Monitoring and sensing method for thermal runaway of lithium battery in underground coal mine

Abstract

The invention relates to the technical field of lithium battery safety, and discloses a coal mine underground lithium battery thermal runaway monitoring and sensing method, which comprises the following steps: deploying multi-element sensing hardware in a charging chamber, and simulating and optimizing gas sensor layout based on computational fluid dynamics; and a hierarchical distributed network is constructed through a cluster head election mechanism to realize real-time convergence and transmission of data. And the ground center executes Kalman filtering denoising and normalization processing on the data, and uses a time sequence potential Dirichlet allocation model to extract implicit theme features of the multi-source data. And then, outputting a thermal runaway risk probability mean value and a variance of quantitative uncertainty through a Bayesian deep learning model. And the system triggers graded early warning according to the information, and fire extinguishing and ventilation equipment is linked through an OPC UA protocol. According to the method, an uncertainty quantification mechanism is introduced, underground environment noise and real fault signals are effectively distinguished, and the accuracy and robustness of thermal runaway early warning are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery safety, in particular to a monitoring and sensing method for thermal runaway of lithium batteries in coal mine. BACKGROUND

[0002] Lithium-ion batteries have gradually been applied to trackless rubber-tyred vehicles and backup power systems in coal mines due to their high energy density and long cycle life. However, lithium batteries are prone to thermal runaway under abnormal conditions such as overcharging, external impact or internal short circuit, which generates a large amount of heat and releases flammable and explosive gases such as carbon monoxide and hydrogen. In the special environment of coal mine, early monitoring and safety warning for lithium battery thermal runaway is a key link to ensure the safety of coal mine production.

[0003] In the prior art, the monitoring of lithium batteries in coal mine mainly relies on the joint application of battery management system and external environment monitoring equipment. The commonly used method is to collect real-time data such as voltage, current, temperature of the battery, smoke concentration and characteristic gas concentration in the charging chamber, and transmit them to the ground monitoring center through wired or wireless network. The system generally sets fixed safety thresholds, when a single parameter or combined parameter exceeds the preset value, or the change rate of the parameter is abnormal, the system will determine that there is a risk and trigger an alarm signal, so as to realize the monitoring of the battery state.

[0004] However, the environment in coal mine is extremely complex, with high humidity, high dust and strong electromagnetic interference, and many other interference factors. The time series data collected by the sensor often accompanies significant random noise and non-stationary fluctuations. The existing monitoring methods are mostly based on deterministic threshold discrimination or traditional machine learning models, which lack the ability to deeply mine the potential time series theme features in multi-source heterogeneous data, and cannot quantify the uncertainty of the prediction results. This limitation makes it difficult for the system to effectively distinguish between false signals caused by environmental interference and real early signs of thermal runaway, and is prone to false alarms when facing complex working conditions, which cannot meet the actual needs of high reliability and accurate early warning in coal mine. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a monitoring and sensing method for thermal runaway of lithium batteries in coal mine, which solves the problems mentioned in the background art.

[0006] To achieve the above object, the present application is implemented by the following technical solutions: A monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines, comprising a test platform construction and data acquisition stage, by establishing a simulated charging chamber environment in underground coal mines, synchronously collecting BMS data, external sensor data, environmental data and video stream data, constructing a thermal runaway feature data set of lithium batteries in underground coal mines containing different battery states and environmental conditions, and providing a data basis for subsequent model training.

[0007] In the sensor network configuration stage, a hierarchical-distributed architecture is adopted to deploy the sensor network. Among them, the arrangement position of the gas sensor array is determined based on computational fluid dynamics (CFD) simulation. The simulation process adopts a standard k-ε turbulence model to calculate the diffusion path of the thermal runaway characteristic gas and the airflow field distribution, and the sensor is deployed upstream of the airflow, near the battery box pressure relief valve and on the top of the chamber. The determination of the cluster head node adopts a cluster head election mechanism, and the fitness value is calculated according to the residual power and received signal strength indication value of the node, and the node with the highest fitness value is selected to be responsible for data aggregation.

