A mine goaf fire extinguishing method based on image recognition

Abstract

The present application relates to the technical field of mine fire extinguishing, and discloses a mine goaf fire extinguishing method based on image recognition. The method acquires image data sequences of the mine goaf in real time, extracts initial space-time feature sequences from the image data sequences, and extracts fire coupling features therefrom; generates a multi-resolution fire feature distribution based on the fire coupling features and the resolution of the image data sequences, analyzes the time sequence of the distribution to obtain a fire risk spectrum; calculates the fire source position and the spreading trend using a fire spreading model according to the fire risk spectrum, identifies fire key points according to the spreading trend; predicts the fire propagation path based on the fire key points, generates an adaptive fire extinguishing threshold; and triggers the fire extinguishing operation in combination with the predicted fire propagation path and the adaptive fire extinguishing threshold. The method can capture goaf fire information in real time and accurately, scientifically predict the development of the fire, and make the fire extinguishing operation more targeted, and is suitable for efficient prevention and control of mine goaf fires.

Description

Technical Field

[0001] This invention relates to the field of mine fire extinguishing technology, specifically a method for extinguishing fires in mine goaf areas based on image recognition. Background Technology

[0002] Mining goafs, as special spaces formed after coal mining, have complex internal structures and contain a large amount of residual coal, making them highly susceptible to fire due to oxidation and heating. Once a fire breaks out, the flames can spread within the concealed spaces of the goaf, consuming large amounts of oxygen, producing toxic and harmful gases, threatening the lives of underground workers, damaging mine production facilities, causing huge economic losses, and even triggering more serious accidents such as gas explosions.

[0003] For the detection and extinguishing of fires in mine goaf areas, the industry has developed various technical solutions, but these solutions still have significant limitations in practical applications. In the fire detection stage, traditional methods mostly rely on manual inspections or fixed sensor monitoring. Manual inspections are limited by the complex terrain and harsh environment of goaf areas, making it difficult to comprehensively cover all areas. Furthermore, the inspection cycle is long, often failing to detect early fire hazards in time, causing the fire to miss the optimal control window. While fixed sensors can achieve real-time monitoring at fixed points, their monitoring range is limited. For large, irregular goaf areas, a large number of sensors are needed to achieve full coverage, increasing equipment costs and maintenance difficulty. Furthermore, sensor malfunctions or signal transmission interference can lead to missed or false alarms.

[0004] In fire suppression, existing technologies typically employ a uniform firefighting strategy after a fire is detected, lacking precise assessment of the specific fire situation. Due to the inability to accurately determine the fire source and its spread, firefighting operations are often indiscriminate, potentially leading to extinguishing agents failing to target the core fire area accurately, reducing extinguishing efficiency. Furthermore, the overuse of extinguishing agents can waste resources and even adversely affect subsequent treatment of the goaf. In addition, existing firefighting methods struggle to effectively predict fire propagation paths and cannot take preventative measures at critical points in the fire's spread, allowing the fire to easily escalate and increasing the difficulty of fire control.

[0005] As mining depths increase and goaf areas expand, the limitations of traditional fire detection and extinguishing methods become increasingly apparent, failing to meet the high demands of mine safety for fire prevention and control. Therefore, the industry urgently needs a technological solution capable of accurately detecting, analyzing in real-time, scientifically predicting, and efficiently extinguishing fires in mine goafs, to address the shortcomings of existing technologies in terms of timeliness of fire monitoring, accuracy of fire assessment, and targeted firefighting operations. Summary of the Invention

[0006] The purpose of this invention is to provide a method for extinguishing fires in mine goaf areas based on image recognition, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for extinguishing fires in mine goaf areas based on image recognition, the method comprising:

[0008] Real-time acquisition of image data sequences of the goaf in a mine; extraction of an initial spatiotemporal feature sequence from the image data sequence; extraction of fire coupling features from the initial spatiotemporal feature sequence;

[0009] Based on the fire coupling features and the resolution of the image data sequence, a multi-resolution fire feature distribution is generated; the time series of the multi-resolution fire feature distribution is analyzed to obtain a fire risk spectrum.

[0010] Based on the fire risk spectrum, a fire spread model is used to calculate the location of the fire source and the spread trend; based on the spread trend, key fire points are identified.

[0011] Based on the aforementioned key fire points, predict the fire propagation path; generate an adaptive fire extinguishing threshold.

[0012] Fire suppression operations are triggered based on the predicted fire propagation path and the adaptive fire suppression threshold.

[0013] Preferably, the step of extracting the initial spatiotemporal feature sequence from the image data sequence includes:

[0014] The image data sequence is standardized to obtain standardized image data;

[0015] The initial spatiotemporal feature sequence is obtained by extracting color features, temperature features, and smoke features from the standardized image data using an image feature extraction algorithm.

[0016] Preferably, the step of extracting fire coupling features from the initial spatiotemporal feature sequence includes:

[0017] A preliminary feature set, which couples color features, temperature features, and smoke features, is extracted from the initial spatiotemporal feature sequence;

[0018] When the feature dimension of the preliminary feature set is greater than a preset dimension threshold, the preliminary feature set is subjected to dimensionality reduction processing to obtain dimensionality-reduced feature data;

[0019] The time series of the reduced-dimensional feature data is analyzed, and the dynamic interaction features of color features, temperature features, and smoke features are enhanced to obtain the fire coupling features.

