Coal gangue mountain internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing

Abstract

The invention relates to the technical field of mine safety and environmental protection, and discloses a coal gangue mountain internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing, and the method comprises the steps: controlling an unmanned aerial vehicle to collect visible light and thermal infrared images along a simulated flight route, and carrying out the atmospheric correction; constructing a target gangue dump three-dimensional geometric model, and mapping the thermal infrared image to generate a three-dimensional thermal texture model; a pre-trained 3DU-Net deep neural network is utilized to extract surface temperature field spatial features, and internal high-temperature area voxel distribution is inverted and output; and according to the voxel distribution, generating a three-dimensional visual early warning graph and marking a dangerous area. According to the method, the nonlinear mapping relation between the surface thermal texture and the internal heat source is established through deep learning, terrain interference is eliminated in combination with a ground-imitating flight strategy, non-contact accurate positioning and quantitative grading early warning of hidden fire sources in the gangue dump are achieved, and a scientific basis is provided for fire prevention and extinguishing treatment.

Description

Technical Field

[0001] This invention relates to the field of mine safety and environmental protection technology, specifically a three-dimensional inversion method for internal fire sources in coal gangue piles based on UAV remote sensing. Background Technology

[0002] Spontaneous combustion of coal gangue piles is a long-standing major safety hazard and source of environmental pollution in the mining industry. During the accumulation of coal gangue, the residual coal and pyrite, among other combustible materials, undergo an exothermic oxidation reaction with oxygen. The accumulated heat causes the temperature to rise, ultimately leading to spontaneous combustion. Accurate and timely detection of the location and development of fire sources within the gangue pile is crucial for the implementation of fire prevention and extinguishing projects.

[0003] For a long time, monitoring methods for spontaneous combustion areas in coal gangue hills have mainly relied on contact or close-range manual operations, but these methods have many limitations in practical applications. Traditional manual on-site inspections depend primarily on the sensory experience of inspectors, who judge the fire situation by observing surface smoke, smelling odors, or measuring temperature by hand. This method is not only highly subjective and inefficient, but also limited by terrain conditions, making it difficult to cover steep slopes or areas with developed cracks. More seriously, this method requires personnel to enter dangerous environments, facing extremely high safety risks such as high temperatures, landslides, and poisoning from toxic gases, and it is also difficult to detect hidden underground fire sources.

[0004] To obtain quantitative data, existing technologies have attempted to employ methods such as deploying fixed temperature probes or drilling gas detection. However, embedding sensors inside or on the surface of loose and continuously settling rock formations is not only difficult and costly to construct, but also makes the equipment highly susceptible to damage from high-temperature corrosion or rock deformation, resulting in extremely difficult maintenance. Furthermore, monitoring at fixed points cannot reflect the continuous temperature field distribution across the entire rock formation, creating blind spots. While laboratory analysis of carbon monoxide, ethylene, and other indicators by drilling shallow holes to extract gas can determine the spontaneous combustion process, this process involves long sampling and analysis cycles, poor timeliness, and shallow gases are easily affected by environmental wind speed, air humidity, and atmospheric pressure, leading to distorted test results.

[0005] In summary, existing monitoring methods generally suffer from high operational risks, low efficiency, and discrete and incomplete data, making it difficult to achieve early, quantitative, and three-dimensional visualization-based accurate early warning of hidden high-temperature areas inside coal gangue piles by penetrating complex surface environmental interference. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a three-dimensional inversion method for internal fire sources in coal gangue piles based on UAV remote sensing. This method solves the problems of high risks associated with manual inspections, incomplete coverage of fixed measuring points, discontinuous data, and difficulty in early and accurate location and quantitative inversion of hidden internal fire sources in existing coal gangue pile spontaneous combustion monitoring technologies.

[0007] To achieve the above objectives, the present invention provides a three-dimensional inversion method for internal fire sources of coal gangue mountains based on UAV remote sensing, which mainly includes four stages: data acquisition, model construction, intelligent inversion, and early warning visualization.

