Lithium battery early thermal runaway early warning management system based on temperature detection

Abstract

The invention discloses a lithium battery early-stage thermal runaway early warning management system based on temperature detection, and relates to the technical field of lithium battery thermal runaway management, and the system comprises a data collection module which collects fire related data in real time through a sensor array; the data processing module is used for analyzing the collected data flow in real time; the disaster prediction module is used for inputting the processed data into a model for prediction; the grading early warning module performs self-adaptive grading on the disaster situation according to the predicted data trend in combination with the comprehensive disaster situation index; the task generation module generates a corresponding task chain according to the grading result; the fire extinguishing execution module starts a liquid nitrogen fire extinguishing unit according to the execution command; the disaster tracing module is responsible for recording operation logs of the whole system module. The technical problems that an existing thermal runaway management system of the lithium battery cabin cannot perform early warning on early-stage thermal runaway, and fire spreading cannot be blocked in time when thermal runaway occurs can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery thermal runaway management, and particularly relates to a lithium battery early thermal runaway early warning management system based on temperature detection. BACKGROUND

[0002] A lithium battery cabin, also known as an energy storage prefabricated cabin, is a large fixed energy storage solution that has rapidly emerged with the large-scale development of renewable energy and the demand for smart grid construction. Its background of birth is mainly due to the inherent intermittency and instability of new energy such as wind energy and solar energy. They are difficult to directly match the requirements of stable operation of the power grid, thus giving rise to the urgent need for large-scale and high-efficiency electric energy storage technology. Its core function is to realize the "spatial and temporal translation" of electric energy - when the power generation is greater than the power consumption, the surplus power is stored; when the power consumption is high or the power generation is insufficient, the stored electric energy is released back to the power grid, thereby effectively performing important tasks such as "peak clipping and valley filling", smoothing new energy output, providing emergency backup power, and adjusting the frequency of the power grid.

[0003] When the lithium battery is overcharged, overdischarged, internally short-circuited, or mechanically damaged, it will trigger an internal chain heat release reaction, causing a rapid rise in temperature and pressure. This process is called "thermal runaway". Once a single cell experiences thermal runaway, the huge amount of heat released will quickly spread to adjacent batteries like a domino effect, triggering "thermal runaway spread" of the entire battery module or battery cabin, thereby instantly exploding a large amount of high-temperature, toxic and flammable gas accompanied by a fire. This fire develops rapidly and is extremely destructive. More seriously, lithium battery fires are essentially internal chemical reactions, and traditional water-based extinguishing agents cannot penetrate the battery interior to effectively cool it, making it prone to "rekindling" - after the fire is extinguished, the residual heat in the battery may again trigger a fire. Therefore, lithium battery cabins often have a fire extinguishing system that can quickly extinguish flames and prevent the occurrence of catastrophic chain reactions through active intervention, ensuring the safety of the entire energy storage facility and surrounding personnel and property.

[0004] However, the existing lithium battery thermal runaway management system has multiple defects. In terms of monitoring and early warning, the system relies heavily on the lagging changes in conventional parameters such as voltage and temperature for judgment, lacks effective in-situ detection capabilities for the core precursors of thermal runaway such as irreversible micro-short circuits and lithium dendrite growth in the battery, resulting in a very short warning window period, and often the internal chain reaction cannot be reversed when the system alarms. In terms of thermal spread barrier, the mainstream fire extinguishing system usually adopts a full submersion extinguishing strategy, which can extinguish the fire but cannot effectively penetrate the battery interior for deep cooling, resulting in frequent rekindling of the extinguished battery due to internal residual high temperature. Therefore, the existing lithium battery cabin thermal runaway management effect is not ideal. SUMMARY

[0005] The application provides a lithium battery early thermal runaway early warning management system based on temperature detection to solve the technical problems that the existing thermal runaway management system of the lithium battery cabin cannot early warn early thermal runaway and cannot timely block the spread of fire when thermal runaway occurs.