[0020] Preferably, generating a multi-resolution fire feature distribution based on the fire coupling features and the resolution of the image data sequence includes:

[0021] The goaf of the mine is divided into multiple sub-regions;

[0022] Construct multiple computation queues, each corresponding to a sub-region;

[0023] Based on the resolution of the image data sequence of each sub-region, determine the cell size, grid density, and number of features for each sub-region;

[0024] The multiple computing queues are controlled to generate target meshes for each sub-region according to the cell size, mesh density, and number of features of each sub-region;

[0025] Based on the fire coupling characteristics and the target grid of each sub-region, the multi-resolution fire feature distribution is obtained.

[0026] Preferably, controlling the plurality of computation queues to generate the target mesh for each sub-region according to the cell size, mesh density, and number of features of each sub-region includes:

[0027] The multiple computing queues are controlled to generate preliminary mesh surface information according to the cell size, mesh density, and number of features of each sub-region;

[0028] The multiple computing queues are controlled to generate 3D meshes for each sub-region based on the mesh surface information and obtain mesh volume information;

[0029] Based on the mesh information, obtain the vertices and volume elements of each 3D mesh; and number the vertices and volume elements of each 3D mesh to obtain a unique identifier for each vertex and a unique identifier for each volume element.

[0030] The multiple computing queues are controlled to communicate synchronously so that the volume units and unique identifiers between adjacent sub-regions are projected onto each other to obtain projection information;

[0031] The 3D mesh of each sub-region is optimized based on the projection information to obtain the target mesh of each sub-region.

[0032] Preferably, the step of analyzing the time series of the multi-resolution fire feature distribution to obtain the fire risk spectrum includes:

[0033] Based on the multi-resolution fire feature distribution, a time series analysis algorithm is used to extract the feature change trend;

[0034] Calculate the probability distribution of fire occurrence based on the aforementioned characteristic change trends;

[0035] By combining historical fire data, the probability distribution is corrected to generate the fire risk spectrum.

[0036] Preferably, the step of calculating the fire source location and spread trend using a fire spread model based on the fire risk spectrum includes:

[0037] Construct a three-dimensional spatial model of the mine goaf;

[0038] External incentives are constructed based on the aforementioned fire risk spectrum;

[0039] Based on the three-dimensional spatial model and the external excitation, fire spread simulation is performed in combination with preset boundary conditions to obtain the fire source location and spread trend.

[0040] Analyze the temporal and spatial changes in the location and spread trend of the fire source;

[0041] Based on the aforementioned time and spatial changes, a dynamic fire spread map is obtained.

[0042] Preferably, predicting the fire propagation path based on the fire key points includes:

[0043] Based on the aforementioned key fire points and combined with the topological information of the mine goaf, each key fire point is mapped to three-dimensional spatial coordinates.

[0044] Spatial interpolation algorithms are used to interpolate discrete fire key points to generate preliminary fire propagation paths;

[0045] Based on the preliminary fire propagation path, a time series prediction algorithm combined with historical propagation data is used to dynamically correct the fire propagation path, resulting in the predicted fire propagation path.

[0046] Preferably, the generation of the adaptive fire extinguishing threshold includes:

[0047] Based on historical normal image data, a normal working condition generation model is trained to generate simulated data that conforms to the normal operating characteristics of the goaf area in the mine.

[0048] Construct a real-time discrimination model by inputting the predicted fire propagation path and the simulation data generated by the normal operating condition generation model;

[0049] The boundary features between normal and abnormal states are learned through dynamic game theory, and the discrimination result is output.

[0050] Based on the discrimination results, a clustering algorithm is used to divide the discrimination results into multiple levels, and combined with the dynamic changes of the predicted fire propagation path, the adaptive fire extinguishing threshold is generated.

[0051] Preferably, triggering the fire extinguishing operation based on the predicted fire propagation path and the adaptive fire extinguishing threshold includes:

[0052] The real-time fire characteristics are compared with the adaptive fire extinguishing threshold.

[0053] Based on the comparison results, a decision-making model is used to dynamically adjust the fire extinguishing level.

[0054] Based on the adjusted fire suppression level, the corresponding fire suppression operation is triggered.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] From the perspective of fire detection timeliness, this method, by acquiring real-time image data sequences of the mine goaf, breaks through the limitations of traditional manual inspections and fixed sensor monitoring. Traditional methods often suffer from monitoring lag or incomplete coverage, while real-time image data sequences can continuously capture environmental changes within the goaf. By extracting initial spatiotemporal feature sequences from the image data and further extracting fire coupling features, abnormal signals can be detected in the early stages of a fire, preventing the expansion of fire hazards due to untimely monitoring and buying valuable time for subsequent fire prevention and control.

[0057] Regarding the accuracy of fire situation assessment, this method generates a multi-resolution fire feature distribution based on fire coupling characteristics and image data sequence resolution, enabling a comprehensive characterization of fire features at different scales. Traditional detection methods often rely on single-dimensional monitoring data, which can easily lead to misjudgments or omissions due to incomplete information. In contrast, the multi-resolution fire feature distribution can cover the characteristics of fires at different spatial ranges and time stages. Combined with the analysis of time series to obtain a fire risk spectrum, it can clearly present the development status of the fire, providing a comprehensive basis for accurate fire situation assessment and reducing control deviations caused by misjudgments.

