A coal mine underground lithium battery thermal runaway gas early warning system and method

Abstract

The present application relates to the technical field of coal mine lithium battery early warning, and discloses a kind of based on coal mine lithium battery thermal runaway gas early warning system and method.The method includes real-time acquisition of the multidimensional monitoring data in the lithium battery pack inside coal mine, including characteristic gas concentration, temperature change rate, voltage fluctuation value and internal resistance offset;The multidimensional monitoring data is input into the pre-trained thermal runaway grading model, the highest priority is given to the characteristic gas concentration by dynamic weight distribution algorithm, and the comprehensive risk score under the current data window is calculated;According to the preset threshold interval where the comprehensive risk score is located, the corresponding first attention instruction, second intervention instruction or third emergency disposal instruction is triggered;When triggering the second intervention instruction or the third emergency disposal instruction, the location identification information of the lithium battery pack is automatically associated, and the emergency disposal plan containing the location identification is generated, which can effectively cope with the thermal runaway risk of lithium battery in coal mine and ensure the safety of underground operation.

Description

Technical Field

[0001] This invention relates to the field of early warning technology for lithium batteries in coal mines, specifically to an early warning system and method for thermal runaway gas from underground lithium batteries in coal mines. Background Technology

[0002] As coal mining technology upgrades towards intelligence and mechanization, lithium batteries, due to their high energy density, excellent charge and discharge efficiency, and low environmental pollution, are widely used in powering equipment in underground coal mines, such as underground inspection robots, explosion-proof transport vehicles, and emergency lighting devices, becoming an important energy component supporting continuous underground operations. However, the underground environment of coal mines is unique and complex, with problems such as high humidity, high dust, confined spaces, and limited ventilation. These environmental factors accelerate the aging of internal electrode materials and electrolyte decomposition in lithium batteries, increasing the risk of thermal runaway. Lithium battery thermal runaway is a multi-stage evolutionary process, from internal micro-short circuits causing localized temperature rises, to electrolyte decomposition producing characteristic gases, and then to a sudden temperature rise and voltage drop accompanied by open flames or explosions. The entire process can take anywhere from a few seconds to several minutes. Once it occurs, it not only damages the power supply equipment but may also trigger a chain reaction of accidents such as underground gas explosions and dust combustion, posing a serious threat to the lives of underground workers and mine property.

[0003] Current methods for monitoring the safety of lithium batteries in underground coal mines still have significant limitations. Some monitoring schemes rely on only a single parameter for early warning, such as monitoring only the battery surface temperature or terminal voltage. These schemes cannot comprehensively reflect the evolution of thermal runaway. While temperature parameters can intuitively reflect the dramatic temperature rise in the later stages of thermal runaway, surface temperature changes are often not obvious in the early stages of thermal runaway, when electrolyte decomposition and characteristic gas generation have already occurred inside the battery, leading to delayed early warnings. Voltage parameters are easily affected by equipment load fluctuations, and relying solely on voltage fluctuations to judge thermal runaway is prone to misjudgment or omission, failing to accurately identify risks. Other schemes attempt to use multi-parameter monitoring, integrating data such as temperature, voltage, and current, but the parameter weighting often adopts a fixed pattern, failing to consider the differences in early warning value of different parameters at different stages of thermal runaway. In fact, characteristic gases (such as carbon monoxide, hydrogen fluoride, methane, etc.) are the most sensitive signals in the early stages of thermal runaway in lithium batteries. Before the battery temperature rises significantly and the voltage fluctuates significantly, characteristic gases have already begun to be released continuously. However, existing multi-parameter monitoring schemes do not take the concentration of characteristic gases as a core early warning indicator, which leads to the failure to capture early risks in time and miss the best intervention opportunity.

[0004] Existing early warning systems lack effective correlation with equipment location information after triggering an alert. Coal mine roadways are intricately interwoven and have complex spatial layouts, with equipment scattered across different areas. When an alert is issued, it is difficult for personnel to quickly locate the specific position of the faulty lithium battery, leading to prolonged emergency response time. Some emergency plans only include general procedures and lack specific solutions tailored to the actual location of the faulty equipment. For example, they cannot adjust ventilation strategies based on the fault location, nor can they quickly deploy nearby firefighting equipment and rescue personnel, further reducing emergency response efficiency and increasing the possibility of accident escalation. Simultaneously, some outdated monitoring equipment underground suffers from low data acquisition frequency and high transmission latency, making it difficult to capture sudden parameter changes during thermal runaway in real time. This fails to provide a continuous and accurate data source for the early warning model, resulting in insufficient accuracy of the risk assessment results output by the model and failing to meet the high safety standards required for monitoring underground. These problems make the current early warning system for thermal runaway of lithium batteries in coal mines inadequate to effectively address the growing safety demands, necessitating a more comprehensive, accurate, and efficient early warning method to fill the technological gap. Summary of the Invention

[0005] The purpose of this invention is to provide an early warning system and method for thermal runaway gas from lithium batteries in underground coal mines, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for early warning of thermal runaway gas in underground lithium batteries in coal mines, the method comprising:

[0007] Real-time acquisition of multi-dimensional monitoring data inside lithium battery packs in underground coal mines, including characteristic gas concentration, temperature change rate, voltage fluctuation value, and internal resistance offset;

[0008] The multidimensional monitoring data is input into a pre-trained thermal runaway classification model. The characteristic gas concentration is assigned the highest priority through a dynamic weight allocation algorithm to generate a comprehensive risk score for the current data window.

[0009] Based on the preset threshold range where the comprehensive risk score is located, a corresponding early warning level instruction is triggered, which includes a Level 1 attention instruction, a Level 2 intervention instruction, and a Level 3 emergency response instruction.

[0010] When the Level 2 intervention command or Level 3 emergency response command is triggered, the location identification information of the lithium battery pack is automatically associated to generate an emergency response plan that includes the location identification.

[0011] Preferably, the construction process of the pre-trained thermal runaway grading model includes:

[0012] Acquire a full-cycle data sequence of historical thermal runaway events, which includes characteristic gas concentration evolution curves, temperature gradient distributions, and electrical parameter degradation trajectories from the normal state to the thermal runaway outbreak.

[0013] A temporal convolutional network is used to extract multi-scale features from the full-cycle data sequence, separating key modes that are strongly correlated with thermal runaway.