[0006] To solve the above technical problems, the application provides the following technical solutions: The application provides a lithium battery early thermal runaway early warning management system based on temperature detection, comprising: The data acquisition module: real-time acquisition of fire-related data through a sensor array, and encrypted data transmission to the cloud through the Internet of Things; The data processing module: for real-time analysis of the collected data stream, efficient extraction of effective information in the data stream through data processing, and improvement of the operation efficiency of the overall system; The disaster situation prediction module: for inputting the processed data into a model for prediction, and generating a prediction report of future disaster situations according to the output results; The hierarchical early warning module: self-adaptive grade division of the disaster situation according to the predicted data trend and the comprehensive disaster situation index; The task generation module: generating a corresponding task chain according to the grade division results, and converting the task chain into an execution command; The fire extinguishing execution module: starting the liquid nitrogen fire extinguishing unit according to the execution command to quickly extinguish the fire; The disaster situation tracing module: responsible for recording the operation log of the entire system module, and analyzing the causes of the disaster situation through analysis of the log data.

[0007] The technical solutions provided by the application have at least the following beneficial effects: The application can extract effective data information from multi-source data by collecting numerical data of the lithium battery cabin and image data in the lithium battery cabin and performing fusion processing, improve the operation efficiency and accuracy of the overall system, and facilitate subsequent multi-angle analysis of the operation state of the lithium battery cabin.

[0008] The hierarchical early warning module can self-adaptively divide the disaster situation according to the real-time collected data and the trend of the predicted disaster situation, accurately improve the management efficiency of the early thermal runaway of the lithium battery, and generate a corresponding task chain through the task generation module, which is convenient for further processing.

[0009] The fire extinguishing execution module can adopt different fire extinguishing measures according to the fire situation, and the disaster situation tracing module can facilitate subsequent analysis of the causes of the disaster situation, improving the integrity of the system. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort based on these drawings.

[0011] Figure 1 is the system diagram of the early thermal runaway early warning management system of lithium battery based on temperature detection provided by the embodiments of the present application.

[0012] Figure 2 is the overall structure schematic diagram of the liquid nitrogen fire extinguishing unit provided by the embodiments of the present application.

[0013] Figure 3 is the structure schematic diagram of the liquid nitrogen spray head and control assembly part provided by the embodiments of the present application.

[0014] In the figure: 1, bottom plate; 2, liquid nitrogen tank; 3, liquid nitrogen recovery device; 4, liquid outlet pipe; 5, liquid inlet pipe; 6, high-pressure nitrogen gas cylinder; 7, manifold pipe; 8, electromagnetic valve; 9, pressure reducing valve; 10, valve assembly; 11, exhaust port; 12, pressure gauge; 13, liquid nitrogen spray head; 14, control assembly. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present application more clear, the following will further describe the embodiments of the present application in combination with the drawings.

[0016] Please refer to Figures 1-3 is a kind of early warning management system of lithium battery early thermal runaway based on temperature detection provided by the embodiments.

[0017] I. Data acquisition module Sensor array includes numerical sensor and image sensor, numerical sensor includes temperature and humidity sensor and gas sensor, image sensor includes thermal imaging sensor and visible light sensor; It should be noted that the sensor array is redundantly deployed, and a plurality of nodes are spatially redundantly arranged in a grid manner in a single monitoring area, and each sensor node is provided with a sensor array.

[0018] The encryption process of the numerical sensor data acquisition is to fuse a plurality of sampling period numerical data, its corresponding time stamp and device identifier into a data packet, encrypt the entire data packet and generate a message authentication code; The encryption process of the image sensor data acquisition is to take two consecutive image data as an independent load, only encrypt the load part of the image data, and retain the image frame header information for network transmission scheduling.

[0019] It should be noted that the encryption mode for data acquired by numerical sensors and image sensors is based on the national cryptographic algorithm SM4.

[0020] II. Data Processing Module The data processing module performs semantic understanding on the high-dimensional visual data of the image data stream by calling a pre-built lightweight multimodal large model, and compresses the understood high-dimensional visual data into a low-dimensional semantic token sequence, thereby achieving intelligent compression of image data. It should be noted that the lightweight multimodal large model is a neural network model with fewer than 100 million parameters, built using knowledge distillation and model pruning techniques. It employs a variant of the Visual Transformer (ViT) as the visual encoder. By segmenting the input image into image patches and extracting visual features, a linear projection layer maps the high-dimensional visual feature vectors output by the visual encoder to the same low-dimensional semantic space as the text feature vectors, achieving cross-modal feature alignment. The model is trained using a fire safety domain dataset consisting of infrared thermal images, visible light images, and corresponding text labels for supervised fine-tuning. Knowledge distillation is then used, with the output of the LlaVA multimodal large model serving as soft labels to guide the training of the lightweight model.

