagates messages on a cyclic graph until convergence, making it suitable for medium-sized Bayesian networks. Through probabilistic inference, the posterior probability distribution P(H|E) of hidden nodes, such as battery health state, SOC, SOH, internal resistance, and thermal runaway risk, is calculated, where H represents the hidden node and E represents observed evidence. The resulting state probability distribution includes the probability distribution, expected value, variance, and confidence interval of each hidden node, comprehensively reflecting the uncertainty of the battery's current state.

[0087] S4.5: Based on the state probability distribution results, and combined with the preset health score function, quantitative calculation is performed to obtain the battery health score and thermal runaway risk level.

[0088] In this step, quantitative calculations are performed based on the state probability distribution results and a pre-defined health scoring function. The health scoring function maps the probability distribution to a single numerical value, reflecting the overall health status of the battery. The scoring function design considers multiple factors: the posterior probability of the battery's health status, the expected value of SOH (State of Health), the rate of increase in internal resistance, and the probability of thermal runaway risk. A weighted summation method is used to combine each factor according to its importance, resulting in a comprehensive health index. The weights are determined based on expert knowledge or data-driven methods; for example, SOH has a weight of 0.4, internal resistance has a weight of 0.3, and thermal runaway risk has a weight of 0.3. The comprehensive health index is normalized and mapped to a pre-defined scoring range (e.g., 0-100 points) to obtain the battery health score. A higher score indicates a better battery health status; a lower score indicates more severe battery degradation. Based on the battery health score and a historical thermal runaway evolution pattern database, the probability of thermal runaway occurring in the current state is calculated using a risk mapping algorithm. The risk mapping algorithm establishes a non-linear mapping relationship between the health score and the probability of thermal runaway, employing logistic regression, support vector machines, or neural network models. The model takes health scores, temperature, voltage, and gas concentration as inputs and outputs a thermal runaway probability value. This probability value is mapped to a preset risk level range to obtain the thermal runaway risk level. The risk level is divided into five levels: normal (probability <1%), low risk (1%-5%), medium risk (5%-20%), high risk (20%-50%), and extremely high risk (>50%). The risk level directly reflects the battery's safety status, providing a basis for early warning decisions.

[0089] This embodiment constructs a dynamic adaptive Bayesian network health assessment model and combines it with multimodal feature fusion results to achieve probabilistic reasoning and quantitative assessment of battery health status. It can adapt to different operating conditions and battery aging states, thereby improving the stability and accuracy of the early warning system.

[0090] Example 8: This embodiment is a further refinement of step S4 in Embodiment 1, specifically illustrating the detailed process of obtaining the battery health score and thermal runaway risk level, including: S4.5.1: Based on the posterior probability distribution output by the Bayesian network model, extract the probability values ​​of each state variable of the battery to obtain the state probability vector.

[0091] In this step, probability values ​​of key state variables are extracted from the posterior probability distribution output by the Bayesian network model. Key state variables include: battery health status (e.g., healthy, sub-healthy, deteriorated, severely deteriorated), SOC (State of Charge), SOH (State of Health), internal resistance, electrolyte state, separator state, positive and negative electrode states, and thermal runaway risk. For discrete state variables, probability values ​​for each state are extracted to form a probability vector; for continuous state variables, the expected value, variance, and quantiles of the probability distribution are extracted. For example, the probability vector for battery health status is P_health = [P(healthy), P(sub-healthy), P(deteriorated), P(severely deteriorated)], and the SOH statistic is [E(SOH), Var(SOH), Q0]. 05 (SOH), Q0. 95 (SOH)]. The probability values ​​of all key state variables are organized into a state probability vector S∈R^m, where m is the total dimension of the state variables. The state probability vector comprehensively reflects the probability distribution characteristics of the battery's current state, providing basic data for calculating the health score.

[0092] S4.5.2: Perform a weighted summation on the state probability vector, with the weights determined based on the degree of influence of each state variable on battery health, to obtain a comprehensive health index.

[0093] In this step, the state probability vectors are weighted and summed to obtain the comprehensive health index. The weight vector W∈R^m is determined based on the degree of influence of each state variable on battery health. This degree of influence is assessed through expert knowledge, Failure Mode and Effects Analysis (FMEA), or data-driven methods. Expert knowledge, based on the battery thermal runaway mechanism and failure modes, determines the importance ranking of each state variable. The FMEA method analyzes the severity, frequency of occurrence, and detection difficulty of each failure mode, calculating the Risk Priority Number (RPN). The higher the RPN, the greater the weight of the state variable corresponding to the failure mode. The data-driven method uses historical failure data to statistically analyze the correlation between each state variable and thermal runaway events, using logistic regression or random forest models to calculate feature importance as the basis for weighting. Combining multiple methods, the weight vector is determined; for example, SOH weight 0.25, internal resistance weight 0.20, electrolyte state weight 0.15, separator state weight 0.15, and thermal runaway risk probability weight 0.25. The dot product operation is performed on the state probability vector and the weight vector to obtain the comprehensive health index HI=S·W. The comprehensive health index takes into account multiple health dimensions of the battery and can comprehensively reflect the overall health status of the battery.

[0094] S4.5.3: Normalize the comprehensive health index and map it to a preset scoring range to obtain the battery health score.