[0058] In terms of fire source location and spread trend analysis, this method, based on the fire risk spectrum, uses a fire spread model to calculate the location and spread trend of the fire source and identify key fire points, effectively solving the problem of accurately locating the fire source in traditional firefighting methods. Traditional firefighting operations often suffer from low efficiency in the use of extinguishing agents because the core area of ​​the fire source cannot be found. This method, however, can clearly identify the location of the fire source and, by identifying key fire points, provides clear targets for subsequent prevention and control. Furthermore, predicting the fire propagation path based on key fire points breaks through the limitation of traditional methods that cannot predict the direction of the fire in advance, allowing for early knowledge of the possible direction and area of ​​fire spread and preventing the disorderly spread of the fire.

[0059] Regarding the targeting and adaptability of fire suppression operations, this method generates an adaptive fire suppression threshold, which, combined with the predicted fire propagation path, triggers fire suppression operations, overcoming the problem of fixed strategies and lack of flexibility in traditional fire suppression. Environmental conditions and fire development stages vary in different goaf areas, and fixed fire suppression thresholds and strategies are difficult to adapt to complex and changing situations, potentially leading to untimely or excessive fire suppression. In contrast, the adaptive fire suppression threshold can dynamically adjust according to the actual fire situation, ensuring appropriate fire suppression measures are taken at different stages of fire development. This guarantees fire suppression effectiveness while avoiding waste of extinguishing agents. Furthermore, by combining the fire propagation path, fire suppression operations can be precisely applied to critical areas, improving fire suppression efficiency and reducing the threat of fire to mine facilities and personnel safety. Attached Figure Description

[0060] Figure 1 This is a schematic diagram illustrating the working principle of the image recognition-based fire extinguishing method for mine goaf areas according to the present invention.

[0061] Figure 2 A flowchart illustrating the workflow for extracting initial spatiotemporal feature sequences from image data sequences;

[0062] Figure 3 A flowchart illustrating the workflow for generating multi-resolution fire feature distributions;

[0063] Figure 4 This is a flowchart illustrating the workflow for defining the fire risk spectrum. Detailed Implementation

[0064] 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.

[0065] Please see Figure 1 This invention provides a method for extinguishing fires in mine goaf areas based on image recognition, the method comprising:

[0066] Image data sequences are acquired in real time by image acquisition equipment deployed in the goaf of the mine. These sequences are then transmitted to a processing unit, from which an initial spatiotemporal feature sequence is extracted. This initial spatiotemporal feature sequence is processed by a feature processing module to extract fire coupling features. Based on the fire coupling features and the resolution of the image data sequence, a multi-resolution fire feature distribution is generated. The time series of this multi-resolution fire feature distribution is analyzed to obtain a fire risk spectrum. Based on the fire risk spectrum, a fire spread model is used to calculate the location and spread trend of the fire source. According to the spread trend, key fire points are identified. Based on these key fire points, the fire propagation path is predicted. An adaptive extinguishing threshold is generated. Based on the predicted fire propagation path and the adaptive extinguishing threshold, extinguishing operations are triggered. This method achieves automatic monitoring and intelligent fire suppression of mine fires.

[0067] Example 1: See Figure 2 High-definition visible light cameras and infrared thermal imagers deployed on the roof walls of mine roadways and around goaf areas acquire continuous image data of the mine environment at a rate of 25 frames per second. These devices, through explosion-proof encapsulation and adaptive lighting systems, ensure clear image sequences even in the low-light, high-dust environment of the mine. The image data is transmitted to the image processing server at the ground monitoring center via the mine's industrial Ethernet. The image data sequence first undergoes standardization processing. Due to uneven lighting and dust interference in the mine environment causing image quality fluctuations, the standardization process includes three main steps: denoising the image sequence using a median filtering algorithm to eliminate pixel anomalies caused by dust reflection; adjusting the image brightness distribution through histogram equalization to enhance the visibility of details in dark areas; and uniformly scaling images of different resolutions and scales to a standard size of 1024×768 pixels and converting them to the RGB color space. After these operations, standardized image data is obtained, providing a consistent data foundation for subsequent feature extraction.

[0068] Image feature extraction algorithms are employed to extract multiple features from standardized image data. Color feature extraction calculates the RGB color values ​​of each pixel in the image and generates a 256-level color histogram. Simultaneously, color moment features are calculated, including the first moment (mean), second moment (variance), and third moment (skewness). These features effectively characterize the red-orange hue distribution and anomalous color shifts in the flame region. Temperature feature extraction is based on infrared thermal imaging data. By analyzing the temperature matrix data output by the thermal imager, the temperature value of each pixel region is calculated, and a temperature distribution map is generated. Temperature gradient features are also calculated, and the Sobel operator is used to detect regions of abrupt temperature changes. These features are used to identify abnormal temperature rise points. Smoke feature extraction uses the Local Binary Pattern (LBP) algorithm to analyze image texture and identify the unique softened edge texture features of smoke. Combined with optical flow methods, motion vectors of pixels between consecutive frames are calculated to detect the diffusion pattern of smoke. All these features are organized in a time series to form an initial spatiotemporal feature sequence, which contains time-synchronized data for color, temperature, and smoke features.

[0069] When extracting fire-coupled features from the initial spatiotemporal feature sequence, the three types of features—color, temperature, and smoke—are first aligned by timestamps and combined into a multi-dimensional preliminary feature set. The dimensionality of this feature set depends on the number of features and the size of the time window; for example, when using a 30-second time window, the feature dimension may exceed 1000. A preset dimensionality threshold of 500 dimensions is set. When the dimensionality of the preliminary feature set exceeds this threshold, dimensionality reduction processing is initiated. Dimensionality reduction employs principal component analysis (PCA) to extract the top 50 principal components with the largest eigenvalues ​​by calculating the covariance matrix of the feature vectors. These principal components retain more than 95% of the variance information in the original feature set, forming the dimensionality-reduced feature data.