[0014] A hierarchical decision tree is constructed based on the key modes. The root node of the hierarchical decision tree uses the abrupt change in characteristic gas concentration as the primary splitting condition, and the secondary nodes sequentially introduce a composite judgment logic of temperature change rate and voltage fluctuation value.

[0015] Preferably, the execution process of the dynamic weight allocation algorithm includes:

[0016] The cumulative deviation of the characteristic gas concentration is calculated in real time, and the cumulative deviation is the multiple of the standard deviation between the current concentration value and the baseline concentration value;

[0017] The weighting coefficient of the characteristic gas concentration in the comprehensive risk score is dynamically adjusted according to the cumulative deviation. When the cumulative deviation exceeds the first critical value, the weighting coefficient is increased to a preset upper limit.

[0018] The acceleration of the rate of temperature change is monitored synchronously. If the acceleration remains positive and exceeds the second critical value, the weighting coefficient of the rate of temperature change is increased linearly.

[0019] Preferably, the process of generating the comprehensive risk score includes:

[0020] Each parameter in the multidimensional monitoring data is normalized to obtain a standardized parameter value;

[0021] The initial risk value is obtained by multiplying the standardized parameter value by the corresponding dynamic weight coefficient and then summing the results.

[0022] The current charge / discharge rate of the lithium battery pack is introduced as a correction factor to perform nonlinear correction on the initial risk value, and the final comprehensive risk score is output.

[0023] Preferably, the triggering process of the warning level instruction includes:

[0024] When the comprehensive risk score is in the first range, only a level 1 attention instruction is sent to the local monitoring terminal;

[0025] When the comprehensive risk score enters the second interval, a secondary intervention command is simultaneously sent to the local monitoring terminal and the regional security system, and the forced heat dissipation device of the lithium battery pack is activated.

[0026] When the comprehensive risk score reaches the third interval, a level-three emergency response instruction is broadcast to all emergency terminals downhole, and the power supply circuit of the lithium battery pack is cut off at the same time.

[0027] Preferably, the process of generating the emergency response plan includes:

[0028] The preset mine roadway topology map is retrieved based on the location identification information, and the nearest evacuation route and the coordinates of the explosion-proof facilities at the location of the lithium battery pack are marked.

[0029] Based on the current distribution data of personnel underground, calculate the optimal evacuation plan and emergency material allocation route;

[0030] The evacuation routes, coordinates of explosion-proof facilities, and emergency material allocation routes are integrated into a structured instruction set and sent to the target terminal.

[0031] Preferably, the method further includes a process of verifying the credibility of the multidimensional monitoring data:

[0032] A physical correlation constraint is established between the concentration of the characteristic gas and the rate of temperature change. When the trend of the two changes violates the physical correlation constraint, redundant sensor data comparison is initiated.

[0033] If the verification result of redundant data is confirmed to be abnormal, the current data window is marked as low confidence, and the data from the previous valid window is used for interpolation compensation.

[0034] Preferably, the method further includes a thermal runaway tracing analysis process:

[0035] After each Level 3 emergency response command is triggered, extract the sequence of changes in multi-dimensional monitoring data within the set time period prior to the trigger;

[0036] The parameter type and time point at which the earliest anomaly occurred were identified based on the change sequence;

[0037] The parameter types and time points are associated with the production batch and usage logs of the lithium battery pack to generate a thermal runaway cause analysis report.

[0038] Preferably, the method further includes an online model optimization process:

[0039] Collect data on the discrepancy between the actual handling results and the model prediction results of all events that have triggered early warnings;

[0040] When the cumulative frequency of the same type of deviation data exceeds a set threshold, the parameter fine-tuning of the thermal runaway classification model is initiated.

[0041] The bias data is transformed into new training samples using an incremental learning algorithm, and the judgment threshold of the hierarchical decision tree is updated.

[0042] Preferably, the present invention also includes an early warning system for thermal runaway gas from lithium batteries in underground coal mines. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described method for early warning of thermal runaway gas from lithium batteries in underground coal mines.

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

[0044] By collecting multi-dimensional monitoring data from inside the lithium battery pack, the limitations of traditional single-parameter monitoring are overcome, enabling comprehensive perception of the lithium battery thermal runaway process. Characteristic gas concentration, temperature change rate, voltage fluctuation, and internal resistance shift correspond to key signals at different stages of thermal runaway. Characteristic gas concentration reflects the internal chemical reactions in the early stages of thermal runaway, temperature change rate reflects the energy release intensity in the middle stages, and voltage fluctuation and internal resistance shift reveal the degree of damage to the battery's internal structure. The coordinated acquisition of multiple parameters covers the entire stage of thermal runaway from its inception to its full outbreak, avoiding omissions or misjudgments caused by incomplete information from a single parameter. This allows the monitoring data to more comprehensively reflect the actual safety status of the lithium battery, meeting the high accuracy requirements of the complex environment in coal mines.

[0045] In the data processing stage, the dynamic weight allocation algorithm assigns the highest priority to the concentration of characteristic gases, fully aligning with the evolutionary pattern of lithium battery thermal runaway. The early stage of thermal runaway is a critical phase for mitigating risks. At this time, the battery surface temperature has not yet risen significantly, and the voltage has not fluctuated drastically, but characteristic gases have already begun to be generated internally. If this signal can be captured in time, it can buy more time for intervention. The dynamic weight allocation is not fixed but adjusts the weight ratio according to the changing trends of each parameter within the current data window. When the concentration of characteristic gases rises abnormally, the algorithm automatically increases its weight, ensuring that the model prioritizes generating risk scores based on the most sensitive early signals. When thermal runaway enters the middle and late stages, and the changes in parameters such as the rate of temperature change and voltage fluctuations intensify, the algorithm will also appropriately adjust the weights of each parameter to ensure that the comprehensive risk score can dynamically match the real-time stage of thermal runaway, making the risk assessment more in line with the actual situation and avoiding delayed warnings or misjudgments caused by fixed weights.