[0095] In this step, the comprehensive health index is normalized and mapped to a preset scoring range. Since the numerical range of the comprehensive health index is not fixed, it needs to be normalized to a unified scoring range for easier understanding and comparison. The preset scoring range is typically 0-100 points, where 100 points indicates the battery is in optimal health and 0 points indicates complete battery failure. Normalization methods include linear and nonlinear mapping. Linear mapping uses the formula: Score = 100 × (HI - HI_min) / (HI_max - HI_min), where HI_min and HI_max are the minimum and maximum values ​​of the comprehensive health index, determined through historical data statistics. Nonlinear mapping uses the Sigmoid function or a piecewise function, which can provide higher resolution in critical intervals. For example, using Sigmoid mapping: Score = 100 / (1 + exp(-k × (HI - HI_0))), where k is the steepness parameter and HI_0 is the center point. After normalization, the battery health score Score ∈ [0, 100] is obtained. The physical meaning of the rating: 90-100 points indicates a healthy battery, 70-90 points indicates a sub-healthy battery, 50-70 points indicates a degraded battery, 30-50 points indicates a severely degraded battery, and 0-30 points indicates a battery nearing failure. The health rating provides an intuitive quantitative indicator for battery management and maintenance decisions.

[0096] S4.5.4: Based on the battery health score and the historical thermal runaway evolution pattern database, calculate the probability of thermal runaway occurring in the current state to obtain the thermal runaway probability value.

[0097] In this step, the probability of thermal runaway occurring in the current state is calculated based on the battery health score and a historical thermal runaway evolution pattern database. The historical thermal runaway evolution pattern database contains a large amount of data on the evolution trajectory of batteries from normal state to thermal runaway, recording the occurrence of thermal runaway under different health scores, temperatures, voltages, gas concentrations, and other characteristics. A mapping model between the health score and the thermal runaway probability is established using data mining and machine learning methods. The mapping model employs logistic regression, support vector machines, random forests, or deep neural networks. Model inputs include features such as the current health score, temperature change rate, voltage drop magnitude, gas release rate, and strain rate increase; the model output is the thermal runaway probability value P_TR∈[0,1]. Model training uses historical data, including normal operation data and thermal runaway event data, and model parameters are optimized through cross-validation to ensure the model's generalization ability. For the current battery state, the health score and related features are input into the mapping model to calculate the thermal runaway probability value. The probability value reflects the likelihood of thermal runaway occurring within a certain time window (e.g., 1 hour, 24 hours) in the current state. To improve prediction accuracy, an ensemble learning approach is used to combine the prediction results of multiple base models and obtain the final probability value through voting or weighted averaging.

[0098] S4.5.5: Based on the thermal runaway probability value, classify the risk level and map the thermal runaway probability value to a preset risk level range to obtain the thermal runaway risk level.

[0099] In this step, the risk level is categorized based on the probability value of thermal runaway, mapping the probability value to a preset risk level range. The risk level categorization is based on risk management theory and safety standards, comprehensively considering the probability of thermal runaway and the severity of its consequences. The preset risk level range includes five levels: Level 0 (Normal): Thermal runaway probability P_TR < 1%, the battery is in a safe state and requires no special intervention; Low risk level (Level 1): 1% ≤ P_TR < 5%, the battery has a slight abnormality, and enhanced monitoring is recommended; Medium risk level (Level 2): ​​5% ≤ P_TR < 20%, the battery shows obvious abnormalities, it is recommended to reduce the load and start active cooling; High-risk level (Level 3): 20% ≤ P_TR < 50%, the battery is in a dangerous state; it is recommended to limit current and enhance cooling. Extremely high risk level (Level 4): P_TR≥50%, the battery is about to thermal runaway, it is recommended to immediately disconnect the power and start the emergency response.

[0100] Based on the calculated thermal runaway probability value, the risk level is determined through interval judgment. For example, if P_TR = 15%, the risk level is medium risk (Level 2). The threshold for classifying the risk level can be adjusted according to the application scenario and safety requirements. For scenarios with high safety requirements (such as manned electric vehicles), the threshold can be appropriately lowered to increase the safety margin; for scenarios with relatively low safety requirements (such as stationary energy storage), the threshold can be appropriately increased to reduce false alarms. After obtaining the thermal runaway risk level, the system transmits the risk level information to the early warning generation module, triggering the corresponding level of early warning signal and control strategy.

[0101] This embodiment transforms the probabilistic inference results of Bayesian networks into intuitive health scores and risk levels through detailed probability calculations and risk mapping processes, providing an actionable decision-making basis for battery safety management.

[0102] Example 9: This embodiment further refines steps S5 and S6 of Embodiment 1. Based on the battery health score and the thermal runaway risk level, a graded early warning signal is obtained through multi-level early warning threshold comparison. Based on the graded early warning signal, the prevention and control strategy knowledge base is queried, and the intervention measures corresponding to the current state are matched to obtain prevention and control suggestions, including: S5.1: Based on the battery health score and the thermal runaway risk level, combined with historical warning data, an adaptive threshold adjustment algorithm is used to set multi-level warning thresholds, obtaining dynamically adjusted multi-level warning thresholds.