[0021] The data processing module uses the Kalman filter algorithm to optimize the collected numerical data, reducing fluctuations and errors caused by environmental noise and electromagnetic interference. The data processing unit maps the standardized point set representation of the numerical sensor and the semantic token sequence of the visual data to a unified feature space through an asymmetric dense projection network, and generates a deeply fused multimodal feature representation through a gated fusion mechanism. It should be noted that the symmetric dense projection network includes a numerical feature projection path and a visual feature projection path. The numerical feature projection path adopts a wide and shallow fully connected layer structure to project the standardized point set representation of the numerical sensor onto a high-dimensional feature space. The visual feature projection path adopts a deep and narrow neural network structure to project the semantic token sequence of the visual data onto a feature space of the same dimension as the numerical features. The numerical feature projection path has a larger number of parameters than the visual feature projection path, forming an asymmetric projection structure.

[0022] The gating fusion mechanism includes computational fusion gating, weighted fusion execution, and residual connection. The computational fusion gating steps are as follows: based on the stitched numerical enhancement features and visual enhancement features, the fusion weight of each feature dimension is learned. The weighted fusion execution steps are as follows: the numerical enhancement features and visual enhancement features are dynamically weighted and summed according to the fusion weights. The residual connection mechanism steps are as follows: the original projection features and the fusion features are weighted and combined to retain the unique information of each modality.

[0023] III. Disaster Prediction Module The disaster prediction module predicts the trend of key disaster indicators within the next 30 minutes by inputting multimodal feature representations into the time series prediction model. It should be noted that the training process of the time-series prediction model is as follows: Historical time-series data from various battery compartment scenes, internal environments, and public databases are collected to form a pre-training dataset (including numerical sensor data and corresponding image and video data). Then, the data undergoes automated spatiotemporal alignment and annotation to generate unified spatiotemporal graph data samples. Strong time-series data augmentation techniques (random time window slicing, sensor node masking, Gaussian noise injection, and pattern perturbations simulating sensor failures) are applied to improve the model's robustness. Using the constructed pre-training dataset as input, multi-task self-supervised learning objectives are designed to pre-train the large-scale time-series prediction model. The learning objectives include mask reconstruction, time-contrast prediction, and graph structure restoration (the mask reconstruction task randomly masks data from some sensor nodes at certain time steps in the input sequence, and the model reconstructs the masked data based on contextual information; the time-contrast prediction task...). By augmenting data across different time windows of the same sequence, the model is trained to make its output representations as similar as possible in the latent space, while being as different as possible from the representations of different sequences. In the graph structure recovery task (randomly shuffling or disrupting the spatial adjacency matrix of sensor nodes), the model is trained to infer and recover the correct spatial relationships based on the time-series dynamics of the nodes. After self-supervised pre-training, physical commonsense constraints are introduced to fine-tune the model. By comparing the sequence predicted by the model with the approximate results calculated based on simplified physical equations, and using this consistency as an auxiliary loss function, the model is guided to learn a dynamic evolution mechanism that conforms to physical laws through joint optimization of data-driven loss and physical consistency loss. Simultaneously, a fine-tuning dataset of the output results is constructed to perform supervised fine-tuning of the pre-trained model, enabling the model to understand and output results such as "predicting temperature field changes in the next 30 minutes" and "outputting a fire spread probability map."

[0024] The temporal prediction model adopts a hierarchical spatiotemporal Transformer architecture, which includes a local temporal encoder, a spatial relationship encoder, and a global spatiotemporal fusion unit. The local temporal encoder is used to capture the high-frequency and short-term dynamics of a single sensor node, and the spatial relationship encoder is used to model the spatial dependencies between nodes based on a graph attention mechanism. The global spatiotemporal fusion unit integrates local temporal features with global spatial context through a cross-scale attention mechanism to form a unified scene understanding. The disaster prediction module is based on the Bayesian state-space model, calculates and outputs the uncertainty range of the prediction trend, and generates a disaster prediction report based on the prediction trend and uncertainty. It should be noted that the Bayesian state-space model includes stochastic state transition equations and stochastic observation equations. The state transition equations are used to learn the dynamic evolution from the current state to the next state and output the probability distribution of the state; the observation equations are used to learn the mapping relationship from the system state to the observations of multimodal sensors.