[0070] Time series analysis of dimensionality-reduced feature data is performed using a Long Short-Term Memory (LSTM) network model. This model takes continuous feature data as input via a sliding window and learns the temporal dependencies between features. Dynamic interactions between color, temperature, and smoke features are captured through the model's hidden states. For example, when the color feature shows an increase in red hue while the temperature feature exhibits a gradient increase, the model strengthens the weight of the feature for that time period. An attention mechanism is further employed to enhance these interactive features. This mechanism calculates the importance scores of features at different time points, highlighting periods with abnormal multi-feature collaboration, ultimately outputting enhanced fire coupling features. These features comprehensively reflect the multimodal signal interaction patterns during a fire, providing a highly discriminative feature representation for subsequent fire identification.

[0071] The entire implementation process is achieved through a dedicated image processing module integrated into the computing unit of the mine monitoring system. This module employs a multi-threaded parallel processing approach to ensure real-time processing of image data streams. The feature extraction and processing algorithms are developed based on the PyTorch framework, supporting GPU-accelerated computation to meet the real-time requirements of mine monitoring. The processed fire coupling features are then transmitted to the subsequent analysis module to generate a multi-resolution fire feature distribution.

[0072] Example 2: See Figure 3 The goaf in a mine is first divided into multiple sub-regions based on its spatial structure and monitoring needs. The division process considers the actual geometry of the goaf, the roadway distribution pattern, and the deployment location of image acquisition equipment. For example, in a typical coal mine goaf, the division might include the main haulage roadway area, the return air roadway area, the working face goaf area, and the connecting roadway area. The extent of each sub-region is defined using a three-dimensional coordinate system, and its boundary coordinates are stored in a spatial configuration database. The functional characteristics and fire risk differences of each region are fully considered during the division process to ensure that each sub-region has relatively independent characteristic representation.

[0073] Multiple computing queues are constructed to achieve parallel processing, with each queue dedicated to computing tasks in a specific sub-region. These queues are deployed in a distributed computing cluster, with each queue consisting of a group of computing nodes. These nodes are equipped with high-performance graphics processors (GPUs) capable of processing large-scale spatial data in parallel. The mapping between computing queues and sub-regions is managed through a configuration table, ensuring accurate task allocation and system scalability. Each computing queue operates independently but maintains a communication connection with the master node to receive computing instructions and transmit results.

[0074] The grid parameters for each sub-region are determined based on the resolution of the image data sequence. The resolution information of the image data sequence is obtained from the specifications of the image acquisition equipment, including spatial resolution and thermal resolution. For high-resolution areas, such as the main roadway area near the image acquisition equipment, a smaller unit size is set, typically 0.5 m × 0.5 m; for low-resolution areas, such as deep areas of the goaf far from the camera, a larger unit size is set, up to 2 m × 2 m. The grid density is dynamically adjusted according to the number of features; dense grids are used in areas with abundant features, and sparse grids are used in areas with sparse features. The number of features is obtained through statistical analysis of fire coupling features and is determined by calculating the spatial distribution density of the feature values.

[0075] Multiple computation queues are controlled to initially generate mesh surface information according to the parameter requirements of each sub-region. Each computation queue reads the spatial extent data of its corresponding sub-region and generates a planar mesh based on the element size parameters. The mesh surface information includes the vertex coordinates of each mesh surface, the connectivity between surface elements, and the orientation information of the surface. The Delaunay triangulation algorithm is used during the generation process to ensure the quality and uniformity of the mesh. Each surface element is assigned a unique identifier, and its spatial location attributes are recorded.

[0076] Based on the generated surface information, the computation queue further generates 3D meshes for each sub-region. The 3D mesh is constructed by stretching the planar mesh vertically to form volume elements; the stretching height is determined based on the mine roadway height data. The mesh volume information includes the volume calculation of each volume element, its spatial coordinates, and its adjacency relationships with other volume elements. Hexahedral element structures are used to generate the volume elements, ensuring computational stability and accuracy. Each volume element contains detailed geometric attribute records.

[0077] Based on the mesh information, the system acquires vertex and volume element data for each 3D mesh. Vertices are numbered using a spatial partitioning encoding method, dividing the 3D space into several cubic regions, with vertices within each region numbered according to their spatial coordinates. Volume elements are numbered systematically according to their generation order and spatial location. Unique identifiers are generated using a hash algorithm to ensure that each vertex and volume element has a unique identity throughout the system. These identifiers use a 64-bit encoding format and include information such as region code, type identifier, and sequence number.

[0078] Controlling synchronous communication between multiple computation queues is a crucial step in achieving grid system unification. Synchronous communication is implemented through a message passing interface, establishing high-speed network connections between computation nodes. Volume cell data between adjacent sub-regions are projected onto each other using a boundary matching algorithm. The projection process includes coordinate transformation, data mapping, and consistency checks. Projection information records the correspondence between volume cells in adjacent regions, the matching accuracy of overlapping areas, and the connection status at boundaries. This information is stored in a shared memory area and accessible to all computation nodes.