[0103] In this step, based on the battery health score and the thermal runaway risk level, combined with historical warning data, an adaptive threshold adjustment algorithm is used to set multi-level warning thresholds. Traditional warning systems use static thresholds and cannot adapt to battery aging and working condition changes, resulting in a high false alarm rate or a high miss alarm rate. The adaptive threshold adjustment algorithm dynamically adjusts the warning threshold according to the battery's usage history, aging status, current working condition, and environmental conditions to achieve a balance between sensitivity and specificity. The algorithm process: First, analyze historical warning data, count the warning accuracy rate, false alarm rate, and miss alarm rate under different threshold settings, draw the ROC curve (Receiver Operating Characteristic curve), and determine the optimal operating point; Second, evaluate the aging degree according to the battery's SOH. For severely aged batteries (SOH < 80%), appropriately lower the warning threshold to improve sensitivity and reduce miss alarms; For new batteries (SOH > 95%), appropriately increase the threshold to reduce false alarms; Third, adjust the threshold according to the current working condition, lower the threshold under high-risk working conditions such as fast charging and high-current discharging, and increase the threshold under static and low-current working conditions; Finally, adjust the threshold according to the environmental temperature, lower the threshold in a high-temperature environment (> 35°C), and increase the threshold in a low-temperature environment (< 0°C). Adaptive adjustment uses fuzzy logic or reinforcement learning methods to establish a threshold adjustment rule base or policy network. Obtain dynamically adjusted multi-level warning thresholds, including: attention threshold T_attention, warning threshold T_warning, emergency threshold T_emergency, danger threshold T_danger. The thresholds are represented in the form of a combination of health scores and risk levels. For example: the attention threshold is (Score < 85 or Risk ≥ Level 1), the warning threshold is (Score < 70 or Risk ≥ Level 2), the emergency threshold is (Score < 50 or Risk ≥ Level 3), and the danger threshold is (Score < 30 or Risk ≥ Level 4).

[0104] S5.2: Compare the battery health score and the thermal runaway risk level with the multi-level warning thresholds to determine the current warning level, obtaining a warning level determination result.

[0105] In this step, compare the current battery health score and the thermal runaway risk level with the dynamically adjusted multi-level warning thresholds to determine the current warning level. The comparison uses logical judgment rules and judges in order of priority from high to low: If Score < T_danger or Risk ≥ Level 4, the warning level is "Danger level"; Otherwise, if Score < T_emergency or Risk ≥ Level 3, the warning level is "Emergency Level"; Otherwise, if Score < T_warning or Risk ≥ Level 2, the warning level is "Warning Level"; Otherwise, if Score < T_attention or Risk ≥ Level 1, the warning level is "Attention Level"; Otherwise, the warning level is "Normal Level".

[0106] To avoid frequent jumps in the warning level, a hysteresis mechanism and time-window smoothing are introduced. The hysteresis mechanism sets different rising and falling thresholds. A lower threshold is used when the warning level rises, and a higher threshold is used when it falls, forming a hysteresis loop. Time-window smoothing averages or median filters the health score and risk level within a certain time window (e.g., 5 minutes) to reduce the impact of instantaneous fluctuations. The warning level determination result is obtained, including the current warning level, trigger time, duration, and confidence level. The confidence level is calculated based on the distance between the health score and the threshold, and the farther the distance, the higher the confidence level.

[0107] S5.3: Generate a warning signal containing risk information and confidence level based on the warning level determination result to obtain the classified warning signal.

[0108] In this step, a detailed warning signal is generated based on the warning level determination result. The warning signal adopts a structured data format and contains the following information: Warning Level: Normal Level, Attention Level, Warning Level, Emergency Level or Danger Level; Trigger Time: The timestamp when the warning is triggered; Duration: The duration for which the warning state persists; Confidence Level: The credibility of the warning determination, with a value ranging from 0 to 1; Health Score: The current battery health score value; Risk Level: The current thermal runaway risk level; Key Features: The key feature values that trigger the warning, such as the temperature peak, voltage drop amplitude, gas concentration, etc.; Abnormal Mode: The type of identified abnormal mode, such as temperature abnormality, voltage abnormality, gas leakage, strain abnormality, etc.; Predicted Trend: The state evolution trend in the future for a certain period (e.g., 1 hour, 24 hours); Suggested Measures: Preliminary response suggestions.

[0109] Warning signals are encoded in JSON, XML, or Protocol Buffers formats for easy transmission and parsing between systems. The transmission method differs depending on the warning level: Attention level signals are recorded via data logs; Warning level signals are delivered via interface prompts and audible alarms; Emergency level signals are sent to relevant personnel via SMS, email, and telephone; Danger level signals trigger emergency response procedures and automatically execute safety measures. Obtaining tiered warning signals provides input for subsequent control strategy matching and execution.

[0110] S5.4: Based on the tiered early warning signal, query the prevention and control strategy knowledge base, retrieve intervention measures that match the current early warning level and risk type, and obtain a set of candidate intervention measures.

[0111] In this step, the prevention and control strategy knowledge base is queried based on the tiered early warning signals to retrieve matching intervention measures. The prevention and control strategy knowledge base is a structured representation of expert knowledge and historical experience, containing intervention measures corresponding to different early warning levels, risk types, and anomaly patterns. The knowledge base is organized using a rule base or decision tree format, with the rule format: IF (Early Warning Level = X) AND (Risk Type = Y) AND (Anomaly Pattern = Z) THEN (Intervention Measures = M). Example of knowledge base content: Alert Level + Temperature Abnormalities: Increase monitoring frequency and check the cooling system; Warning level + abnormal voltage: Reduce charging / discharging power to 70% of rated power and check battery connections; Warning level + gas leak: Activate active cooling, increase ventilation, and check battery seals; Emergency Level + Rapid Temperature Rise: Limit current to 50% of rated current, activate maximum power cooling, and prepare for emergency power outage; Emergency level + multimodal collaborative anomaly: Immediately reduce load, activate emergency cooling, and notify maintenance personnel; Hazard level + any abnormality: Immediately cut off power, activate emergency response, evacuate personnel, and prepare fire-fighting measures.

[0112] The query process employs a pattern matching algorithm. Based on the warning level, risk type, and anomaly pattern in the warning signal, it retrieves all matching rules from the knowledge base to obtain a set of candidate intervention measures. Candidate measures may include multiple options, such as "reduce power," "initiate cooling," "limit current," and "power off." Each intervention measure is accompanied by attribute information, including: measure type, execution parameters (such as power limitation ratio and cooling intensity level), expected effect, execution cost, impact on system performance, and execution priority.