[0046] The three-tiered early warning system enables differentiated and precise risk management, avoiding resource waste or inadequate response caused by a "one-size-fits-all" approach. The Level 1 "Attention" command corresponds to a low overall risk score. At this level, although the lithium battery exhibits minor abnormalities, it has not reached an emergency state. Workers can respond gently by increasing monitoring frequency and adjusting equipment load, without needing to suspend normal underground operations, thus ensuring production continuity. The Level 2 "Intervention" command corresponds to medium risk, indicating a clear tendency for thermal runaway in the lithium battery. Immediate proactive intervention measures are required, such as activating the cooling system and cutting off the battery power supply circuit, to prevent further escalation of the risk. The Level 3 "Emergency Response" command corresponds to high risk, meaning thermal runaway is imminent or has initially occurred. The highest level of emergency response must be activated, such as evacuating surrounding personnel and using fire extinguishers, to minimize the impact of the accident. The direct link between different levels of early warning commands and specific response measures allows workers to quickly identify the appropriate course of action, improving the efficiency and targeted nature of risk management.

[0047] When a Level 2 or Level 3 warning is triggered, the system automatically associates the location identification information of the lithium battery pack and generates an emergency response plan, effectively solving the problems of difficult equipment location and slow emergency response in underground mines. Coal mine roadways are complex and equipment is scattered. Associating location identification information allows workers to quickly obtain information such as the specific roadway number and equipment location of the faulty lithium battery through the warning system, eliminating the need for extensive investigation and significantly shortening emergency response time. Simultaneously, the emergency response plan, which includes location identification, is tailored to the actual environmental characteristics of the fault location. For example, if the fault location is near an area with high methane concentration, the plan prioritizes measures such as enhanced ventilation and prevention of open flames; if the fault location is in a densely populated area, the plan clearly defines evacuation routes and assembly points. This location-linked personalized plan makes emergency response measures more aligned with the actual situation on-site, avoiding the inadequacy of general plans in different scenarios, further improving the effectiveness of emergency response, and ensuring the safety of underground workers and mine property.

[0048] This method is based on a pre-trained thermal runaway grading model. The model can be iteratively optimized by continuously accumulating actual underground monitoring data, constantly improving its ability to identify thermal runaway characteristics of lithium batteries of different brands and specifications, thus adapting to the needs of upgrading and replacing lithium battery equipment in coal mines. Simultaneously, the entire early warning process requires no manual intervention. From data collection and risk score calculation to early warning command triggering and emergency plan generation, everything is automated, reducing errors caused by manual operation and lowering reliance on the professional skills of personnel. It can operate stably in unattended underground areas or harsh working environments, providing a more reliable and sustainable technical solution for the safety monitoring of lithium batteries in coal mines. Attached Figure Description

[0049] Figure 1A weighted graph of risk assessment parameters;

[0050] Figure 2 The flowchart shows the construction and feature extraction of a thermal runaway staging model.

[0051] Figure 3 Dynamic evaluation diagram for early warning system of thermal runaway risk of lithium battery in underground coal mine;

[0052] Figure 4 This is a flowchart for the generation and distribution of emergency response plans. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides a method for early warning of thermal runaway gas in lithium batteries in underground coal mines. The method includes: real-time acquisition of multi-dimensional monitoring data inside the lithium battery pack through an integrated sensor system. This data includes characteristic gas concentration, temperature change rate, voltage fluctuation value, and internal resistance offset. The acquired data is transmitted to a processing unit, where a pre-trained thermal runaway classification model analyzes the data. This model employs a dynamic weight allocation algorithm, assigning the highest priority to characteristic gas concentration during calculation to ensure that abnormal gas signals are identified first. The model outputs a comprehensive risk score for the current data window, which is compared with a preset threshold range. When the score falls into different ranges, corresponding early warning level instructions are triggered, including a Level 1 warning instruction, a Level 2 intervention instruction, and a Level 3 emergency response instruction. When a Level 2 intervention instruction or a Level 3 emergency response instruction is triggered, the system automatically associates the location identification information of the lithium battery pack and generates an emergency response plan containing location details, thereby achieving rapid response.

[0055] Example 1: See Figure 2The pre-trained thermal runaway grading model is built upon a complete historical thermal runaway event data sequence. This data sequence originates from a real-world case library recorded during the operation of lithium-ion battery packs in underground coal mines, covering various operating conditions and battery aging states. The complete historical thermal runaway event data sequence includes characteristic gas concentration evolution curves, temperature gradient distributions, and electrical parameter degradation trajectories throughout the entire continuous period from normal stable operation to thermal runaway outbreak. The characteristic gas concentration evolution curves record the complete trend of the generation rate and accumulation of key gases such as hydrogen, carbon monoxide, and hydrocarbons over time. The temperature gradient distribution depicts the heat transfer and accumulation process between different measuring points inside the battery pack. The electrical parameter degradation trajectory focuses on tracking the decay paths of operating voltage, internal resistance, and charge / discharge efficiency. The data acquisition system synchronously records these parameters at high frequency, ensuring that the complete historical thermal runaway event data sequence is strictly aligned on the time axis, forming the original material for model training.

[0056] The multi-scale feature extraction stage employs a temporal convolutional network (TCNN) to process the full-cycle data sequence of historical thermal runaway events. The TCNN utilizes its dilated causal convolutional structure to capture latent patterns across different time scales, ranging from seconds to hours. The input layer of the TCNN receives a pre-processed, standardized data stream. Deeper layers focus on identifying macroscopic trends spanning longer time periods, such as the slow accumulation of characteristic gas concentrations, while shallower layers focus on short-term transient phenomena, such as sudden voltage drops or rapid temperature increases. Through this hierarchical processing, the TCNN can separate key modes strongly correlated with thermal runaway from the complex full-cycle data sequence of historical thermal runaway events. These key modes can be understood as a series of highly abstract feature vectors that collectively characterize the essential changes in the system state before thermal runaway occurs.

[0057] Based on the key modes extracted by the temporal convolutional network, the construction process enters the stage of building a hierarchical decision tree. The design philosophy of the hierarchical decision tree is to simulate the logical sequence of expert risk assessment, decomposing complex multi-parameter decisions into a series of hierarchical judgment steps. The root node of the hierarchical decision tree uses the abrupt change in characteristic gas concentration as the primary splitting condition. This abrupt change is obtained by calculating the slope of the gas concentration change per unit time. This design is based on the physical law that the initial stage of thermal runaway is always accompanied by anomalies in characteristic gases. When the judgment condition of the root node meets the threshold, the data flow enters the secondary nodes of the hierarchical decision tree. The secondary nodes sequentially introduce the composite judgment logic of temperature change rate and voltage fluctuation value. The composite judgment logic of temperature change rate and voltage fluctuation value is not a simple comparison of the thresholds of independent parameters, but rather examines the temporal coupling relationship between the temperature change rate and voltage fluctuation value. For example, a rapid rise in temperature accompanied by violent voltage oscillations is considered a higher risk indicator. The leaf nodes of the hierarchical decision tree ultimately correspond to different preliminary risk level judgments.