[0079] The 3D mesh of each sub-region is optimized based on the projection information. The optimization process includes adjusting the position coordinates of vertices at the boundaries to smoothly connect meshes in adjacent regions; modifying the connection relationships of volume elements to eliminate mesh overlaps or gaps; and unifying the numbering system to ensure consistency of identifiers across regions. The optimization algorithm employs an iterative relaxation method to progressively improve mesh quality. The final target meshes for each sub-region are obtained, which maintain the characteristics of each region while ensuring the integrity and consistency of the entire goaf mesh system.

[0080] The target grid system provides a spatial framework for the distribution of multi-resolution fire features. Fire coupling features are mapped to corresponding volume elements, forming a feature distribution map with spatial resolution. This distribution map can accurately reflect the fire feature status of different areas in the mine, providing a precise spatial data foundation for subsequent fire risk analysis. The entire implementation process achieves efficient processing of large-scale spatial data through a distributed computing architecture, meeting the dual requirements of real-time performance and accuracy for mine fire monitoring.

[0081] Example 3: See Figure 4 An improved time-series processing algorithm was employed to analyze the time series of multi-resolution fire feature distributions. This algorithm uses a sliding window approach to process feature data, with the window size set to 60 seconds based on the characteristics of the mining environment, and a sliding step of 5 seconds. Within each window, statistical measures of the fire feature values ​​were calculated, including the mean, variance, and trend slope. Feature change trends were extracted using a polynomial fitting method, employing a quadratic polynomial to fit the feature changes within the window to obtain the coefficient parameters of the trend line. These parameters reflect the acceleration and direction of feature value changes over time, and are used to identify abnormal rapid change patterns.

[0082] The probability distribution calculation for fire occurrence is based on characteristic change trends. The system establishes a probability model that considers the joint distribution of multiple characteristic parameters. For each spatial unit, its fire occurrence probability value is calculated. The probability calculation employs a historical statistical method, where the deviation of the current characteristic value from historical normal values ​​is quantified as a probability index. During the calculation process, different characteristics are assigned different weights, with temperature typically having a higher weight than other characteristics. The probability value is calculated using the following formula:

[0083]

[0084] in: This represents the probability of a fire occurring in the k-th spatial unit at time t, where α is the proportionality coefficient and w i Let i be the weight of the i-th feature. Let i be the eigenvalue of the k-th unit at time t. This represents the historical mean of the i-th feature of the unit. Here, is the corresponding historical standard deviation, and n is the total number of features. This formula combines the anomalies of multiple features into a single probability value, reflecting the likelihood of a fire occurring.

[0085] The probability distribution is corrected by incorporating historical fire data, including records of past fire events and their corresponding feature data. The correction process uses a Bayesian update method, combining the prior probability distribution with new observation data to obtain the posterior probability distribution. The corrected probability distribution more accurately reflects the actual fire risk, generating the final fire risk spectrum. The fire risk spectrum is represented as a three-dimensional probability field, with each spatial unit bearing a timestamp-marked probability value, forming four-dimensional risk distribution data.

[0086] Based on the fire risk spectrum, a fire spread model is used to calculate the location and spread trend of the fire source. First, a three-dimensional spatial model of the mine goaf is constructed, based on mine geological mapping data and real-time monitoring data. The model includes detailed information such as roadway geometry, ventilation network, and coal face distribution. The model achieves centimeter-level accuracy, accurately reflecting the complex structure of the mine space. The three-dimensional model is built using the finite element method, discretizing the mine space into numerous small elements, each with material properties and environmental parameters. External excitation inputs are constructed based on the fire risk spectrum. The probability values ​​in the fire risk spectrum are converted into heat release rate parameters, with high-risk areas corresponding to higher heat release rates. External excitations also include environmental parameters such as ventilation intensity and oxygen concentration, which are obtained in real-time from the mine environmental monitoring system. These external excitations are applied to the three-dimensional spatial model in the form of field variables, driving the fire spread simulation process.

[0087] Fire spread simulation was conducted based on a three-dimensional spatial model and external excitation, combined with preset boundary conditions. These boundary conditions included thermal boundary conditions of the tunnel walls, mass flow boundary conditions of ventilation openings, and chemical reaction boundary conditions of the coal face. Computational fluid dynamics was employed to solve the energy equation, mass conservation equation, and momentum equation. The simulation process considered physical phenomena such as turbulent combustion, radiative heat transfer, and multiphase flow. Numerical calculations yielded key parameters such as the temperature field, flue gas concentration field, and flame propagation velocity, thereby determining the fire source location and spread trend. The temporal and spatial variations of the fire source location and spread trend were analyzed. Temporal variation analysis calculated the fire spread rate and intensity change rate by comparing simulation results at different time steps. Spatial variation analysis used gradient calculation methods to identify the main direction of fire spread and obstruction areas. The variation analysis also considered the influence of ventilation airflow on the fire and the constraint effect of the tunnel structure on flame propagation.

[0088] Based on temporal and spatial variations, a dynamic fire spread map is generated. This map visually displays the fire's development, including the current fire zone extent, predicted spread path, and marked hazardous areas. The map uses a color-coding system: red indicates high-temperature areas, orange indicates the frontal spread, and yellow indicates potentially hazardous areas. The dynamic map updates over time, reflecting real-time changes in the fire situation and providing intuitive reference for firefighting decisions. The entire implementation process is achieved through a high-performance computing platform, ensuring the timeliness and accuracy of the simulation calculations and meeting the real-time requirements of mine fire emergency response.