[0008] In the software system and early warning model implementation stage, the collected data is first preprocessed. Kalman filtering is used to remove random noise, and the state equation and the observation equation are used to describe the system state; by calculating the Kalman gain , the posterior state estimation and the posterior error covariance , the optimal estimation of the true measurement value is realized; then the min-max normalization is used to eliminate the dimension effect. The feature extraction process adopts the TLDA model. This model maps multi-source time series data into implicit topic distribution, and each topic corresponds to a different evolution stage of thermal runaway. The model assumes that each time window generates a topic distribution , and each data point is generated by the word distribution corresponding to the topic . By introducing the variational distribution , the lower bound of the evidence is maximized using variational inference:

[0009] By iteratively optimizing the variational parameters, the topic feature vector reflecting the internal structure of the data is extracted.

[0010] The risk assessment process adopts a Bayesian deep learning model. The model takes the topic feature vector as input, and its weight parameters and bias parameters follow a Gaussian prior distribution. In the reasoning process, the model outputs the mean and variance of the thermal runaway risk probability, where the mean is used for risk warning and the variance is used to quantify the prediction uncertainty, so as to provide confidence evaluation while outputting the risk probability.

[0011] During the early warning triggering and linkage mechanism phase, when the probability of thermal runaway risk exceeds the set threshold, a tiered early warning is executed. The first-level early warning activates the on-site audible and visual alarms, while the second-level linkage controls the surface and underground equipment via the OPCUA protocol to perform operations such as reporting information, cutting off power, activating fire extinguishing devices, and adjusting ventilation equipment.

[0012] A second aspect of the present invention provides a monitoring and sensing system for thermal runaway of lithium batteries in coal mines. This system is based on the above-described method and includes: The sensing layer hardware, deployed in the underground charging chamber of the coal mine, includes a battery management system for collecting internal battery parameters, an explosion-proof gas sensor array, a temperature sensor, a smoke detector, and an infrared camera located at key positions determined by CFD simulation. The network layer hardware consists of a self-organizing network composed of several cluster nodes and cluster head nodes, configured to aggregate and transmit data collected by the perception layer to the ground monitoring center. The cluster head nodes are elected based on fitness values. The data processing module, configured in the ground monitoring center, is used to perform time alignment, Kalman filtering for noise reduction, and normalization on the received data. The model inference module integrates a TLDA model and a Bayesian deep learning model; the TLDA model is configured to extract topic feature vectors from time-series data based on maximizing the lower bound of evidence; the Bayesian deep learning model is configured to output the mean and variance of the probability of thermal runaway risk based on Gaussian prior distribution and variational inference. The linkage control module is configured to trigger graded early warnings and equipment linkage control based on the comparison results between the probability of thermal runaway risk and the early warning threshold.

[0013] This invention optimizes sensor layout through CFD simulation, solving the problem of blind spots in gas monitoring caused by complex downhole gas flow; it eliminates environmental noise interference by using Kalman filtering algorithm; and by combining TLDA model and Bayesian deep learning model, it not only achieves effective fusion and feature extraction of multi-source heterogeneous data, but also effectively distinguishes between system interference and real thermal runaway signals by introducing uncertainty quantification mechanism, thereby improving the accuracy of monitoring and early warning.

[0014] This invention provides a method for monitoring and sensing thermal runaway of lithium batteries in underground coal mines. It has the following beneficial effects: 1. This invention achieves deep feature extraction and risk assessment of complex downhole time-series monitoring data through the joint application of a temporal latent Dirichlet assignment model and a Bayesian deep learning model. This method utilizes variational inference to uncover hidden theme patterns in the thermal runaway evolution process and simultaneously outputs risk probability and prediction uncertainty based on probability distribution. This mechanism can effectively distinguish between environmental noise interference and real thermal runaway signals, significantly reducing the false alarm rate while ensuring a high detection rate, thus solving the problem of weak generalization ability of existing single-threshold alarm methods in complex downhole environments.

[0015] 2. This invention optimizes the physical layout of the sensor array through computational fluid dynamics simulation technology, solving the problem of blind spots in gas monitoring caused by forced ventilation and complex spatial structures in underground charging chambers. This method uses a fluid turbulence model to deduce the diffusion path and accumulation area of ​​characteristic gases in thermal runaway, guiding the deployment of sensors upstream of the airflow and at key accumulation points. This layout significantly improves the system's ability to capture early-stage, minute leaks of gas, shortens the early warning response time, and ensures accurate detection and triggering of safety mechanisms in the early stages of thermal runaway diffusion.