[0113] S5.5: Prioritize and assess the feasibility of the candidate intervention set, select the most suitable intervention for the current state, and obtain the prevention and control recommendations.

[0114] In this step, the candidate intervention set is prioritized and its feasibility is assessed to select the optimal intervention. Prioritization considers multiple factors: Effectiveness: The effectiveness of intervention measures in reducing the risk of thermal runaway is evaluated through historical data statistics or simulation experiments; Urgency: The degree of urgency of the current risk level; the higher the risk level, the higher the priority. Feasibility: Whether the conditions for implementing the intervention measures are met, such as whether the cooling system is available and whether a power outage is permissible; Cost: The cost of implementing intervention measures, including energy consumption, equipment wear and tear, and impact on user experience; Side effects: The intervention may have negative consequences, such as reduced power affecting system performance, and power outages causing data loss.

[0115] Multi-attribute decision-making methods (such as Analytic Hierarchy Process (AHP), TOPSIS, or multi-objective optimization) are used to comprehensively evaluate candidate measures, calculating a comprehensive score for each measure. The comprehensive score is calculated as follows: Comprehensive Score = w1 × Effectiveness + w2 × Urgency + w3 × Feasibility - w4 × Cost - w5 × Side Effects, where w1-w5 are weighting coefficients determined based on the application scenario. The intervention measure with the highest comprehensive score is selected as the recommended measure. For hazard-level warnings, the most effective measure is prioritized, even if the cost is high or the side effects are significant; for attention-level or warning-level warnings, measures with low cost and minimal side effects are prioritized. A feasibility assessment examines the implementation conditions of the intervention measure, such as: Cooling measures: Check the status of the cooling system, and whether the coolant temperature and flow rate are normal; Power limiting measures: Check the current load demand to determine if a power reduction is permissible; Power outage measures: Check whether there are any critical tasks and whether a power outage would have serious consequences.

[0116] If the optimal measure is not feasible, the second-best measure is selected. Prevention and control recommendations are obtained and output in structured data format, including: Recommended measures: Name and description of the intervention; Execution parameters: Specific execution parameters, such as "limit charging power to 50% of rated power" and "start cooling system to maximum level"; Execution priority: execute immediately, execute as soon as possible, or execute according to plan; Expected results: The expected changes in health scores and a reduction in risk level after implementation; Alternative solutions: If the recommended measures cannot be implemented, alternative measures are provided; Implementation Guidelines: Detailed implementation steps and precautions.

[0117] Preventive control recommendations are displayed to operators via a human-machine interface or sent to the battery management system (BMS) for execution via an automated control interface. For automatically executed measures, the system records an execution log, including execution time, results, and effectiveness evaluation, providing data for subsequent system optimization.

[0118] This embodiment achieves differentiated responses to different risk levels through a multi-level early warning mechanism and intelligent control strategy matching, which ensures safety, avoids excessive intervention, and improves the system's usability and user acceptance.

[0119] Example 10: This embodiment is a further extension of embodiment 9, and the method further includes: Step A: Based on the battery state change data after the implementation of the prevention and control recommendations, calculate the early warning accuracy and intervention effectiveness to obtain the early warning effect evaluation results.

[0120] In this step, after the implementation of the prevention and control recommendations, the battery state changes are continuously monitored, and multimodal sensor data are collected after implementation, including the temporal changes of parameters such as temperature, voltage, gas concentration, and strain pressure. By comparing the state data before and after implementation, the actual effect of the intervention measures is evaluated. The calculation of the early warning accuracy includes: statistically analyzing all early warning events within a certain period (e.g., one month), recording the number of true positives (TP, correctly warning of thermal runaway risk), false positives (FP, false alarms), true negatives (TN, correctly judging the normal state), and false negatives (FN, missed detections), and calculating the accuracy = (TP+TN) / (TP+TN+FP+FN), precision = TP / (TP+FP), recall = TP / (TP+FN), and F1 score F1 = 2 × Precision × Recall / (Precision+Recall). The calculation of intervention effectiveness includes: for each intervention, comparing the changes in health scores and risk levels before and after implementation; if the health score increases or the risk level decreases, the intervention is considered effective; the ratio of effective interventions to total interventions is calculated, yielding the intervention effectiveness rate: Effectiveness = N_effective / N_total. Further analysis of the effectiveness of different interventions is conducted, calculating the average effect of each measure, such as the "power reduction" measure increasing the health score by an average of 5 points, and the "cooling activation" measure decreasing the temperature by an average of 3°C. For cases of failed early warnings (false positives or false negatives), root cause analysis is performed to identify the reasons for failure, such as sensor malfunction, inaccurate model parameters, or improper threshold settings. The early warning effectiveness evaluation results are obtained, including various performance indicators, failure case analysis, and improvement suggestions, providing a basis for system optimization.

[0121] Step B: Based on the evaluation results of the early warning effect, update the multi-level early warning thresholds and intervention strategy priorities using reinforcement learning methods to obtain the system optimization parameters.