[0025] The calculation process of the uncertainty interval is as follows: by performing multiple forward propagations on the same input sequence and activating dropout in each layer of the neural network, and by performing T random forward propagations, T sets of predicted trajectories of key indicators of future disasters are obtained. Based on the distribution of these T sets of trajectories, the expected value, variance and quantile of each prediction time step are calculated, thereby obtaining the prediction interval at the specified confidence level.

[0026] The prediction interval is decomposed into cognitive uncertainty and random uncertainty through uncertainty decomposition. Cognitive uncertainty is measured by calculating the output variance caused by the randomness of model parameters in T forward propagations, while random uncertainty is measured by calculating the noise variance inherent in the data itself. It should be noted that the decomposed uncertainty information is associated with the nodes in the spatiotemporal graph model to locate the sensor region or mode that contributes the most to the uncertainty and take it as the root cause of uncertainty. The specific steps of uncertainty decomposition are as follows: by calculating the contribution of different nodes or edges in the spatiotemporal graph neural network to the total uncertainty, the spatial source of uncertainty is realized, and the key uncertainty sources in the system are identified.

[0027] The steps for generating a disaster prediction report are as follows: Construct an uncertainty-aware report generation template library, with each template having reserved variable slots for inserting quantitative prediction values, uncertainty intervals, confidence levels, and trend qualitative terms. The most matching template is dynamically selected from the template library based on the severity of the predicted trend, the level of uncertainty, and the main sources of uncertainty. The final natural language disaster prediction report is automatically generated by filling the variable slots of the template with the results of the uncertainty interval and uncertainty decomposition steps.

[0028] IV. Tiered Early Warning Module The graded early warning module uses an adaptive clustering model to classify disaster levels based on the comprehensive disaster index and the predicted trend. The classification steps of the adaptive clustering model are as follows: unsupervised clustering of the comprehensive disaster index and the short-term predicted trend is performed by using a Gaussian mixture model, and the boundary thresholds of different disaster levels are dynamically determined based on the clustering results and the scenario type. It should be noted that when the slope of the future trend of the comprehensive disaster index is found to exceed the preset acceleration threshold, the warning level will be automatically upgraded by one level.

[0029] The comprehensive disaster index is obtained by constructing a multi-dimensional parameter set that includes factors such as fire intensity inside the cabin, surrounding environment, battery density, and asset impact. It is obtained by weighting and integrating each factor with its corresponding dynamic weight coefficient and combining it with the diffusion coefficient.

[0030] It should be noted that the weight coefficients of each factor adopt an attention mechanism and are dynamically adjusted according to real-time scene information.

[0031] V. Task Generation Module The task generation module is based on the constructed emergency task knowledge graph. It generates the optimal task chain corresponding to the disaster level through a resource game algorithm. The nodes of the emergency task knowledge graph represent task units, and the edges represent the logical and resource association relationships between tasks. It should be noted that the construction process of the emergency task knowledge graph is as follows: define the schema layer of the knowledge graph, including multiple core entity types and semantic relationship types between core entity types, acquire knowledge data from multiple heterogeneous data sources, and perform structured extraction and fusion of knowledge data based on the schema layer to form the initial dataset of the knowledge graph. By injecting real-time system status data into the corresponding entities of the initial dataset, a knowledge graph instance with dynamic attributes is formed.

[0032] The task chain generation process is as follows: After receiving the determined disaster level, the Monte Carlo tree search algorithm is used to search the knowledge graph in combination with the real-time system resource status. By evaluating the execution success rate, resource conflicts and total time of different task sequences, the task chain with the highest success probability is finally output as the execution plan.

[0033] It should be noted that for low-level warnings, a single task order is generated that includes management personnel verification and equipment inspection. For high-level warnings, a parallel task sequence is generated that includes shutting down critical facilities, initiating fire extinguishing procedures, and issuing alarm notifications, and warning information is sent to management personnel.