[0058] The training process of the hierarchical decision tree relies on the full-cycle data sequence of historical thermal runaway events with accurate labels. The training algorithm recursively selects the optimal splitting features and splitting points starting from the root node to maximize the data purity of the child nodes. The abrupt change in the characteristic gas concentration is forcibly set as the splitting feature of the root node to ensure the priority of gas parameters in the decision path. The depth and number of nodes of the hierarchical decision tree are controlled through pruning techniques to prevent overfitting to the full-cycle data sequence of historical thermal runaway events and ensure that the hierarchical decision tree has good generalization ability. The trained hierarchical decision tree and the front-end temporal convolutional network together constitute the thermal runaway classification model. The temporal convolutional network is responsible for extracting key information from the raw data, and the hierarchical decision tree makes interpretable risk classifications based on the key information.

[0059] The validation phase of the thermal runaway grading model uses a full-cycle data sequence of historical thermal runaway events that were not used in the training. The validation process evaluates the accuracy and timeliness of the model in identifying different stages before thermal runaway occurs. The parameters of the thermal runaway grading model, especially the judgment thresholds of each node in the hierarchical decision tree, need to be fine-tuned based on the validation results. This fine-tuning aims to balance the false alarm rate and the false negative rate to adapt to the actual needs of underground coal mine safety management. The finalized thermal runaway grading model is deployed to the underground real-time computing unit. The model continuously receives multi-dimensional data streams from battery monitoring sensors and outputs a comprehensive risk score. The online operation mechanism of the thermal runaway grading model is periodic. It reads the latest multi-dimensional monitoring data using a sliding time window. For each data segment within a time window, the model calls a temporal convolutional network module for feature extraction to obtain the key modality representation for the current window. The hierarchical decision tree module receives the key modality representation and judges layer by layer starting from the root node, outputting the risk level corresponding to that time window. The calculation frequency of the thermal runaway grading model is synchronized with the data acquisition frequency to ensure real-time response to changes in battery state.

[0060] Example 2: The execution process of the dynamic weight allocation algorithm begins with the continuous monitoring and calculation of the characteristic gas concentration parameter, which is one of the most sensitive indicators in thermal runaway early warning. The dynamic weight allocation algorithm calculates the cumulative deviation of the characteristic gas concentration in real time. The cumulative deviation is a statistic used to quantify the degree of deviation of the current gas concentration reading from the historical normal baseline. The specific calculation method for the cumulative deviation is to take the difference between the current characteristic gas concentration value and the pre-established baseline concentration value, and then divide this difference by the standard deviation of the baseline concentration dataset. The resulting factor is the cumulative deviation of the characteristic gas concentration. The baseline concentration value is determined by collecting characteristic gas concentration data over a long period under normal lithium battery pack conditions and calculating its average value. This baseline value is updated periodically to reflect the slow drift caused by battery aging or environmental changes. The dynamic weight allocation algorithm dynamically adjusts the weight coefficient of the characteristic gas concentration in the subsequent comprehensive risk score calculation based on the calculated cumulative deviation value. The weight coefficient is an amplification factor used to adjust the contribution ratio of different monitoring parameters in the final risk score. When the cumulative deviation of the characteristic gas concentration exceeds a preset first threshold, the dynamic weighting algorithm immediately increases the weighting coefficient of the characteristic gas concentration to a preset upper limit. The first threshold is typically set to 2 to 3 times the standard deviation of the baseline concentration value; this threshold signifies a statistically significant anomaly in the characteristic gas concentration. Increasing the weighting coefficient of the characteristic gas concentration to the preset upper limit ensures that the characteristic gas concentration parameter dominates the overall risk score when a clear gas anomaly signal is detected, thereby ensuring the system's sensitivity to the most dangerous signs.

[0061] The dynamic weight allocation algorithm does not only focus on the characteristic gas concentration parameter; it also simultaneously monitors the dynamic changes in the temperature change rate parameter. The temperature change rate parameter describes how quickly the battery's temperature rises per unit time. The algorithm pays particular attention to the acceleration of the temperature change rate, which refers to the rate of change of the temperature itself—whether the temperature rise is continuously accelerating. The algorithm calculates the acceleration value of the temperature change rate in real time and judges its sign and magnitude. If the acceleration value remains positive and exceeds a preset second threshold, the algorithm increases the weight coefficient of the temperature change rate parameter in the comprehensive risk score according to a preset linear relationship. The second threshold is set based on extensive historical data analysis, indicating that the temperature has entered a non-linear, potentially runaway, accelerating state. The adjustment of the weight coefficient of the temperature change rate by the dynamic weight allocation algorithm is gradual, and the adjustment magnitude is proportional to the degree to which the acceleration value of the temperature change rate exceeds the second threshold. This design gives greater weight to the temperature change rate parameter in risk assessment when the accelerating upward trend in temperature is more pronounced. The dynamic weight allocation algorithm adjusts the weights of characteristic gas concentration and temperature change rate independently but in a related manner. The core logic of the dynamic weight allocation algorithm is to assign the highest basic priority to the characteristic gas concentration, while the weight increase of the temperature change rate is a superimposed enhancement mechanism when it exhibits an accelerating deterioration trend.