[0122] In this step, based on the evaluation results of the early warning effect, reinforcement learning is used to optimize system parameters and achieve closed-loop adaptive optimization. Reinforcement learning models the early warning system as a Markov Decision Process (MDP). The state space includes battery health scores, risk levels, sensor data, etc., while the action space includes adjusting the early warning threshold and selecting intervention measures. The reward function is designed based on early warning accuracy and intervention effectiveness. The reward function is in the form: R = α × (TP - FP) - β × FN + γ × Effectiveness, where α, β, and γ are weight coefficients, encouraging the improvement of true positives and intervention effectiveness, and penalizing false positives and false negatives. Q-learning, SARSA, or deep reinforcement learning (DQN, A3C) algorithms are used to learn the optimal strategy through interaction with the environment. The learning process involves the system collecting state-action-reward data during actual operation, updating the Q-value function or policy network parameters, and gradually optimizing the early warning threshold and intervention strategy. Specific optimization content includes: Multi-level warning threshold optimization: Based on historical warning effects, adjust the thresholds for attention, warning, emergency, and danger levels to maximize the AUC (area under the curve) of the ROC curve; Intervention strategy prioritization optimization: Adjust the priority weights of each intervention measure based on its actual effectiveness, so that effective measures receive higher priority; Adaptive adjustment rule optimization: Optimize the rule parameters for adjusting thresholds based on SOH, operating conditions, and ambient temperature to improve adaptability; Feature fusion weight optimization: Adjust the weight coefficients of feature fusion according to the actual contribution of each modality feature in the early warning.

[0123] To ensure system stability, a gradual update strategy is adopted, with limited parameter changes in each update to avoid drastic fluctuations. A / B testing is used to trial the optimized parameters on a subset of batteries, comparing performance before and after optimization to verify the effectiveness before full rollout. The resulting optimized system parameters include updated warning thresholds, intervention strategy weights, and adaptive adjustment rules. These optimized parameters are saved and applied to the next round of warnings, enabling continuous improvement. Through this closed-loop optimization mechanism, the system can continuously learn and adapt; as runtime increases, warning accuracy gradually improves, false alarm rate gradually decreases, and system performance is continuously optimized.

[0124] This embodiment forms a closed-loop adaptive system through early warning effect evaluation and reinforcement learning optimization, which can continuously improve based on actual operating data, thereby improving the long-term stability and reliability of the early warning system.

[0125] Example 11: This embodiment is a further extension of Embodiment 1, and the method further includes: Step A: Analyze the data flow characteristics and processing requirements of the multi-source heterogeneous dataset, allocate edge computing resources to various types of sensor data, and obtain a resource allocation scheme.

[0126] This step analyzes the data flow characteristics of the multi-source heterogeneous dataset, including data generation rate, data volume, data type, and real-time requirements. Temperature data has a low generation rate (1-10Hz), small data volume, and medium real-time requirements; voltage data has a high generation rate (10-100Hz), medium data volume, and high real-time requirements; gas concentration data has a very low generation rate (0.1-1Hz), small data volume, and low real-time requirements; strain and pressure data have a medium generation rate (1-10Hz), small data volume, and medium real-time requirements. The data processing requirements are analyzed, including the computational complexity and latency requirements of tasks such as time synchronization, spatial alignment, noise reduction, and feature extraction. Time synchronization and spatial alignment have low computational complexity but require real-time processing; noise reduction (low-rank approximation of the Hankel matrix) has medium computational complexity and high latency requirements; feature extraction (deep learning network) has high computational complexity and medium latency requirements. Based on the data flow characteristics and processing requirements, an edge computing resource allocation scheme is designed. Edge computing nodes are deployed close to the sensors, such as inside or near the battery pack, to reduce data transmission latency. Resource allocation adopts a layered architecture: First layer (sensor layer): Sensor nodes are responsible for data acquisition and preliminary quality assessment, and are configured with low-power microcontrollers (MCUs); The second layer (edge ​​node layer): Edge computing nodes are responsible for time synchronization, spatial alignment, noise reduction, and preliminary feature extraction, and are configured with embedded processors (such as ARM Cortex-A series) or edge AI chips (such as NVIDIA Jetson, Google Edge TPU); The third layer (cloud layer): cloud servers are responsible for complex feature fusion, model training and knowledge base updates, and are configured with high-performance GPU servers.

[0127] Edge computing resources are allocated to various sensor data: temperature and strain data are assigned to edge node 1, voltage data to edge node 2, and gas data to edge node 3. Resource allocation considers load balancing to avoid overloading any single node. The resulting resource allocation scheme includes the hardware configuration of each edge node, processing task allocation, data flow routing, and communication protocols.

[0128] Step B: Configure edge computing nodes based on the resource allocation scheme, establish an edge computing node network, and perform time synchronization, spatial location correspondence processing, and noise reduction processing on the edge computing node network.

[0129] In this step, edge computing nodes are configured based on a resource allocation scheme. Configuration includes: installing an operating system (such as Linux or FreeRTOS), deploying data processing software (such as a time synchronization module, noise reduction algorithm library, and feature extraction model), configuring network communication (such as Ethernet, Wi-Fi, and CAN bus), and setting security mechanisms (such as data encryption and access control). An edge computing node network is established, with nodes interconnected via a high-speed network to form a distributed computing system. The network topology uses a star or mesh structure to ensure reliability and fault tolerance. Data processing tasks are then executed on the edge computing node network. Time synchronization: Each edge node synchronizes with the master clock via the PTP protocol to achieve microsecond-level time alignment; each node timestamps the received sensor data to ensure time consistency of multi-source data; Spatial location mapping: Each edge node establishes a mapping relationship between data and spatial location based on sensor deployment information; spatial mapping information is exchanged between nodes to construct a global spatiotemporal data model; Noise reduction: Each edge node applies the Hankel matrix low-rank approximation algorithm to the local sensor data for noise reduction; the noise reduction algorithm is executed efficiently on the embedded processor or AI chip of the edge node, and the processing latency is controlled in milliseconds; the noise-reduced data is shared among the edge nodes to provide input for subsequent feature extraction and fusion.

[0130] Through edge computing architecture, most data processing tasks are completed at edge nodes, with only processed feature data or anomaly events uploaded to the cloud. This significantly reduces communication bandwidth requirements (by more than 80%) and end-to-end latency (to 1 / 10 of the original). Edge computing nodes possess local decision-making capabilities and can still perform basic early warning functions in the event of network outages, improving system reliability and robustness. Edge nodes adopt a low-power design, supporting battery power or energy harvesting power, making them suitable for distributed battery systems.