[0016] 3. This invention improves the operational stability and data quality of the monitoring system in underground environments with high humidity, high dust levels, and electromagnetic interference by employing a hierarchical distributed sensor network architecture and a Kalman filter algorithm. This method utilizes an adaptive cluster head election mechanism to ensure network communication robustness, prevent data interruptions caused by single-point failures, and effectively removes random noise from the acquired signals through filtering algorithms. Combined with a tiered early warning and equipment linkage control strategy, it achieves closed-loop management from anomaly detection to proactive handling, thereby enhancing the overall safety protection level of underground charging facilities in coal mines. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the data fusion and early warning model of the present invention. Figure 2 This is a block diagram of the overall system architecture of the present invention; Figure 3 This is a schematic diagram of the sensor network layout in the charging chamber according to the present invention; Figure 4 This is a schematic diagram of the downhole lithium battery charging thermal runaway test platform of the present invention. Detailed Implementation

[0018] The technical solutions in 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.

[0019] Example: Please see the appendix Figure 1 -Appendix Figure 4 This invention provides a monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines, including the following hardware deployment: Basic conditions of the monitoring area.

[0020] The underground charging chamber in the coal mine was selected as the monitoring area. The spatial structure of the chamber conforms to the characteristics of the underground decentralized charging layout. There are environmental factors such as ventilation, airflow, humidity, and dust, and there may be flammable gases. The deployment environment needs to be adapted to such complex conditions.

[0021] Hardware configuration of the multi-signal monitoring system.

[0022] Internal monitoring hardware, including the battery management system (BMS).

[0023] Each charging cabinet is equipped with a mining-grade explosion-proof battery management system, the core function of which is to collect internal status parameters of the lithium battery in real time during the charging process. Data collected: current, voltage, cell surface temperature, cell internal temperature, state of charge (SOC); Acquisition frequency: Ensures real-time capture of parameter changes; Installation method: Integrated installation with lithium battery pack, communicating with external sensor network through explosion-proof interface, and data transmission adopts bus protocol adapted to downhole environment.

[0024] External monitoring hardware.

[0025] Temperature sensor: Explosion-proof PT100 sensor or explosion-proof thermocouple is selected. Multiple sensors are arranged in each charging cabinet and installed at key points on the surface of the battery box and near the charging interface. The probe is tightly attached to the surface of the battery box with thermally conductive silicone. The lead wire is led out through the explosion-proof junction box. The junction box has the corresponding protection level and temperature resistance performance. Gas sensor array: Each charging cabinet is equipped with one explosion-proof gas sensor array, including four core sensors: CO2, H2, CO, and VOCs. Specific selection requirements: The CO2 sensor adopts a high-sensitivity non-dispersive infrared type, which has excellent gas selectivity and long-term stability; the H2 sensor adopts a MEMS thermal conductivity type; the CO sensor adopts an electrochemical type; and the VOCs sensor adopts a semiconductor type. All gas sensors must be able to adapt to the background gas interference environment downhole. Array installation location: near the pressure relief valve of the battery box, the upper part of the charging cabinet, and the top of the chamber where combustible gases are likely to accumulate. Smoke detectors: Explosion-proof photoelectric or ionization smoke detectors are selected, with one unit installed above each charging cabinet. The installation height is adapted to the smoke rising and capturing requirements, and it has an anti-dust interference design. Infrared cameras: Explosion-proof infrared cameras are used, and the number deployed meets the requirement of full coverage of all charging points. They are installed in suitable locations in the chamber and support real-time monitoring of changes in battery appearance such as bulging and smoke.

[0026] Environmental monitoring hardware.

[0027] Each charging chamber is equipped with multiple sets of environmental monitoring sensors, which are respectively arranged on the air inlet side and air return side of the chamber entrance. Specific configuration: Temperature and humidity sensor: It has the corresponding measurement range and accuracy, the sampling frequency is adapted to the environmental parameter monitoring requirements, the protection level meets the high humidity environment downhole, and it has an anti-condensation design; Dust sensor: It adopts the laser scattering type, has the corresponding particle size and concentration measurement range, the sampling frequency is adapted to the dust concentration monitoring needs, and has a built-in automatic cleaning device to reduce the impact of dust adhesion.