[0062] The execution of the dynamic weight allocation algorithm is a continuous cyclical process. In each calculation cycle, the algorithm rereads the latest multidimensional monitoring data and recalculates the cumulative deviation of characteristic gas concentrations and the acceleration of temperature change rates. Based on the latest calculation results, the algorithm updates the dynamic weight coefficients of each parameter in real time, ensuring that the weight assigned to each parameter most accurately reflects the actual risk status of the battery at the current moment. The parameters used in the dynamic weight allocation algorithm, such as the first and second critical values, are not fixed. These thresholds can be set within a reasonable range and fine-tuned according to different battery models and specific downhole environments to optimize the early warning performance of the algorithm in different application scenarios. The output of the dynamic weight allocation algorithm is a set of real-time changing dynamic weight coefficients, which are directly passed to the comprehensive risk score generation module. In the comprehensive risk score generation module, the normalized values ​​of each monitoring parameter are multiplied by the corresponding dynamic weight coefficients provided by the algorithm. The magnitude of the dynamic weight coefficient directly determines the parameter's influence on the final comprehensive risk score. Through the operation of the dynamic weight allocation algorithm, the system achieves dynamic shifts in the risk assessment focus. When the concentration of characteristic gases shows significant anomalies, the early warning system quickly enters a high-alert state. When the rate of temperature change shows accelerated characteristics, the system further strengthens its focus on heat accumulation risks. The dynamic weight allocation algorithm transforms the early warning logic from a static comparison of parameter thresholds into an intelligent weighting process capable of capturing the dynamic correlation and evolution trend between parameters, thereby improving the accuracy and foresight of the early warning. The effective operation of the dynamic weight allocation algorithm relies on high-quality, high-frequency real-time data streams. The reliability of the data acquisition system is the foundation for the dynamic weight allocation algorithm to make correct weight allocation decisions. The code for the entire dynamic weight allocation algorithm is integrated into the data processing unit of the underground monitoring host and runs periodically as a software module. The computational load is optimized to adapt to the limitations of underground computing resources in coal mines.

[0063] See Figure 3In the design of the dynamic evaluation mechanism for the early warning system of lithium battery thermal runaway risk in coal mines, the evolution process of the comprehensive risk score is visualized through a dual-graph coupling. Specifically, the comprehensive risk score change curve uses time as the horizontal axis (0-120 minutes) and risk level quantification as the vertical axis (0-100 points). The black solid line represents the score curve generated based on a dynamic weight allocation algorithm, and its trend follows an exponential growth model, breaking through the 80-point secondary intervention threshold at 20 minutes, and finally converging to the 100-point tertiary emergency threshold. The change in warning level adopts a discrete state transition model, with the vertical axis defining the warning state machine: Normal → Level 1 Attention → Level 2 Intervention → Level 3 Emergency. The system completes the three-level transition within 20 minutes: the initial steady state remains normal, Level 1 Attention is triggered at 5 minutes, Level 2 Intervention is upgraded at 15 minutes, and Level 3 Emergency is reached at 20 minutes. The spatiotemporal alignment mechanism of the dual graphs reveals the dynamic coupling between the risk score gradient and the early warning response logic, and the temporal synchronization between abnormal parameters such as empirical characteristic gas concentration and threshold triggering. In the parameter configuration, the risk scoring thresholds are strictly set to 80 points (Level 2) and 100 points (Level 3), the time axis sampling interval is 20 minutes, and the sensitivity of the weighting algorithm to the priority of gas concentration is verified by the change of curve slope.

[0064] Example 3: The generation process of the comprehensive risk score is the core computational step of the early warning method. This process receives two inputs from the front-end module: standardized parameter values ​​obtained through normalization and dynamic weight coefficients calculated in real-time by a dynamic weight allocation algorithm. The characteristic gas concentration, temperature change rate, voltage fluctuation value, and internal resistance offset included in the multi-dimensional monitoring data need to be normalized. Normalization maps the original monitoring data of different physical dimensions and orders of magnitude to a unified [0,1] interval, transforming them into standardized parameter values. The normalization benchmark for the characteristic gas concentration is its maximum value within its historical safe operating range and the theoretical alarm threshold. The normalization reference for the temperature change rate is the maximum allowable temperature rise rate of the battery under normal operating conditions. The voltage fluctuation value and internal resistance offset are scaled according to the normal fluctuation range defined in their technical specifications. Normalization allows these fundamentally different parameters—characteristic gas concentration, temperature change rate, voltage fluctuation value, and internal resistance offset—to be weighted, compared, and comprehensively calculated on a common mathematical basis.

[0065] Each standardized parameter value is multiplied by its corresponding dynamic weighting coefficient, which is dynamically determined by the dynamic weighting algorithm based on the real-time anomaly level of the parameter. The dynamic weighting coefficient for characteristic gas concentration remains at its highest level in most cases; however, as the cumulative deviation of the characteristic gas concentration increases, its dynamic weighting coefficient rapidly climbs to a preset upper limit. The dynamic weighting coefficient for temperature change rate gains a linear gain when its acceleration is significant, while the dynamic weighting coefficients for voltage fluctuation and internal resistance offset remain relatively stable or undergo only minor adjustments. All weighted parameter values—the product of the standardized parameter values ​​and the dynamic weighting coefficients—are fed into an adder for summation, generating an initial risk value. This initial risk value is a continuous, dimensionless value that preliminarily characterizes the overall risk situation of the lithium battery pack within the current data window. The initial risk value monotonically increases as the anomaly level of each monitored parameter intensifies.

[0066] The initial risk value needs to be corrected by a nonlinear correction step before it can be output as the final comprehensive risk score. This nonlinear correction step introduces a key correction factor: the current charge / discharge rate of the lithium battery pack. The charge / discharge rate is a crucial parameter describing the battery's operating intensity; high charge / discharge rates drastically increase the rate of electrochemical reactions within the battery, generating a large amount of heat and significantly increasing the risk of thermal runaway. The correction factor's adjustment of the initial risk value is nonlinear, reflecting the general trend of exponential growth in thermal runaway risk under high-load conditions. The nonlinear correction process is defined by the following mathematical expression:

[0067]

[0068] in: This represents the final output comprehensive risk score after nonlinear correction; it is a dimensionless numerical value. This represents the initial risk value before nonlinear correction, and is also a dimensionless value. Represents the absolute value of the current charge / discharge rate read in real time from the battery management system, with dimensions of... . It is a reference ratio, and its dimensions are the same as... Consistent, typically taking a value of 1C, its introduction makes the ratio It becomes a dimensionless quantity, representing the intensity of the current magnification relative to the reference magnification. It is a dimensionless proportionality coefficient used to adjust the weight of the charge / discharge rate in the overall risk score, and its value is related to the battery's chemical system. It is a dimensionless shape parameter used to control the overall risk score as a function of the normalization factor. The formula describes the increasing nonlinear curvature. Both sides of the formula are dimensionless values ​​with strictly consistent dimensions. This formula ensures a comprehensive risk score even at low magnification. It depends mainly on the initial risk value. However, when operating at high magnification, the overall risk score... This will be significantly amplified through the index term, thus more accurately reflecting the true risk level of the battery.