[0131] This embodiment achieves distributed and localized data processing through an edge computing architecture, solving the problems of large response latency and high communication bandwidth requirements of centralized processing architecture, and significantly improving the real-time performance and reliability of the system.

[0132] Example 12: This embodiment provides a multimodal fusion lithium iron phosphate battery thermal runaway early warning system, corresponding to the method described in Embodiment 1, including: The data acquisition module is used to collect temperature data, voltage data, gas concentration data, and shell strain and pressure data in real time through a heterogeneous sensor network deployed on the surface and inside the battery, so as to obtain a multi-source heterogeneous dataset.

[0133] The data acquisition module comprises a heterogeneous sensor network and a data acquisition unit. The heterogeneous sensor network includes a temperature sensor array (at least 8 temperature measurement points, using K-type thermocouples or Pt100, temperature range -40℃ to 200℃, accuracy ±0.5℃), a voltage sensor (high-precision voltage acquisition chip, measurement range 0-5V, accuracy ±1mV), gas sensors (CO, CO2, H2 sensors, detection range 0-1000ppm, response time <30 seconds), and strain pressure sensors (thin-film pressure sensors or fiber Bragg gratings, measurement range 0-10MPa, accuracy ±0.1%). The data acquisition unit includes signal conditioning circuitry, a high-precision ADC of 16 bits or more, a FIFO buffer, and a communication interface (CAN bus, Modbus, or Ethernet). The data acquisition module enables multi-channel synchronous acquisition, with the sampling frequency set according to the sensor type. It also performs noise assessment and quality labeling, outputting a multi-source heterogeneous dataset with quality labels.

[0134] The data preprocessing module is used to perform time synchronization and spatial location correspondence processing on the multi-source heterogeneous dataset, and to perform noise reduction and temporal feature extraction using the sublinear time low-rank approximation algorithm of the Hankel matrix to obtain preprocessed multimodal feature data.

[0135] The data preprocessing module is deployed on edge computing nodes and includes a time synchronization unit, a spatial mapping unit, a denoising unit, and a feature extraction unit. The time synchronization unit uses the PTP protocol to achieve microsecond-level time alignment and resamples data at different sampling frequencies. The spatial mapping unit establishes a mapping relationship between data and spatial location based on sensor deployment information, constructing a spatiotemporally aligned data model. The denoising unit implements a sublinear time-low-rank approximation algorithm for the Hankel matrix, with a time complexity of O(N log N), enabling efficient execution on embedded processors or edge AI chips. The feature extraction unit applies wavelet transform and empirical mode decomposition to extract multi-scale temporal features, outputting preprocessed multimodal feature data.

[0136] The feature fusion module is used to extract single-modal deep features from the preprocessed multimodal feature data and construct a cross-modal feature association network using a quadratic time algorithm for the maximum weight sparse subgraph problem to obtain a fused feature vector.

[0137] The feature fusion module includes a single-modal feature extraction unit and a cross-modal fusion unit. The single-modal feature extraction unit comprises four dedicated networks: a thermal feature extraction network (1D-CNN architecture, processing temperature data), an electrical property analysis network (LSTM architecture, processing voltage data), a chemical fingerprint recognition model (multi-task learning framework, processing gas data), and a stress analysis model (physically constrained neural network, processing strain data). The cross-modal fusion unit implements the maximum weight sparse subgraph algorithm with a time complexity of O(m). 2The algorithm identifies the most informative correlation structure between modalities, calculates feature importance weights, integrates multimodal features using an adaptive fusion algorithm, and outputs a low-dimensional, high-information fused feature vector through PCA or autoencoder dimensionality reduction.

[0138] The health assessment module is used to input the fused feature vector into the Bayesian network health assessment model for probabilistic inference calculation to obtain the battery health score and thermal runaway risk level.

[0139] The health assessment module comprises a Bayesian network construction unit, a probabilistic inference unit, and a scoring unit. The Bayesian network construction unit uses PC and hill-climbing algorithms to learn the network topology, calculates conditional probability distribution parameters using maximum likelihood estimation, and establishes a parameterized Bayesian network model. The probabilistic inference unit uses the belief propagation algorithm to calculate the posterior probability distribution of hidden nodes and outputs a state probability vector. The scoring unit performs a weighted summation of the state probability vectors, normalizes and maps them to a 0-100 score range to obtain the health score; based on the health score and a historical thermal runaway evolution pattern database, it calculates the thermal runaway probability using logistic regression or a neural network model, mapping it to a five-level risk level (normal, low risk, medium risk, high risk, and extremely high risk).

[0140] The early warning generation module is used to obtain graded early warning signals based on the battery health score and the thermal runaway risk level by comparing multiple early warning thresholds.

[0141] The early warning generation module includes a threshold adjustment unit, an early warning determination unit, and a signal generation unit. The threshold adjustment unit employs an adaptive threshold adjustment algorithm to dynamically adjust multi-level early warning thresholds (Caution, Warning, Emergency, Danger) based on battery SOH, current operating conditions, and ambient temperature. The early warning determination unit compares the health score and risk level with the thresholds, using a lag mechanism and time window smoothing to determine the current early warning level. The signal generation unit generates structured early warning signals, including information such as early warning level, trigger time, confidence level, key features, anomaly patterns, and predicted trends, outputting them in JSON or XML format.

[0142] The control strategy module is used to query the prevention and control strategy knowledge base based on the graded early warning signal, match the intervention measures corresponding to the current state, and obtain prevention and control suggestions.