[0028] Hierarchical-distributed sensor network hardware configuration: Perception layer hardware.

[0029] Composed of the aforementioned sensor nodes, all nodes adopt a mine-grade explosion-proof design with a protection level of no less than IP65. The shell material is corrosion-resistant and adaptable to the high humidity and dust environment underground. The nodes are powered by intrinsically safe power supplies to ensure long-term stable operation.

[0030] Network layer hardware.

[0031] Cluster head node: An embedded industrial control computer is selected, and the number of configurations is determined according to the distribution of charging points. Each cluster head node is responsible for aggregating sensor data from charging points in the corresponding area, has multi-protocol data access capabilities, and runs an adapted operating system. Cluster nodes: Employ low-power microcontrollers, with each sensor node corresponding to one cluster node, responsible for data acquisition and preliminary processing, and power supply supporting long-term battery life or charging battery life; Communication module: Supports multiple communication methods, including wired and wireless, between nodes within a cluster and the cluster head node. The appropriate type is selected based on the actual deployment distance underground to ensure stable communication. Industrial Ethernet Switch: Adopting a mining-grade explosion-proof design, the number of ports meets network access requirements and supports redundancy backup. Cluster head nodes connect to the mine's industrial ring network through the switch to realize data transmission to the ground monitoring center.

[0032] Application layer hardware.

[0033] The ground monitoring center is equipped with hardware such as servers, display devices, audible and visual alarms, and SMS modules. The servers need to have the computing power for model training and real-time inference, the audible and visual alarms need to have sufficient sound intensity and obvious visual prompts, and the SMS module needs to support sending early warning information to management personnel.

[0034] Based on the detailed hardware deployment described above, the specific implementation process of this technical solution is as follows: S1 test platform setup and data acquisition.

[0035] S1.1 Test Platform Construction.

[0036] A thermal runaway test platform simulating an underground charging chamber in a coal mine was built in a ground laboratory to ensure that the test environment was consistent with the actual underground conditions. The platform consisted of the following components: Simulated chamber: The size is adapted to the typical specifications of underground charging chambers, and it has the functions of adjusting temperature, humidity, wind speed and dust concentration, which can simulate complex underground environmental conditions; Charging system: Equipped with explosion-proof charging cabinets, mining lithium battery packs and programmable chargers consistent with those used underground. The chargers support overcharge mode and can precisely control thermal runaway trigger conditions. Safety devices: The simulation chamber is equipped with pressure relief valves, fire extinguishing devices, and gas collection devices to prevent the leakage of harmful gases and safety risks during the test; Data acquisition system: Selects multi-channel data acquisition equipment, configured with analog input module, digital input module, counter module, and video acquisition module, supports synchronous acquisition of all sensor data, BMS data, charger data and video stream, and has real-time data storage and offline export functions.

[0037] S1.2 Thermal runaway test and data acquisition.

[0038] Test parameter settings: By setting different wind speed gradients through a controllable ventilation system, the ambient temperature, humidity, and dust concentration were adjusted to different working conditions to cover common underground environmental conditions; Thermal runaway triggering: Start the charger to charge the lithium battery pack at a constant current. After charging to full charge, continue to trigger thermal runaway with different overcharge currents. Repeat the test multiple times for each condition. Data acquisition scope: The acquisition time is from the start of charging until the end of thermal runaway, covering the normal charging stage, the pre-thermal runaway stage, the thermal runaway outbreak stage, and the cooling stage. The acquired data includes internal data: BMS current, voltage, cell temperature, and SOC timing data; external data: timing data from external temperature sensors, gas sensor arrays, smoke detectors, and infrared camera video streams; and environmental data: temperature, humidity, and dust concentration timing data. Dataset preparation: After the experiment, the collected data were denoised, time-aligned, and format-standardized to form a dataset of thermal runaway characteristics of underground lithium batteries in coal mines with different battery states, environmental conditions, and overcharge currents, which was used for model training and validation.

[0039] S2 sensor network configuration.

[0040] S2.1 Network Structure Configuration.