[0069] Once the comprehensive risk score is generated, it is immediately sent to the triggering logic unit of the warning level instruction. The triggering process maps continuous comprehensive risk score values ​​to discrete warning instructions with clearly corresponding actions. The system internally presets two thresholds, dividing the comprehensive risk score range into three consecutive intervals: the first interval, the second interval, and the third interval. When the comprehensive risk score is in the lowest first interval, the system determines that the current risk is at the attention level, and the warning level instruction triggering process generates and sends a Level 1 attention instruction to the local monitoring terminal. The Level 1 attention instruction is typically presented as a screen message prompt or a gentle audio-visual signal, designed to inform inspection personnel or operators to pay attention to the battery pack's operating status; no automated control actions are initiated at this time.

[0070] When the comprehensive risk score rises and crosses the first threshold into the second interval, it indicates that the risk has reached a level requiring proactive system intervention. The triggering process for the warning level instruction simultaneously executes two operations. One operation is to send a Level 2 intervention instruction to the local monitoring terminal and the broader regional safety system. This Level 2 intervention instruction has a higher alarm priority and stronger alert intensity, aiming to elicit immediate attention and coordinated response from safety personnel at multiple levels. The other operation is to automatically send an activation signal to the forced cooling device of the lithium battery pack. This device typically includes a high-efficiency fan or liquid-cooled pump, and its activation aims to suppress the rising battery temperature by enhancing heat dissipation, thereby controlling and reducing the risk. When the comprehensive risk score continues to rise and exceeds the second threshold to reach the highest level, the third interval, it signifies that the risk of thermal runaway is extremely critical. The triggering process for the warning level instruction will initiate the highest level of emergency response. A Level 3 emergency response instruction is generated and immediately broadcast to all emergency terminals downhole, covering the hazardous area and all relevant personnel downhole, ensuring that no warning information is missed. Simultaneously, the system sends a high-priority hardwired safety command to directly cut off the power supply circuit of the lithium battery pack, physically disconnecting the battery's energy output path. This is the last automatic electrical protection measure to prevent thermal runaway from escalating into a fire or explosion. The triggering process of the warning level command is a deterministic, automated process based on preset threshold logic. Each command level is associated with progressively escalating information notification scope and equipment handling actions, collectively forming a hierarchical and orderly safety protection system. The generation process of the comprehensive risk score is seamlessly integrated with the triggering process of the warning level command, forming a complete closed loop from multi-parameter perception and fusion assessment to graded response.

[0071] Example 4: See Figure 4 The emergency response plan generation process begins immediately upon the early warning system triggering a Level 2 intervention command or a Level 3 emergency response command. The core basis for this process is the location identification information of the lithium battery pack that triggered the alarm. This location identification information is typically provided by RFID tags or Bluetooth beacons pre-installed on the battery pack. The system accurately determines the battery's spatial coordinates underground by reading the unique codes of these tags. Based on the location identification information, the system retrieves a pre-set mine roadway topology map stored in a central database. This map is a digital map that details the routing and connections of all main roadways, auxiliary roadways, connecting roadways, chambers, safety exits, hoisting equipment, and ventilation, drainage, and compressed air pipelines. The mine roadway topology map forms the geospatial basis for generating all subsequent action plans. On the loaded mine roadway topology map, the system automatically marks the specific coordinates of the lithium battery pack that triggered the alarm. Centered on the coordinates of the lithium battery pack, the system uses a path planning algorithm to calculate and mark the nearest evacuation route. The nearest evacuation route is the optimal path from the location of the battery pack to the nearest safe exit or refuge chamber. The path planning algorithm comprehensively considers the accessibility of the tunnels, the type of tunnels, and real-time environmental parameters. Simultaneously, the system highlights the coordinates of nearby explosion-proof facilities on the mine tunnel topology map, including the specific locations of fire extinguisher boxes, fire hydrants, sandboxes, and explosion-proof doors. This marking information provides on-site personnel with the most direct guidance on the location of emergency resources.

[0072] The emergency response plan generation process then involves data interaction with the underground personnel positioning system to obtain current underground personnel distribution data. This data reflects the real-time location of all workers carrying positioning cards on the mine roadway topology map at the moment the alarm is triggered. The system combines the lithium battery pack coordinates, the nearest evacuation route, the coordinates of explosion-proof facilities, and the underground personnel distribution data to calculate the optimal evacuation plan. The optimal evacuation plan not only includes planning the shortest path to a safe location for personnel in hazardous areas but also avoids guiding personnel from different work faces to the same potentially congested roadway. The calculation of the optimal evacuation plan follows basic principles of coal mine safety regulations, such as "evacuation against the wind flow" and "evacuation towards the intake airway." Simultaneously, the system generates one or more emergency material allocation routes, indicating the best path for rescue personnel or vehicles to safely and efficiently transport emergency supplies such as fire extinguishers and respirators from material warehouses or backup sites to the incident site.

[0073] All path, coordinate, and plan information will be integrated into a structured instruction set, which uses a standardized data format that is both machine-readable and easily understood by humans. The structured instruction set typically includes the following main parts: precise identification and location description of the alarm battery pack; a list of evacuation routes planned for personnel in different areas; a list of nearby available explosion-proof facilities and their functional descriptions; and details of the origin-destination-route for emergency material allocation. The structured instruction set is distributed to target terminals via mining industrial Ethernet or an emergency broadcast system. Target terminals include computers in the regional safety monitoring center, explosion-proof smartphones of on-site personnel, audible and visual alarms in the tunnels, and information displays in refuge chambers. The information displayed on different target terminals may vary; for example, the monitoring center may see a complete mine tunnel topology map overlaid with dynamic command information, while on-site personnel's mobile phones may receive concise text evacuation instructions.