[0089] Example 4: Based on the fire spread trend analysis results, the system identified multiple key fire points. These key points include confirmed fire sources, high-temperature accumulation areas, abnormal smoke concentration points, and the location of the spread front. Each key point is marked with a timestamp and feature description, including parameters such as temperature value, smoke concentration, and rate of change. The identification of key points adopts a multi-feature fusion algorithm, which comprehensively evaluates multiple indicators such as temperature gradient, smoke diffusion pattern, and color feature changes.

[0090] Each key fire point was mapped to a three-dimensional spatial coordinate system. The mapping process was based on the mine's global coordinate system, established using a total station and covering the entire mine area. The position and orientation parameters of each image acquisition device were precisely calibrated to ensure accurate conversion from image coordinates to spatial coordinates. The mapping algorithm employed perspective projection transformation, combined with camera intrinsic and extrinsic parameter matrices, to calculate the precise location of the key points in three-dimensional space. The mapped coordinates included components in the X, Y, and Z directions, accurate to the centimeter level.

[0091] By combining the topological structure information of the mine's goaf, spatial analysis is performed on the mapped key points. The topological structure information is extracted from the mine's digital twin model, including data on roadway connections, ventilation paths, and obstacle distribution. The system establishes a topological graph, where nodes represent key spatial locations and edges represent connecting paths. Graph theory algorithms are used to analyze the topological structure and identify potential fire propagation channels and obstruction zones.

[0092] Spatial interpolation algorithms are employed to process discrete fire critical points. Considering the non-uniformity of the spatial distribution of critical points, Kriging interpolation is selected for spatial filling. The interpolation process considers spatial autocorrelation, and the spatial autocorrelation structure is described based on a variogram model. The interpolation results generate a continuous fire risk surface, representing the fire development trend of the entire goaf. By extracting contour lines from the risk surface, preliminary fire propagation paths are obtained, representing the most likely direction of fire development.

[0093] The initial fire propagation path needs further refinement to improve accuracy. A time-series prediction algorithm is used to process the path data and establish a time-series model of path evolution. The algorithm input includes the current path morphology, historical path change trends, and time series of environmental parameters. A Long Short-Term Memory (LSTM) network is used to capture the long-term dependencies of path changes. The network is trained based on historical fire propagation data to learn the fire propagation patterns under different mine environments. The refinement process also incorporates real-time environmental parameters, such as ventilation speed and humidity changes, which are obtained in real-time from the mine environmental monitoring system.

[0094] The process of generating adaptive fire extinguishing thresholds begins with the collection and analysis of historical normal image data. A large amount of image data is collected during normal mine operation, including scenes under different seasons and production intensities. This data is cleaned and labeled to form a training dataset. A generative adversarial network (GAN) is used to train a normal operating condition generation model. The generator learns the data distribution of normal states, while the discriminator distinguishes between real normal data and generated data. After adversarial training, the generator can produce simulated data that conforms to the characteristics of normal mine operation. A real-time discriminative model is constructed to process the current fire state data. The discriminative model adopts a deep neural network architecture, with inputs including predicted fire propagation path features and simulated data generated by the normal operating condition generation model. The model is designed as a two-stream structure: one stream processes real-time fire data, and the other processes normal state data, learning abnormal patterns through feature comparison. The network output is a discriminant score, representing the degree of deviation between the current state and the normal state.

[0095] The discrimination process is optimized through a dynamic game mechanism, employing a reinforcement learning framework. The agent learns strategies to distinguish between normal and abnormal states under different mine environments. The reward function is designed based on discrimination accuracy, encouraging the agent to discover subtle anomalies. The game process takes place in a simulated environment, where the agent interacts with the environment to learn the optimal discrimination strategy. After training, the system can accurately identify early signs of fire. Hierarchical clustering is used to divide the discrimination scores into multiple risk levels, each corresponding to a different degree of fire risk. Cluster centers are adjusted based on historical fire data to ensure the rationality of the level division. Combining the dynamic changes in predicted fire propagation paths, the risk levels are dynamically adjusted to form the final adaptive fire extinguishing threshold. Table 1 shows the main parameter characteristics of key fire points.

[0096] Table 1: Characteristics of Key Fire Parameters

[0097]

[0098] The entire implementation process is achieved through a distributed computing platform, with multiple computing nodes collaboratively handling spatial mapping, path prediction, and threshold generation tasks. The system adopts a microservice architecture, with each functional module deployed independently and exchanging data via message queues. A real-time data processing pipeline ensures that the entire process from key point identification to threshold generation is completed within seconds, meeting the real-time requirements of mine fire prevention and control. The generated predicted fire propagation paths and adaptive extinguishing thresholds provide decision-making support for subsequent firefighting operations, forming a complete fire emergency response closed loop.

[0099] Example 5: Real-time fire characteristics are acquired through a multimodal sensor network deployed across various areas of the mine. This network includes high-definition visible light cameras, infrared thermal imagers, temperature sensors, and smoke concentration detectors. These sensors collect environmental data at a sampling frequency of 30 times per second, and the data is transmitted to the central processing unit via industrial Ethernet. Real-time fire characteristics include multi-dimensional parameters such as temperature spatiotemporal distribution, smoke diffusion patterns, and flame optical characteristics. Each characteristic parameter carries precise timestamps and spatial coordinate information, forming a complete real-time monitoring dataset.