[0041] Sensor networks adopt a layered-distributed architecture, consisting of a perception layer, a network layer, and an application layer. Specific configurations are as follows: The sensing layer consists of all explosion-proof sensor nodes. Each node has a built-in basic data preprocessing unit, supports plug-and-play, and the node number corresponds one-to-one with the charging point. Network layer: It adopts a self-organizing wireless network architecture of cluster head-internal nodes. The cluster head node is responsible for data aggregation, preprocessing, storage and uploading within the cluster, while the internal nodes are responsible for data collection and transmission. The network self-organizing protocol supports automatic node joining, dynamic cluster head election and automatic replacement of faulty nodes. Application layer: Includes ground monitoring center server and early warning software module, responsible for data reception, model inference, early warning issuance and linkage control.

[0042] S2.2 Cluster head election and data transmission.

[0043] Cluster head election mechanism: A fixed election cycle is set. Nodes within the cluster calculate their fitness value by monitoring their remaining power and received signal strength indicators. The fitness value is positively correlated with the remaining power and signal strength. The node with the highest fitness value is elected as the cluster head. After being elected, it sends a broadcast notification to other nodes in the cluster. After the other nodes confirm, they join the cluster and begin data transmission. Data transmission process: Cluster nodes collect data at a set frequency, and after simple filtering, send the data to the cluster head node. After receiving the data, the cluster head node performs alignment, filtering, and normalization preprocessing, and uploads it to the ground monitoring center through the mine industrial ring network at a set period. At the same time, it stores the raw data for a certain period locally to prevent data loss due to network interruption.

[0044] S2.3 Sensor array layout optimization.

[0045] The specific steps for optimizing sensor placement based on computational fluid dynamics (CFD) simulation are as follows: CFD simulation parameters: Professional fluid simulation software is used, the simulation area is the three-dimensional space of the charging chamber, the appropriate number of grids are divided, the standard k-ε turbulence model is adopted, and inlet velocity boundary, outlet pressure boundary, and wall no-slip boundary are set to simulate the characteristic gas diffusion path and airflow field distribution during thermal runaway; Layout scheme: Determine the placement of each sensor based on simulation results: Gas sensors: positioned upstream of the airflow, near the battery box pressure relief valve, and at the top of the chamber to ensure early detection of gas release signals; Temperature sensors: placed on hot spots on the surface of the battery box and in key areas with weak heat dissipation inside the charging cabinet; Smoke detector: Installed at the top center of the charging cabinet to ensure a rapid response when smoke rises; Environmental sensors: One set is installed on the air inlet side and one on the air return side to monitor background environmental parameters and calibrate sensor readings; Infrared cameras: installed in suitable locations within the chamber to ensure complete coverage of all charging cabinets without any blind spots.

[0046] Implementation of S3 software system and early warning model.

[0047] S3.1 software system architecture.

[0048] The monitoring center software adopts a B / S architecture, supports multi-terminal access, and its specific components are as follows: Vue.js front-end module development: Data display module: Real-time display of monitoring data from each sensor, sensor network operating status, and charging point status; Early warning module: Displays early warning information in the form of sound and light, and pop-up windows, including early warning location, early warning time, risk probability, confidence interval, and trigger parameters; System configuration module: Supports setting early warning thresholds, calibrating sensor parameters, configuring network parameters, and managing user permissions; Historical data query module: Supports querying historical data by time range, monitoring parameters, and charging points; reports can be exported; video playback is supported. Backend module Python Django development: Data receiving module: Supports multi-protocol data access, parses sensor data and stores it in the database; Data preprocessing module: Implements data alignment, filtering, and normalization functions; Model inference module: Integrates the TensorFlow deep learning framework, runs a latent Dirichlet Allocation (TLDA) Bayesian deep learning model, and outputs the probability of thermal runaway risk in real time; Linkage control module: It achieves linkage control with underground fire extinguishing devices, charging power supplies, and ventilation equipment through the OPCUA protocol.

[0049] S3.2 Data Preprocessing.