[0143] The control strategy module includes a knowledge base query unit, a measure evaluation unit, and a suggestion generation unit. The knowledge base query unit retrieves matching intervention measures from the prevention and control strategy knowledge base based on the warning level, risk type, and anomaly pattern, obtaining a set of candidate measures. The measure evaluation unit uses multi-attribute decision-making methods (AHP, TOPSIS) to comprehensively evaluate the effectiveness, urgency, feasibility, cost, and side effects of each measure, and prioritizes them. The suggestion generation unit selects the optimal measure and generates structured control suggestions, including recommended measures, execution parameters, execution priority, expected effects, alternatives, and execution guidelines, which are output through a human-machine interface or an automatic control interface.

[0144] The system in this embodiment also includes an effect evaluation module and a parameter optimization module. The effect evaluation module monitors changes in battery status after the intervention measures are implemented, calculates the early warning accuracy and intervention effectiveness, and performs root cause analysis on failure cases. The parameter optimization module uses reinforcement learning to update the early warning threshold, intervention strategy priority, and feature fusion weights based on the evaluation results, achieving closed-loop adaptive optimization.

[0145] The system in this embodiment adopts a modular design, with each module functioning independently and featuring standardized interfaces, facilitating system integration and functional expansion. The system supports distributed deployment, with data acquisition and preprocessing performed on edge nodes, feature fusion and health assessment on edge nodes or in the cloud, and early warning generation and control strategies executed locally or remotely, flexibly adapting to different application scenarios. The system boasts high reliability, employing redundant design and fault detection mechanisms, with key modules supporting hot backup to ensure continuous system operation. The system is scalable, supporting the addition of new sensor types, feature extraction algorithms, and intervention measures, with functional expansion achieved through configuration files or plugin mechanisms.

[0146] The system in this embodiment realizes comprehensive monitoring, accurate early warning and intelligent intervention for thermal runaway of lithium iron phosphate batteries. It solves the technical problems of limited accuracy of monitoring a single physical quantity and response delay of centralized processing, and significantly improves the safety and reliability of the battery system. It is suitable for various application scenarios such as electric vehicles, energy storage systems and electric ships.

[0147] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A multimodal fusion method for early warning of thermal runaway in lithium iron phosphate batteries, characterized in that, Includes the following steps: By deploying a heterogeneous sensor network on and inside the battery, temperature data, voltage data, gas concentration data, and casing strain and pressure data are collected in real time to obtain a multi-source heterogeneous dataset. The multi-source heterogeneous dataset is processed for time synchronization and spatial location correspondence, and the sublinear time low-rank approximation algorithm of Hankel matrix is ​​used for noise reduction and temporal feature extraction to obtain preprocessed multimodal feature data. The preprocessed multimodal feature data is subjected to single-modal deep feature extraction, and a cross-modal feature association network is constructed using a quadratic time algorithm for the maximum weight sparse subgraph problem to obtain a fused feature vector. The fused feature vector is input into a Bayesian network health assessment model for probabilistic inference calculation to obtain a battery health score and thermal runaway risk level. Based on the battery health score and the thermal runaway risk level, a graded early warning signal is obtained by comparing multiple early warning thresholds; Based on the tiered early warning signal, the knowledge base of prevention and control strategies is queried, and the intervention measures corresponding to the current state are matched to obtain prevention and control suggestions.

2. The method according to claim 1, characterized in that, By deploying a heterogeneous sensor network on and inside the battery, real-time data such as temperature, voltage, gas concentration, and casing strain and pressure are collected to obtain a multi-source heterogeneous dataset, specifically including: Temperature sensor arrays, voltage sensors, gas sensors, and strain pressure sensors are deployed at key locations on the battery surface and inside to obtain a configured heterogeneous sensor network. The heterogeneous sensor network is used to synchronously acquire signals from various sensors to obtain the original multi-source heterogeneous data stream; The original multi-source heterogeneous data stream is subjected to noise assessment and signal quality index calculation, and outliers are filtered out to obtain the multi-source heterogeneous dataset.

3. The method according to claim 1, characterized in that, The time synchronization and spatial location mapping of the multi-source heterogeneous dataset specifically includes: The data from different types of sensors in the multi-source heterogeneous dataset are timestamped to obtain a time-synchronized dataset. Spatial mapping processing is performed on the time-synchronized dataset to establish the correspondence between different sensor data and battery spatial location, resulting in a spatiotemporally aligned multi-source heterogeneous dataset. The sublinear time low-rank approximation algorithm of the Hankel matrix is ​​used for noise reduction and temporal feature extraction to obtain preprocessed multimodal feature data, specifically including: Hankel matrix representations were constructed on spatiotemporally aligned multi-source heterogeneous datasets to obtain the matrix form of time-series data. The Hankel matrix is ​​decomposed using a sublinear time low-rank approximation algorithm to obtain denoised structured data. The structured data after noise reduction is subjected to a multi-scale analysis method to extract signal features at different time scales, thereby obtaining the preprocessed multimodal feature data.