[0074] The process of verifying the credibility of multi-dimensional monitoring data runs in parallel with the generation of emergency response plans. This verification process is crucial for ensuring the authenticity and reliability of early warning signals. The system internally establishes a physical correlation constraint rule base for characteristic gas concentration and temperature change rate. This constraint is based on the mechanism of lithium battery thermal runaway: internal short circuits generate heat, leading to temperature rise, which accelerates electrolyte decomposition and volatilization, producing characteristic gases. Therefore, in the early stages of actual thermal runaway, the upward trend of characteristic gas concentration and the increasing trend of temperature change rate should be synchronous or positively correlated in time. The system compares the characteristic gas concentration change curve and the temperature change rate change curve in the current data window in real time. When the trends of both violate the physical correlation constraint between characteristic gas concentration and temperature change rate—for example, a sharp increase in characteristic gas concentration while the temperature change rate remains stable or even decreases—the system will determine that the current data is abnormal and initiate a redundant sensor data comparison process.

[0075] The redundant sensor data comparison process cross-validates data from redundant sensors deployed within or adjacent to the same lithium battery pack that measure the same physical quantity. Referring to Table 1, a typical redundant data comparison scenario is illustrated: the system detects an abnormally high reading from gas sensor #1, but the temperature sensor #1, installed alongside it, does not show a corresponding temperature rise.

[0076] Table 1: Comparison Table of Redundant Sensor Data

[0077]

[0078] If the redundancy data verification results confirm that the anomaly is caused by a single sensor failure or interference, the system will mark the current data window as low confidence. The system will automatically discard the low confidence data in the current window and use the readings from the previous valid data window for interpolation compensation as input to the thermal runaway grading model. The interpolation compensation algorithm typically uses linear extrapolation or maintains the previous valid value to ensure the continuity of the model input data. The process of verifying the confidence of multi-dimensional monitoring data effectively reduces the probability of false alarms triggered by sensor false alarms or temporary interference, improving the reliability of the entire early warning system. The emergency response plan generation process works in conjunction with the multi-dimensional monitoring data confidence verification process to ensure the accuracy and necessity of the emergency response, avoiding resource misuse and personnel panic caused by data distortion.

[0079] Example 5: The thermal runaway source tracing analysis process is automatically initiated after each warning system triggers a Level 3 emergency response command. This process aims to trace the root cause of thermal runaway events. The system extracts a complete multidimensional monitoring data change sequence from the historical data storage server within a set timeframe prior to triggering the Level 3 emergency response command. This set timeframe typically covers tens of minutes to an hour before the thermal runaway critical point, ensuring the earliest abnormal signs are captured. The multidimensional monitoring data change sequence includes continuous records of parameters such as characteristic gas concentration, temperature change rate, voltage fluctuation value, and internal resistance offset on the time axis. The data is acquired at a high frequency, forming a high-resolution time series dataset. Based on the extracted multidimensional monitoring data change sequence, the system runs an anomaly identification algorithm to identify the earliest abnormal parameter type and time point. The anomaly identification algorithm performs independent trend analysis and abrupt change point detection on each parameter sequence, calculating the degree of deviation and duration of deviation for each parameter relative to its historical baseline value. For example, the system might identify a sustained upward trend in the characteristic gas concentration parameter, exceeding the normal fluctuation range, 23 minutes before the Level 3 emergency response command is triggered, while the temperature change rate parameter remains within the normal range at 20 minutes. This earliest identified time point, the moment the characteristic gas concentration anomaly began, is marked as the potential earliest anomalous time point. The identification of parameter types and time points needs to be accurate to the second in order to construct a clear timeline of event development.

[0080] The earliest identified abnormal parameter type and time point need to be associated with the lithium battery pack's identity information. The system retrieves the corresponding production batch and usage logs through the lithium battery pack's unique code. The production batch information records key data such as the battery manufacturer, manufacturing date, electrode material formulation, and initial capacity. The usage log contains the complete history of the battery pack since it was put into operation, such as cumulative cycle count, average charge / discharge depth, historical maximum operating temperature, and whether it has experienced abnormal events such as overcharging or external short circuits. The system performs cross-correlation analysis on the earliest abnormal parameter type and time point with relevant entries in the production batch and usage logs. The correlation analysis process attempts to find potential connections between abnormal parameter patterns and the battery's historical state. For example, the analysis may find that the earliest abnormal characteristic gas concentration parameter is close in time to a recent high-rate discharge event recorded in the battery usage log, or has a statistical correlation with the electrolyte formulation of a specific batch shown in the production batch information. Based on a predefined rule base and pattern matching algorithm, the system integrates parameter type, time point, production batch data, and usage log information to automatically generate a structured thermal runaway cause analysis report. A thermal runaway cause analysis report typically includes an event overview, time-series analysis of abnormal parameters, characteristics of associated production batches, key points of associated usage history, and inferential conclusions regarding possible causes.

[0081] The online model optimization process and the thermal runaway source analysis process are carried out in parallel. The online model optimization process aims to continuously improve the predictive accuracy of the thermal runaway classification model based on actual operational feedback. The system collects complete data packages of all triggered warning events (including Level 1 attention instructions, Level 2 intervention instructions, and Level 3 emergency response instructions). The data packages contain multi-dimensional monitoring data before and after the warning is triggered, the comprehensive risk score predicted by the model, and the actual handling results of the event. The actual handling results include the actual battery status confirmed by on-site personnel. The online model optimization process compares the model's prediction results with the actual handling results, calculating the deviation data between the two. The deviation data quantifies the model's prediction error in a specific event; for example, the model might give a high-risk score but on-site inspection confirms a normal state, or the model might give a low score but the battery subsequently experiences thermal runaway. The deviation data is stored in a dedicated model optimization database, and the system continuously monitors the accumulation of deviation data. When the cumulative frequency of the same type of deviation data exceeds a preset threshold—for example, if the number of false alarms for a specific battery model under specific operating conditions reaches a threshold—the online model optimization process automatically initiates the parameter fine-tuning program for the thermal runaway grading model. The parameter fine-tuning program targets the judgment thresholds of the hierarchical decision tree within the thermal runaway grading model, such as the splitting threshold for sudden changes in characteristic gas concentration and the judgment critical value for the rate of temperature change.