[0050] Data preprocessing includes three core steps: data alignment, filtering, and normalization, to ensure that the data quality meets the model input requirements. Data alignment: A timestamp matching method is adopted, using the synchronization clock provided by the network layer counter module as a reference, and all sensor data are sorted by timestamp to ensure that data at the same time point corresponds to the same row of records, eliminating timing misalignment caused by data transmission delay; Kalman filtering: Removes random noise from data, with the following formula: Equations of state:

[0051] Observation equation:

[0052] in, Let k be the system state vector at time k, i.e., the actual sensor measurement value. Here is the state transition matrix. For process noise, The data is the original sensor data collected at time k. For the observation matrix, To observe noise; Filtering update process: Prior state estimation:

[0053] Prior error covariance:

[0054] Kalman gain:

[0055] Posterior state estimation:

[0056] Posterior error covariance:

[0057] in, For the posterior state estimation at time k-1, Let the posterior error covariance be at time k-1. It is the identity matrix; min-max normalization: Transforms the filtered data to the range of 0-1 to eliminate the influence of units. The formula is:

[0058] in, The original data after filtering. The minimum value of this parameter in the dataset. This is the maximum value of the parameter in the dataset. This is the normalized data.

[0059] S3.3 Temporal Latent Dirichlet Allocation (TLDA) Model.

[0060] The TLDA model is used to extract thematic features from multi-source time-series data and capture the characteristic patterns of different stages of thermal runaway. Specific implementation details are as follows: Model assumptions and parameter settings.

[0061] Number of topics Based on the division of thermal runaway stages, the corresponding stages are: normal charging, precursor stage, early warning, violent reaction, and cooling stage. Hyperparameters: Word distribution hyperparameters Theme distribution hyperparameters All are set values; Timing window length: 60 seconds. Each window contains data points corresponding to the acquisition frequency. The window sliding step size ensures timing continuity.

[0062] Model generation process.

[0063] For each topic ( ), distribution of generated words , The characteristic distribution of each sensor parameter under this topic; For each timing window ( , (Total number of windows), generate topic distribution. , This represents the percentage of each topic within the window; For timing window Each data point in ( , (Number of data points per window), sampling topic That is, to determine the theme to which the data point belongs; For data points Sampling observations , This is the normalized sensor data.

[0064] Model inference (variational inference).

[0065] Introducing variational distribution ,in Thematic distribution Variational parameters (dimensions) ), Assigning topics Variational parameters (dimensions) The goal is to minimize the KL divergence between the variational distribution and the posterior distribution, which is equivalent to maximizing the lower bound of evidence. :

[0066] After unfolding:

[0067] in: (The difference between the prior and variational distributions of topic distributions); (Differences between the prior and variational distributions of topic assignment); (Likelihood contribution of observational data); (The difference between the prior and variational distribution of word distribution); Iterative optimization of variational parameters and Maximize the lower bound of evidence Finally, the topic distribution for each time series window is obtained. Extract topic feature vectors (dimension) ).

[0068] S3.4 Bayesian deep learning model.

[0069] Topic feature vectors extracted using the TLDA model Using the input as input, a Bayesian deep learning model is constructed, which outputs the probability of thermal runaway risk. The specific structure and implementation are as follows: Model structure.

[0070] Input layer: dimension The dimension of the topic feature vector; Hidden layers: 3 fully connected layers, with the number of neurons set to 64, 32, and 16 respectively, and the ReLU activation function is used. Each layer has a Dropout layer added to prevent overfitting; Output layer: Dimension 1, activation function is Sigmoid ( Output thermal runaway risk probability (0≤) ≤1).

[0071] Bayesian inference.

[0072] Prior distribution: all parameters of the model (Weights, biases) follow a Gaussian prior distribution ,in For setting value, It is the identity matrix; Likelihood function: Hypothesis label ( This indicates thermal runaway. (Indicating normality) follows a Bernoulli distribution, and the likelihood function is... ,in The input feature matrix; Posterior distribution: Using variational inference approximation, a variational distribution is introduced. ,in For variational mean, For variational variance (all are learnable parameters); Objective function: Maximize the lower bound of evidence ,in Let be the expectation of the likelihood function. Let KL be the divergence between the variational distribution and the prior distribution.

[0073] Model training and inference.

[0074] Training parameters: Stochastic gradient descent optimizer is used, with learning rate, batch size, and number of training epochs all set values. Early stopping strategy is used to prevent overfitting. Model validation: Five-fold cross-validation is used to ensure the model's generalization ability; Inference output: For new input features The model outputs the mean probability of thermal runaway risk. and variance : mean (Core early warning indicator, namely the probability of thermal runaway risk); variance (Quantification of prediction uncertainty); The confidence interval is set at 95%, that is .