[0100] Real-time fire characteristics are compared and analyzed against adaptive extinguishing thresholds. The comparison process is performed on a distributed computing platform using streaming data processing technology. The system establishes a feature-threshold mapping table, matching each fire characteristic parameter with its corresponding threshold range. The comparison algorithm calculates the deviation between real-time characteristic values ​​and thresholds, considering multiple dimensions such as absolute difference, relative rate of change, and duration. For temperature characteristics, the system compares the current temperature value with the threshold temperature and analyzes whether the temperature rise rate exceeds a preset limit. For smoke characteristics, it assesses whether the smoke concentration reaches the threshold level and detects whether the smoke diffusion rate is abnormal. Regarding optical characteristics, it analyzes whether the flame flicker frequency and color characteristics conform to fire characteristic patterns.

[0101] Based on the comparison results, a decision model is used to dynamically adjust the fire extinguishing level. This model is built upon a fuzzy logic system, and its inputs include the deviation of various features, spatial distribution characteristics, and temporal duration. An internal rule base is established, containing hundreds of empirical rules derived from mine fire prevention expert knowledge and historical accident cases. The rule base covers various possible scenario combinations, such as high temperature accompanied by dense smoke, and rapid temperature rise accompanied by specific smoke patterns. The decision model processes real-time data through an inference engine and outputs preliminary fire extinguishing level recommendations. The fire extinguishing level is divided into three levels: Level 1 is a warning observation state, Level 2 is a local intervention state, and Level 3 is a full fire extinguishing state.

[0102] The decision-making model also integrates information on predicted fire propagation paths for comprehensive judgment. The system analyzes the degree of agreement between the current fire location and the predicted path, assessing whether the fire development conforms to the expected pattern. When the real-time fire deviates from the predicted path, the decision-making model activates an adaptive adjustment mechanism to reassess the fire suppression level. Simultaneously, the model considers real-time changes in the mine environment, such as the status of the ventilation system, personnel distribution, and equipment operating status. These environmental parameters are obtained in real-time from the mine monitoring system. Integrating all input information, the decision-making model ultimately outputs an optimized fire suppression level command.

[0103] Based on the adjusted fire extinguishing level, corresponding fire extinguishing operations are triggered. The system establishes a tiered response mechanism, with a set of operation instructions corresponding to each fire extinguishing level. When the fire extinguishing level is Level 1, the system initiates an early warning procedure, including activating audible and visual alarm devices, fine-tuning the ventilation system, and prompting evacuation of personnel in the area. Simultaneously, the monitoring frequency is increased, and the data collection density is improved.

[0104] When the fire extinguishing level is raised to Level II, the system initiates local intervention measures, including activating the sprinkler system in designated areas, releasing inert gas to locally suppress the fire, and adjusting ventilation direction to control smoke spread. Operational instructions are issued to on-site execution equipment via the industrial control network; each instruction includes specific location coordinates, operational intensity, and duration. The system monitors the intervention effect in real time, evaluates the effectiveness of the measures through feedback sensors, and dynamically adjusts operational parameters as needed.

[0105] When the fire level reaches level three, the system initiates a comprehensive fire suppression procedure. This procedure includes simultaneously activating fire suppression equipment in multiple areas, implementing large-scale spraying and inert gas injection, activating the emergency ventilation system to change airflow direction, and coordinating with the mine safety system to implement area isolation. All operations are executed according to the pre-determined emergency plan, but the order and intensity of operations will be intelligently adjusted based on the real-time fire situation. The system evaluates the fire suppression effect in real time through an execution feedback mechanism and dynamically optimizes subsequent operational strategies.

[0106] The entire fire suppression operation triggering process adopts a closed-loop control architecture. The system monitors the execution effect of fire suppression measures in real time, collects environmental data after fire suppression through a sensor network, and assesses changes in fire characteristics. This feedback data is sent to the decision model for a new round of analysis, forming a continuous monitoring-decision-execution-feedback loop. This closed-loop control ensures that fire suppression operations can be adjusted in real time according to changes in the fire situation, achieving dynamically optimized fire prevention and control.

[0107] The system also establishes an operation log recording mechanism, recording in detail the data status, decision-making process, and execution results at each decision-making point in time. This log data is used for post-event analysis and system optimization, providing data support for improving firefighting strategies. Simultaneously, the system includes a manual intervention interface, allowing professionals to take over control when necessary, ensuring the safe and reliable operation of the system.

[0108] The implementation process employs redundant design and fail-safe mechanisms. All critical components have backup systems, automatically switching to the backup system when the primary system detects an anomaly. The execution equipment utilizes a dual-control design, allowing for both automatic system control and manual on-site operation, ensuring effective firefighting operations under any circumstances. The entire system, through multi-layered safety protection measures, guarantees the reliability and effectiveness of mine fire emergency response.

[0109] 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.

[0110] 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 extinguishing fires in mine goaf areas based on image recognition, comprising: Real-time acquisition of image data sequences of mined-out areas in mines; Extract an initial spatiotemporal feature sequence from the image data sequence; Extract fire coupling features from the initial spatiotemporal feature sequence; Based on the fire coupling features and the resolution of the image data sequence, a multi-resolution fire feature distribution is generated; The fire risk spectrum is obtained by analyzing the time series of the multi-resolution fire feature distribution. Based on the fire risk spectrum, the fire source location and spread trend are calculated using a fire spread model; Based on the described spread trend, identify the key points of the fire; Based on the aforementioned key fire points, predict the fire propagation path; Generate adaptive fire suppression threshold; Fire suppression operations are triggered based on the predicted fire propagation path and the adaptive fire suppression threshold.