[0082] Parameter fine-tuning is achieved using an incremental learning algorithm. This algorithm transforms accumulated bias data into new training samples without retraining with all historical data. The incremental learning algorithm analyzes the bias data, extracts data feature patterns that lead to prediction errors, and combines these patterns with the correct labels to form new training samples. These new training samples are used to locally adjust the hierarchical decision tree, such as fine-tuning the judgment threshold of a node or adding a new judgment branch under specific conditions. The incremental learning algorithm ensures that the model can learn from actual warning events, gradually adapting to new battery aging states or changes in the underground environment, thereby continuously improving the adaptability and warning accuracy of the thermal runaway classification model. The online model optimization process is a continuous, self-improving cycle that feeds back field experience to the model, enabling the warning system to evolve. The thermal runaway source analysis process and the online model optimization process together constitute the closed-loop learning mechanism of the warning system. The thermal runaway source analysis process uncovers deep-seated causes from single events, while the online model optimization process learns general patterns from multiple events. Their collaborative work significantly improves the intelligence and reliability of the entire underground lithium battery safety management system in coal mines.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for early warning of thermal runaway gas from lithium batteries in underground coal mines, characterized in that, include: Real-time acquisition of multi-dimensional monitoring data inside lithium battery packs in underground coal mines, including characteristic gas concentration, temperature change rate, voltage fluctuation value, and internal resistance offset; The multidimensional monitoring data is input into a pre-trained thermal runaway classification model. The characteristic gas concentration is assigned the highest priority through a dynamic weight allocation algorithm to generate a comprehensive risk score for the current data window. The construction process of the pre-trained thermal runaway grading model includes: Acquire a full-cycle data sequence of historical thermal runaway events, which includes characteristic gas concentration evolution curves, temperature gradient distributions, and electrical parameter degradation trajectories from the normal state to the thermal runaway outbreak. A temporal convolutional network is used to extract multi-scale features from the full-cycle data sequence, separating key modes that are strongly correlated with thermal runaway. A hierarchical decision tree is constructed based on the key modes. The root node of the hierarchical decision tree uses the abrupt change in characteristic gas concentration as the primary splitting condition, and the secondary nodes sequentially introduce the composite judgment logic of temperature change rate and voltage fluctuation value. The execution process of the dynamic weight allocation algorithm includes: The cumulative deviation of the characteristic gas concentration is calculated in real time, and the cumulative deviation is the multiple of the standard deviation between the current concentration value and the baseline concentration value; The weighting coefficient of the characteristic gas concentration in the comprehensive risk score is dynamically adjusted according to the cumulative deviation. When the cumulative deviation exceeds the first critical value, the weighting coefficient is increased to a preset upper limit. The acceleration of the rate of temperature change is monitored synchronously. If the acceleration is continuously positive and exceeds the second critical value, the weighting coefficient of the rate of temperature change is increased linearly. Based on the preset threshold range where the comprehensive risk score is located, a corresponding early warning level instruction is triggered, which includes a Level 1 attention instruction, a Level 2 intervention instruction, and a Level 3 emergency response instruction. When the Level 2 intervention command or Level 3 emergency response command is triggered, the location identification information of the lithium battery pack is automatically associated to generate an emergency response plan that includes the location identification.

2. The method for early warning of thermal runaway gas in underground lithium batteries in coal mines according to claim 1, characterized in that, The process of generating the comprehensive risk score includes: Each parameter in the multidimensional monitoring data is normalized to obtain a standardized parameter value; The initial risk value is obtained by multiplying the standardized parameter value by the corresponding dynamic weight coefficient and then summing the results. The current charge / discharge rate of the lithium battery pack is introduced as a correction factor to perform nonlinear correction on the initial risk value, and the final comprehensive risk score is output.

3. The method for early warning of thermal runaway gas in underground lithium batteries in coal mines according to claim 1, characterized in that, The triggering process for the warning level instruction includes: When the comprehensive risk score is in the first range, only a level 1 attention instruction is sent to the local monitoring terminal; When the comprehensive risk score enters the second interval, a secondary intervention command is simultaneously sent to the local monitoring terminal and the regional security system, and the forced heat dissipation device of the lithium battery pack is activated. When the comprehensive risk score reaches the third interval, a level-three emergency response instruction is broadcast to all emergency terminals downhole, and the power supply circuit of the lithium battery pack is cut off at the same time.

4. The method for early warning of thermal runaway gas in underground lithium batteries in coal mines according to claim 3, characterized in that, The process of generating the emergency response plan includes: The preset mine roadway topology map is retrieved based on the location identification information, and the nearest evacuation route and the coordinates of the explosion-proof facilities at the location of the lithium battery pack are marked. Based on the current distribution data of personnel underground, calculate the optimal evacuation plan and emergency material allocation route; The evacuation routes, coordinates of explosion-proof facilities, and emergency material allocation routes are integrated into a structured instruction set and sent to the target terminal.

5. The method for early warning of thermal runaway gas in underground lithium batteries in coal mines according to claim 4, characterized in that, It also includes a process for verifying the credibility of the multidimensional monitoring data: A physical correlation constraint is established between the concentration of the characteristic gas and the rate of temperature change. When the trend of the two changes violates the physical correlation constraint, redundant sensor data comparison is initiated. If the verification result of redundant data is confirmed to be abnormal, the current data window is marked as low confidence, and the data from the previous valid window is used for interpolation compensation.

6. The method for early warning of thermal runaway gas in underground lithium batteries in coal mines according to claim 5, characterized in that, It also includes the process of tracing the source of thermal runaway: After each Level 3 emergency response command is triggered, extract the sequence of changes in multi-dimensional monitoring data within the set time period prior to the trigger; The parameter type and time point at which the earliest anomaly occurred were identified based on the change sequence; The parameter types and time points are associated with the production batch and usage logs of the lithium battery pack to generate a thermal runaway cause analysis report.

7. A method for early warning of thermal runaway gas in underground lithium batteries in coal mines according to claim 6, characterized in that, It also includes the online optimization process of the model: Collect data on the discrepancy between the actual handling results and the model prediction results of all events that have triggered early warnings; When the cumulative frequency of the same type of deviation data exceeds a set threshold, the parameter fine-tuning of the thermal runaway classification model is initiated. The bias data is transformed into new training samples using an incremental learning algorithm, and the judgment threshold of the hierarchical decision tree is updated.

8. A thermal runaway gas early warning system based on lithium batteries in underground coal mines, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for early warning of thermal runaway gas in underground lithium batteries as described in any one of claims 1 to 7.

Images