[0075] S4 Early Warning Triggering and Linkage Mechanism.

[0076] S4.1 Warning threshold settings.

[0077] Based on experimental data and actual downhole needs, an early warning threshold is set when the probability of thermal runaway risk is high. An alert will be triggered if the threshold is exceeded.

[0078] S4.2 Tiered Early Warning and Coordination.

[0079] Level 1 warning, local response: When When the warning threshold is exceeded, the audible and visual alarm at the charging chamber site will be activated immediately, and the warning information will be displayed on the local display screen of the charging cabinet to remind on-site personnel to take emergency measures. Two-level linkage, remote device linkage: The system quickly completes the following operations via the OPCUA protocol: Report early warning information to the ground monitoring center, including location, time, risk probability, confidence interval, sensor data curve, and infrared video screenshot; Send warning text messages to management personnel, including warning locations, risk levels, and handling suggestions; The system is activated by linking the underground fire extinguishing device to the warning location; Disconnect the charging power supply and shut down the charging circuit corresponding to the warning point. Adjust ventilation equipment to increase the ventilation speed in the warning area and accelerate the diffusion of harmful gases.

[0080] S4.3 Continuous monitoring and reset.

[0081] The system updates the timing window according to the set cycle and recalculates the probability of thermal runaway risk. When the risk probability is consistently lower than the set threshold and the on-site personnel confirm that the risk has been eliminated, the warning status can be manually reset through the monitoring center software to restore the charging function. If the risk probability continues to rise, it will be upgraded to an emergency alarm and corresponding emergency measures will be activated.

[0082] S5 actual downhole deployment and operation verification.

[0083] S5.1 deployment and implementation.

[0084] Select a charging chamber at a suitable level underground in the coal mine for actual deployment. Specific steps: Hardware installation: Install the sensor cluster, cluster head node, environmental sensor, and infrared camera according to the hardware deployment requirements, ensuring that the sensor probes are firmly installed, the communication lines are properly protected, and the cluster head node is fixed in the explosion-proof control box on the side wall of the chamber; Network debugging: Start the sensor network, test the node network access, data transmission status, and communication stability, and calibrate the sensor parameters through the monitoring center software; Model Deployment: The trained TLDA-Bayes deep learning model is deployed to the ground monitoring center server, and the coal mine underground thermal runaway feature dataset is loaded for online fine-tuning to ensure that the model is adapted to the specific environment of the chamber.

[0085] S5.2 runtime verification.

[0086] During the formal operation of the system, multiple signals during the charging process are continuously monitored. Through data preprocessing, feature extraction, and model inference, thermal runaway risk warnings are achieved, warning events and handling are recorded, and the system's operational stability and warning effectiveness in the complex downhole environment are verified.

[0087] 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 monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines, characterized in that, Includes the following steps: S1 Test Platform Construction and Data Acquisition: A thermal runaway test platform simulating an underground charging chamber in a coal mine was constructed in a ground laboratory. The thermal runaway test platform includes a simulated chamber, a charging system, safety devices, and a data acquisition system. The data acquisition system collects internal, external, and environmental data during the thermal runaway test, and processes the collected data to form a thermal runaway characteristic dataset of lithium batteries in underground coal mines. S2 sensor network configuration: Configure a hierarchical-distributed sensor network, which includes a sensing layer, a network layer, and an application layer; The network layer adopts a self-organizing wireless network architecture of cluster head-intra-cluster nodes, and the perception layer is composed of explosion-proof sensor nodes; S3 software system and early warning model implementation: The ground monitoring center receives the data transmitted by the hierarchical-distributed sensor network and performs data preprocessing; The topic feature vector is extracted from the preprocessed data using the TLDA model; The topic feature vector is input into a Bayesian deep learning model, which outputs the probability of thermal runaway risk. S4 Early Warning Triggering and Linkage Mechanism: Set an early warning threshold. When the probability of thermal runaway exceeds the early warning threshold, execute graded early warning and linkage. S5 Actual Underground Deployment and Operation Verification: Select an underground charging chamber in a coal mine, perform hardware installation, network debugging, and model deployment. The model deployment includes deploying the Bayesian deep learning model to the ground monitoring center server and loading the underground lithium battery thermal runaway feature dataset for online fine-tuning.

2. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, In the S2 sensor network configuration, the arrangement of the gas sensor array in the explosion-proof sensor node is determined based on computational fluid dynamics (CFD) simulation. The specific steps of the computational fluid dynamics (CFD) simulation include: Using the standard k-ε turbulence model, inlet velocity boundary, outlet pressure boundary, and wall no-slip boundary are set to simulate the characteristic gas diffusion path and gas flow field distribution during thermal runaway. Based on the simulation results, the gas sensor array was arranged upstream of the airflow, near the battery box pressure relief valve, and at the top of the chamber.

3. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, In the S2 sensor network configuration, the cluster head node of the network layer is determined through a cluster head election mechanism; The specific steps of the cluster head election mechanism include: Nodes within a cluster calculate their fitness value by monitoring their remaining battery power and received signal strength indicators. The fitness value is positively correlated with the remaining battery power and the received signal strength indication value; The node with the highest fitness value within the cluster is selected as the cluster head node.

4. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, In the implementation of the S3 software system and early warning model, the data preprocessing includes data alignment, Kalman filtering, and min-max normalization. The specific process of Kalman filtering is as follows: Establish the state equation and the observation equation; Calculate the prior state estimate and prior error covariance using the posterior state estimate and posterior error covariance from the previous time step; Calculate the Kalman gain based on the prior error covariance, observation matrix, and observation noise. The prior state estimate is updated using the Kalman gain to obtain the posterior state estimate.

5. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, In the implementation of the S3 software system and early warning model, the implementation process of the TLDA model includes: The number of topics is determined based on the stages of thermal runaway; For each topic, generate a word distribution; For each time series window, generate a topic distribution; For each data point in the time series window, the topic to which the data point belongs is determined by sampling, and observations are sampled based on the topic to which the data point belongs. The observations are normalized sensor data. Variational inference introduces a variational distribution containing variational parameters of topic distribution and topic assignment, maximizes the lower bound of evidence, obtains the topic distribution for each time series window, and uses the topic distribution as the topic feature vector.

6. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, In the implementation of the S3 software system and early warning model, the Bayesian deep learning model includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is the same as the dimension of the topic feature vector; The hidden layer consists of three fully connected layers, with a Dropout layer added to each fully connected layer, and the ReLU function is used as the activation function. The activation function of the output layer is the Sigmoid function; All weight and bias parameters of the Bayesian deep learning model follow a Gaussian prior distribution.

7. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 6, characterized in that, The inference output process of the Bayesian deep learning model includes: Variational inference is used to approximate the posterior distribution of the Bayesian deep learning model, and a variational distribution is introduced, which includes variational mean and variational variance. For new input features, the Bayesian deep learning model outputs the mean and variance of the probability of thermal runaway risk, where the variance is a quantification of prediction uncertainty.

8. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, In the S4 early warning triggering and linkage mechanism, the tiered early warning and linkage includes a first-level early warning and a second-level linkage; The first-level early warning is as follows: when the warning threshold is exceeded, the audible and visual alarm at the underground charging chamber of the coal mine is activated, and the warning information is displayed on the local display screen of the charging cabinet. The secondary linkage is as follows: the system reports early warning information and infrared video screenshots to the ground monitoring center via the OPCUA protocol, sends early warning text messages to management personnel, activates the underground fire extinguishing device, cuts off the charging power, and adjusts the ventilation equipment.

9. The monitoring and sensing method for thermal runaway of lithium batteries in underground coal mines according to claim 1, characterized in that, The external data used in the construction and data acquisition of the S1 test platform includes external temperature sensor data, gas sensor array data, smoke detector data, and infrared camera video stream. The gas sensor array includes a CO2 sensor, an H2 sensor, a CO sensor, and a VOCs sensor; The CO2 sensor is a high-sensitivity non-dispersive infrared (NDIR) type. The H2 sensor is of the MEMS thermal conductivity type; The CO sensor is of the electrochemical type; the VOCs sensor is of the semiconductor type.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of a monitoring and sensing method for thermal runaway of lithium batteries in coal mines as described in any one of claims 1 to 9.

Images