2. The method for extinguishing fires in mine goaf areas based on image recognition as described in claim 1, characterized in that, The step of extracting the initial spatiotemporal feature sequence from the image data sequence includes: The image data sequence is standardized to obtain standardized image data; The initial spatiotemporal feature sequence is obtained by extracting color features, temperature features, and smoke features from the standardized image data using an image feature extraction algorithm.

3. The method for extinguishing fires in mine goaf areas based on image recognition as described in claim 2, characterized in that, The step of extracting fire coupling features from the initial spatiotemporal feature sequence includes: A preliminary feature set, which couples color features, temperature features, and smoke features, is extracted from the initial spatiotemporal feature sequence; When the feature dimension of the preliminary feature set is greater than a preset dimension threshold, the preliminary feature set is subjected to dimensionality reduction processing to obtain dimensionality-reduced feature data; The time series of the reduced-dimensional feature data is analyzed, and the dynamic interaction features of color features, temperature features, and smoke features are enhanced to obtain the fire coupling features.

4. The method for extinguishing fires in mine goaf areas based on image recognition as described in claim 3, characterized in that, The step of generating a multi-resolution fire feature distribution based on the fire coupling features and the resolution of the image data sequence includes: The goaf of the mine is divided into multiple sub-regions; Construct multiple computation queues, each corresponding to a sub-region; Based on the resolution of the image data sequence of each sub-region, determine the cell size, grid density, and number of features for each sub-region; The multiple computing queues are controlled to generate target meshes for each sub-region according to the cell size, mesh density, and number of features of each sub-region; Based on the fire coupling characteristics and the target grid of each sub-region, the multi-resolution fire feature distribution is obtained.

5. A method for extinguishing fires in mine goaf areas based on image recognition as described in claim 4, characterized in that, The control of the multiple computation queues to generate target meshes for each sub-region according to the cell size, mesh density, and number of features of each sub-region includes: The multiple computing queues are controlled to generate preliminary mesh surface information according to the cell size, mesh density, and number of features of each sub-region; The multiple computing queues are controlled to generate 3D meshes for each sub-region based on the mesh surface information and obtain mesh volume information; Based on the mesh information, obtain the vertices and volume elements of each 3D mesh; and number the vertices and volume elements of each 3D mesh to obtain a unique identifier for each vertex and a unique identifier for each volume element. The multiple computing queues are controlled to communicate synchronously so that the volume units and unique identifiers between adjacent sub-regions are projected onto each other to obtain projection information; The 3D mesh of each sub-region is optimized based on the projection information to obtain the target mesh of each sub-region.

6. A method for extinguishing fires in mine goaf areas based on image recognition as described in claim 5, characterized in that, The analysis of the time series of the multi-resolution fire feature distribution to obtain the fire risk spectrum includes: Based on the multi-resolution fire feature distribution, a time series analysis algorithm is used to extract the feature change trend; Calculate the probability distribution of fire occurrence based on the aforementioned characteristic change trends; By combining historical fire data, the probability distribution is corrected to generate the fire risk spectrum.

7. A method for extinguishing fires in mine goaf areas based on image recognition as described in claim 6, characterized in that, The calculation of the fire source location and spread trend using a fire spread model based on the fire risk spectrum includes: Construct a three-dimensional spatial model of the mine goaf; External incentives are constructed based on the aforementioned fire risk spectrum; Based on the three-dimensional spatial model and the external excitation, fire spread simulation is performed in combination with preset boundary conditions to obtain the fire source location and spread trend. Analyze the temporal and spatial changes in the location and spread trend of the fire source; Based on the aforementioned time and spatial changes, a dynamic fire spread map is obtained.

8. A method for extinguishing fires in mine goaf areas based on image recognition as described in claim 7, characterized in that, The prediction of fire propagation path based on the aforementioned fire key points includes: Based on the aforementioned key fire points and combined with the topological information of the mine goaf, each key fire point is mapped to three-dimensional spatial coordinates. Spatial interpolation algorithms are used to interpolate discrete fire key points to generate preliminary fire propagation paths; Based on the preliminary fire propagation path, a time series prediction algorithm combined with historical propagation data is used to dynamically correct the fire propagation path, resulting in the predicted fire propagation path.

9. A method for extinguishing fires in mine goaf areas based on image recognition as described in claim 8, characterized in that, The generation of the adaptive fire extinguishing threshold includes: Based on historical normal image data, a normal working condition generation model is trained to generate simulated data that conforms to the normal operating characteristics of the goaf area in the mine. Construct a real-time discrimination model by inputting the predicted fire propagation path and the simulation data generated by the normal operating condition generation model; The boundary features between normal and abnormal states are learned through dynamic game theory, and the discrimination result is output. Based on the discrimination results, a clustering algorithm is used to divide the discrimination results into multiple levels, and combined with the dynamic changes of the predicted fire propagation path, the adaptive fire extinguishing threshold is generated.

10. A method for extinguishing fires in mine goaf areas based on image recognition as described in claim 9, characterized in that, The step of triggering a fire suppression operation based on the predicted fire propagation path and the adaptive fire suppression threshold includes: The real-time fire characteristics are compared with the adaptive fire extinguishing threshold. Based on the comparison results, a decision-making model is used to dynamically adjust the fire extinguishing level. Based on the adjusted fire suppression level, the corresponding fire suppression operation is triggered.

Images