4. The method according to claim 1, characterized in that, Single-modal deep feature extraction is performed on the preprocessed multimodal feature data, including: A thermal feature extraction network is applied to the temperature data in the preprocessed multimodal feature data to obtain deep temperature modal features; An electrical characteristic analysis network is applied to the voltage data in the preprocessed multimodal feature data to obtain deep voltage mode features; A chemical fingerprinting model is applied to the gas concentration data in the preprocessed multimodal feature data to obtain deep gas modal features; Applying a stress analysis model to the strain and pressure data in the preprocessed multimodal feature data, deep features of the strain modes are obtained; A quadratic time algorithm for the maximum weight sparse subgraph problem is used to construct a cross-modal feature association network, resulting in a fused feature vector, including: The deep features of the single mode are processed by feature vectorization, and the deep features of temperature mode, voltage mode, gas mode and strain mode are converted into feature vectors of the same dimension to obtain the single mode feature vector set; The correlation metric is calculated between different modal feature vectors in the single-modal feature vector set to obtain the cross-modal correlation matrix; A quadratic time algorithm for the maximum weight sparse subgraph problem is applied to the cross-modal correlation matrix to identify the most informative association structure between different modal features, resulting in a weighted cross-modal feature association graph. Based on the weighted cross-modal feature association graph, the contribution of different features in battery abnormal state identification is calculated by combining historical data, and the feature importance weight vector is obtained. Based on the cross-modal feature association graph and the feature importance weight vector, an adaptive fusion algorithm is used to weight and integrate the multimodal features, and the feature space is optimized by dimensionality reduction to obtain the fused feature vector.

5. The method according to claim 1, characterized in that, The fused feature vector is input into the Bayesian network health assessment model for probabilistic inference calculation, including: Based on the fused feature vectors and historical battery operation data, a structural learning algorithm is used to determine the topology of the Bayesian network, resulting in a Bayesian network topology model. Based on the Bayesian network topology model and historical operating data, the conditional probability distribution parameters of each node in the network are calculated using the maximum likelihood estimation method, resulting in a parameterized Bayesian network model. The fused feature vector is used as observational evidence and input into the parameterized Bayesian network model to obtain the input state representation; A probabilistic reasoning algorithm is applied to the input state representation to calculate the posterior probability distribution of the current state of the battery, and the state probability distribution result is obtained. Based on the state probability distribution results, a quantitative calculation is performed using a preset health scoring function to obtain the battery health score and thermal runaway risk level.

6. The method according to claim 1, characterized in that, The battery health score and thermal runaway risk level are obtained, including: Based on the posterior probability distribution output by the Bayesian network model, the probability values ​​of each state variable of the battery are extracted to obtain the state probability vector. The state probability vectors are weighted and summed, with the weights determined based on the degree of influence of each state variable on battery health, to obtain a comprehensive health index. The comprehensive health index is normalized and mapped to a preset scoring range to obtain the battery health score. Based on the battery health score and the historical thermal runaway evolution pattern database, the probability of thermal runaway occurring in the current state is calculated to obtain the thermal runaway probability value. Based on the thermal runaway probability value, a risk level is determined by mapping the thermal runaway probability value to a preset risk level range.

7. The method according to claim 1, characterized in that, Based on the battery health score and the thermal runaway risk level, a graded warning signal is obtained by comparing multiple warning thresholds, specifically including: Based on the battery health score and the thermal runaway risk level, combined with historical early warning data, an adaptive threshold adjustment algorithm is used to set multi-level early warning thresholds to obtain dynamically adjusted multi-level early warning thresholds. The battery health score and the thermal runaway risk level are compared with the multi-level warning threshold to determine the current warning level and obtain the warning level determination result. Based on the warning level determination result, a warning signal containing risk information and confidence level is generated, thus obtaining the graded warning signal; Based on the tiered early warning signals, the prevention and control strategy knowledge base is queried, and the intervention measures corresponding to the current state are matched to obtain prevention and control suggestions, including: Based on the tiered early warning signal, query the prevention and control strategy knowledge base, retrieve intervention measures that match the current early warning level and risk type, and obtain a set of candidate intervention measures; The candidate intervention set is prioritized and its feasibility is assessed. The intervention most suitable for the current situation is selected to obtain the prevention and control recommendations.

8. The method according to claim 1, characterized in that, The method further includes: Based on the battery state change data after the implementation of the aforementioned prevention and control recommendations, the early warning accuracy and intervention effectiveness are calculated to obtain the early warning effect evaluation results. Based on the evaluation results of the early warning effect, the multi-level early warning thresholds and intervention strategy priorities are updated using reinforcement learning methods to obtain the system optimization parameters.

9. The method according to claim 1, characterized in that, The method further includes: Analyze the data flow characteristics and processing requirements of the multi-source heterogeneous dataset, allocate edge computing resources to various types of sensor data, and obtain a resource allocation scheme. Based on the resource allocation scheme, configure edge computing nodes, establish an edge computing node network, and perform time synchronization, spatial location correspondence processing, and noise reduction processing on the edge computing node network.

10. A multimodal fusion lithium iron phosphate battery thermal runaway early warning system, characterized in that, include: The data acquisition module is used to collect temperature data, voltage data, gas concentration data, and shell strain and pressure data in real time through a heterogeneous sensor network deployed on the surface and inside the battery, so as to obtain a multi-source heterogeneous dataset. The data preprocessing module is used to perform time synchronization and spatial location correspondence processing on the multi-source heterogeneous dataset, and to perform noise reduction and temporal feature extraction using the sublinear time low-rank approximation algorithm of the Hankel matrix to obtain preprocessed multimodal feature data. The feature fusion module is used to extract single-modal deep features from the preprocessed multimodal feature data, and to construct a cross-modal feature association network using a quadratic time algorithm for the maximum weight sparse subgraph problem to obtain a fused feature vector. The health assessment module is used to input the fused feature vector into the Bayesian network health assessment model for probabilistic inference calculation to obtain the battery health score and thermal runaway risk level. The early warning generation module is used to obtain graded early warning signals based on the battery health score and the thermal runaway risk level by comparing multiple early warning thresholds; The control strategy module is used to query the prevention and control strategy knowledge base based on the graded early warning signal, match the intervention measures corresponding to the current state, and obtain prevention and control suggestions.

Images