[0034] VI. Fire Extinguishing Execution Module The liquid nitrogen fire extinguishing unit includes: a base plate 1, a liquid nitrogen tank 2, and a high-pressure nitrogen cylinder 6. A liquid nitrogen recovery device 3 is installed on the upper end of the liquid nitrogen tank 2. The liquid nitrogen recovery device 3 has an outlet pipe 4 on the side near the storage cylinder and an inlet pipe 5 at the rear end. The outlet pipe 4 and the inlet pipe 5 are respectively connected to the inlet and outlet of the liquid nitrogen nozzle 13. Multiple high-pressure nitrogen cylinders 6 are provided, and a pressure reducing valve 9 is provided on the top. The pressure reducing valves 9 are interconnected through a manifold 7. A medium-pressure solenoid valve 8 corresponding to the pressure reducing valve 9 is provided on the manifold 7. The port of the manifold 7 is connected to a valve assembly 10 on the liquid nitrogen tank 2 through an air inlet pipe. A pressure gauge 12 is installed on the valve assembly 10. An exhaust port 11 is opened at the rear end of the valve assembly 10. A control component 14 is installed on the liquid nitrogen nozzle 13.

[0035] The liquid nitrogen fire extinguishing unit receives tasks in real time via the Internet of Things and can realize active and passive activation functions, pulse discharge and continuous discharge functions, and delayed discharge functions through the control component 14. The active start function is triggered by the warning level. When the system issues a high-level warning, the active start nozzle will ensure the smooth flow of the liquid nitrogen delivery pipeline. If the active start function fails, the nozzle will still have a passive start function. When the preset temperature is reached, the nozzle will start, which will still ensure the smooth flow of the liquid nitrogen delivery pipeline. The pulse emission and continuous emission functions determine the emission strategy based on the risk situation. When a high-level warning is generated, the pulse emission strategy is executed. If the high-level warning is not lifted within five minutes, the continuous emission strategy is executed. The delayed eruption function determines the delay time based on personnel information. When a high-level warning is issued, a buzzer is triggered and the emission program is started after a 30-second delay.

[0036] VII. Disaster Situation Tracing Module The disaster tracing module is based on the recorded multidimensional heterogeneous data stream. Through event extraction and entity recognition technology, it constructs an event knowledge graph with internal causal and temporal relationships. The multidimensional heterogeneous data stream includes the operation status logs and performance index data generated by each module, the original sensor data snapshots corresponding to key decision points, user operation audit trajectories, and contextual data of the internal environment of the battery compartment. It should be noted that the steps for constructing the event knowledge graph are as follows: using natural language processing technology to automatically extract key events and entities from unstructured logs, and using events and entities as nodes and the causal, temporal, and data flow relationships between events as edges to construct a graph structure. As the system runs, newly occurring events and relationships are incrementally updated into the graph structure in real time.

[0037] The event knowledge graph serves as a priori structure, integrating a causal inference model. After a disaster occurs, it inputs observed anomalies and automatically calculates and locates the root cause nodes. It should be noted that the causal inference model is a probabilistic inference model based on Bayesian networks. It provides a counterfactual query engine that modifies specific events in a multidimensional heterogeneous data stream according to query conditions as simulation input, and receives counterfactual query conditions submitted by users. It conducts inferences in a digital twin simulation environment constructed from multidimensional heterogeneous data streams to evaluate the event evolution results under different assumptions.

[0038] The disaster tracing module provides an interactive and visual tracing interface, enabling immersive exploration and retrospective analysis of multidimensional heterogeneous data streams, event knowledge graphs, and causal inference results.

[0039] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0040] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0041] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0042] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0043] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A lithium battery early thermal runaway warning and management system based on temperature detection, characterized in that, include: Data acquisition module: Collects fire-related data in real time through a sensor array and transmits the data to the cloud in encryption via the Internet of Things; Data processing module: Used to perform real-time analysis of the collected data stream, and to efficiently extract effective information from the data stream through data processing, thereby improving the overall system operating efficiency; Disaster prediction module: This module is used to input the processed data into the model for prediction and generate a prediction report of future disaster situations based on the output results. Tiered early warning module: Based on predicted data trends and a comprehensive disaster index, the disaster situation is adaptively classified into different levels; Task generation module: Generates corresponding task chains based on the level classification results, and converts the task chains into execution commands; Fire extinguishing module: Activates the liquid nitrogen fire extinguishing unit according to the execution command to quickly extinguish any fires that may occur; Disaster Tracing Module: Responsible for recording the operation logs of the entire system module and analyzing the causes of disasters by analyzing the log data.

2. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The data acquisition module collects fire-related data in real time through a sensor array and transmits the encrypted data to the cloud via the Internet of Things, wherein: The sensor array includes numerical sensors and image sensors. The numerical sensors include temperature and humidity sensors and gas sensors. The image sensors include thermal imaging sensors and visible light sensors. The encryption process for the data collected by the numerical sensor is as follows: the numerical data from multiple sampling periods, their corresponding timestamps, and device identifiers are merged into a single data packet, the entire data packet is encrypted, and a message authentication code is generated. The encryption process for the image sensor data is as follows: two consecutive frames of image data are treated as an independent payload, and only the payload part of the image data is encrypted, while the image frame header information is retained for network transmission scheduling.

3. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The data processing module is used to perform real-time analysis of the collected data stream. By efficiently extracting effective information from the data stream through data processing, it improves the overall system operating efficiency. The data processing module performs semantic understanding on the high-dimensional visual data of the image data stream by calling a pre-set lightweight multimodal large model, and compresses the understood high-dimensional visual data into a low-dimensional semantic token sequence, thereby achieving intelligent compression of image data. The data processing module uses the Kalman filter algorithm to optimize the collected numerical data, reducing fluctuations and errors caused by environmental noise and electromagnetic interference. The data processing unit maps the standardized point set representation of the numerical sensor and the semantic token sequence of the visual data to a unified feature space through an asymmetric dense projection network, and generates a deeply fused multimodal feature representation through a gated fusion mechanism. The gated fusion mechanism includes computational fusion gating, weighted fusion execution, and residual connection. The computational fusion gating step is as follows: based on the stitched numerical enhancement features and visual enhancement features, the fusion weight of each feature dimension is learned. The weighted fusion execution step is as follows: the numerical enhancement features and visual enhancement features are dynamically weighted and summed according to the fusion weight. The residual connection mechanism step is as follows: the original projection features and the fusion features are weighted and combined to retain the unique information of each modality.

4. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The disaster prediction module is used to input the processed data into the model for prediction, and generate a prediction report of future disaster situations based on the output results, wherein: The disaster prediction module inputs multimodal feature representations into a time-series prediction model to predict the trend of changes in key disaster indicators within the next thirty minutes. The temporal prediction model adopts a hierarchical spatiotemporal Transformer architecture, including a local temporal encoder, a spatial relationship encoder, and a global spatiotemporal fusion unit. The local temporal encoder is used to capture the high-frequency, short-term dynamics of a single sensor node. The spatial relationship encoder is used to model the spatial dependencies between nodes based on a graph attention mechanism. The global spatiotemporal fusion unit integrates local temporal features with global spatial context through a cross-scale attention mechanism to form a unified scene understanding. The disaster prediction module is based on the Bayesian state-space model, calculates and outputs the uncertainty interval of the prediction trend, and generates a disaster prediction report based on the prediction trend and uncertainty. The calculation process of the uncertainty interval is as follows: by performing multiple forward propagations on the same input sequence and activating dropout in each layer of the neural network, and by performing T random forward propagations, T sets of predicted trajectories of key indicators of future disasters are obtained. Based on the distribution of these T sets of trajectories, the expected value, variance and quantile of each prediction time step are calculated, thereby obtaining the prediction interval at the specified confidence level. The prediction interval is decomposed into cognitive uncertainty and random uncertainty. The cognitive uncertainty is measured by calculating the output variance caused by the randomness of the model parameters in T forward propagations, and the random uncertainty is measured by calculating the noise variance inherent in the data itself.

5. The disaster prediction module as described in claim 4 is used to input processed data into a model for prediction, and generate a prediction report of future disaster situations based on the output results, characterized in that... in: The steps for generating the disaster prediction report are as follows: Construct an uncertainty-aware report generation template library, and each template reserves variable slots for inserting quantitative prediction values, uncertainty intervals, confidence levels, and trend qualitative terms. The most matching template is dynamically selected from the template library based on the severity of the prediction trend, the level of uncertainty, and the main source of uncertainty. The final natural language disaster prediction report is automatically generated by filling the variable slots of the template with the results of the uncertainty interval and uncertainty decomposition in the steps.

6. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The tiered early warning module adaptively classifies disaster levels based on predicted data trends and a comprehensive disaster index, wherein: The graded early warning module uses an adaptive clustering model to classify disaster levels based on the comprehensive disaster index and the predicted trend. The division steps of the adaptive clustering model are as follows: unsupervised clustering of the comprehensive disaster index and the short-term predicted trend is performed by using a Gaussian mixture model, and the boundary thresholds of different disaster levels are dynamically determined based on the clustering results and the scenario type. The comprehensive disaster index is obtained by constructing a multi-dimensional parameter set that includes factors such as fire intensity inside the cabin, surrounding environment, battery density, and asset impact. The index is obtained by weighting and fusing each factor with its corresponding dynamic weight coefficient and combining it with the diffusion coefficient.

7. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The task generation module generates a corresponding task chain based on the level classification results, and converts the task chain into execution commands, wherein: The task generation module generates the optimal task chain corresponding to the disaster level based on the constructed emergency task knowledge graph and through a resource game algorithm. The nodes of the emergency task knowledge graph represent task units, and the edges represent the logical and resource association relationships between tasks. The process of generating the task chain is as follows: after receiving the determined disaster level, the Monte Carlo tree search algorithm is used to search the knowledge graph in combination with the real-time system resource status. By evaluating the execution success rate, resource conflicts and total time of different task sequences, the task chain with the highest success probability is finally output as the execution plan.

8. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The fire extinguishing execution module activates the liquid nitrogen fire extinguishing unit according to the execution command to quickly extinguish the fire, wherein: The liquid nitrogen fire extinguishing unit includes a base plate (1), a liquid nitrogen tank (2), and a high-pressure nitrogen cylinder (6). A liquid nitrogen recovery device (3) is installed at the upper end of the liquid nitrogen tank (2). The liquid nitrogen recovery device (3) has an outlet pipe (4) on the side near the original gas storage cylinder and an inlet pipe (5) at the rear end. The outlet pipe (4) and the inlet pipe (5) are respectively connected to the inlet and outlet of the liquid nitrogen nozzle (13). Multiple high-pressure nitrogen cylinders (6) are provided, and the top... A pressure reducing valve (9) is provided, and the pressure reducing valves (9) are interconnected through a manifold (7). A medium-pressure solenoid valve (8) corresponding to the pressure reducing valve (9) is provided on the manifold (7). The port of the manifold (7) is connected to the valve assembly (10) on the liquid nitrogen tank (2) through an air inlet pipe. A pressure gauge (12) is installed on the valve assembly (10). An exhaust port (11) is opened at the rear end of the valve assembly (10). A control component (14) is installed on the liquid nitrogen nozzle (13).

9. The fire extinguishing execution module as described in claim 8 activates the liquid nitrogen fire extinguishing unit according to the execution command to quickly extinguish the fire, characterized in that, in: The liquid nitrogen fire extinguishing unit receives tasks in real time via the Internet of Things and can realize active and passive activation functions, pulse spraying and continuous spraying functions, and delayed spraying functions through the control component 14. The active start function is triggered by the warning level. When the system issues a high-level warning, the active start nozzle will ensure the smooth flow of the liquid nitrogen delivery pipeline. If the active start function fails, the nozzle will still have a passive start function. When the preset temperature is reached, the nozzle will start, which will still ensure the smooth flow of the liquid nitrogen delivery pipeline. The pulse emission and continuous emission functions determine the emission strategy based on the risk situation. When a high-level warning is generated, the pulse emission strategy is executed. If the high-level warning is not lifted within five minutes, the continuous emission strategy is executed. The delayed eruption function determines the delay time based on personnel information. When a high-level warning is issued, a buzzer is triggered and the eruption program is started after a 30-second delay.

10. The lithium battery early thermal runaway warning management system based on temperature detection as described in claim 1, characterized in that, The disaster tracing module is responsible for recording the operation logs of the entire system module and analyzing the causes of the disaster by analyzing the log data, including: The disaster tracing module, based on the recorded multidimensional heterogeneous data stream, constructs an event knowledge graph with internal causal and temporal relationships through event extraction and entity recognition technologies. The multidimensional heterogeneous data stream includes the operation status logs and performance index data generated by each module, the original sensor data snapshots corresponding to key decision points, user operation audit trajectories, and contextual data of the internal environment of the battery compartment. The event knowledge graph serves as a priori structure, integrating a causal inference model. After a disaster occurs, it inputs observed anomalies and automatically calculates and locates the root cause nodes. The disaster tracing module provides an interactive and visual tracing interface, enabling immersive exploration and retrospective analysis of multidimensional heterogeneous data streams, event knowledge graphs, and causal inference results.

Images