[0008] During the data acquisition phase, UAVs equipped with visible light and thermal infrared imaging devices were used to perform air-to-ground collaborative operations. Given the complex terrain and significant elevation differences of the coal gangue hill, a terrain-following flight path was established for the UAVs. This path required the UAVs to maintain a constant relative altitude to the surface of the target gangue hill and to cover the survey area in a zigzag pattern. During the operation, the forward and lateral overlap of adjacent visible light and thermal infrared images was strictly controlled to ensure that the image data met the accuracy requirements for 3D reconstruction. Simultaneously, to eliminate thermal interference caused by solar shortwave radiation reflection and non-uniform surface heating, the data acquisition window was limited to nighttime or early morning. While acquiring image data, the ambient temperature and humidity were monitored in real time. These environmental parameters, combined with an atmospheric transmission model, were used to perform radiometric correction on the thermal infrared image data to restore the true surface radiance and temperature.

[0009] In the model building phase, a high-precision 3D geometric model of the target waste rock pile was constructed using photogrammetry with high overlap visible light imagery. Based on this, atmospherically corrected thermal infrared imagery was precisely mapped onto the surface of the 3D geometric model using feature point matching technology. This process achieved the conversion of 2D planar thermal imagery data into 3D spatial data, generating a 3D thermal texture model containing terrain geometry information and surface temperature distribution information.

[0010] In the intelligent inversion stage, a deep neural network model is established and utilized to mine the nonlinear mapping relationship between the surface temperature field and the internal heat source. This deep neural network model adopts the 3DU-Net architecture, which includes an encoder path and a decoder path. The encoder path is responsible for multi-scale downsampling of the input 3D thermal texture model to extract the spatial distribution features of the surface temperature field, including the geometry, continuity, dispersion, and edge gradient of high-temperature regions. The decoder path upsamples and restores the extracted features, and fuses feature information from different levels through skip connections, ultimately reconstructing the output 3D temperature field matrix (voxel distribution) inside the target waste rock pile.

[0011] To ensure the generalization ability and accuracy of the inversion model, the training data for this deep neural network model comes from multiphysics numerical simulations. A virtual 3D model of a waste rock pile is constructed, and various internal heat source boundary conditions with different depths, shapes, and intensities are set according to the heat conduction equation to calculate the corresponding virtual surface temperature field. The virtual surface temperature field is used as the input feature, and the corresponding real distribution of internal heat sources is used as the output label for supervised training of the network until the loss function converges. In the actual simulation process, a preprocessing stage is also included to calculate the background average temperature of the 3D thermal texture model and set a temperature anomaly threshold accordingly. Only regions of interest exceeding the threshold are input into the network, reducing computational load and minimizing background noise interference.

[0012] In the early warning visualization phase, the voxel distribution data of high-temperature areas output by the deep neural network model is analyzed to extract the temperature values ​​and spatial depth coordinates of each voxel. Based on the temperature values, different areas are divided into high-risk, medium-risk, and low-risk levels. In the 3D visualized early warning map, different colors are used to render areas of each risk level, with the geometric center depth of high-risk areas being highlighted. This result directly displays the 3D spatial location of the internal fire source, providing quantitative coordinate guidance for subsequent fire prevention and extinguishing projects such as grouting and drilling.

[0013] Furthermore, this invention introduces a multi-temporal dynamic verification and closed-loop correction mechanism. By repeatedly probing the same target at different time points and comparing the inversion results across multiple periods, regions with overlapping spatial locations and consistently stable temperatures are identified, eliminating transient spurious anomalies caused by sudden environmental changes. Simultaneously, after grouting or drilling, actual internal temperature measurement data is obtained. This measured data is compared with the inversion results to calculate errors, and the measured data is added to the training database for incremental model training, continuously optimizing the model's inversion accuracy.

[0014] Through the above-mentioned technical solution, this invention overcomes the limitation that traditional infrared thermometry can only reflect the shallow surface temperature, establishes an intelligent inversion channel from surface thermal texture to internal temperature field, and realizes qualitative identification, quantitative inversion and location early warning of hidden fire sources inside coal gangue mountains.

[0015] This invention provides a three-dimensional inversion method for identifying internal fire sources in coal gangue piles based on UAV remote sensing. It offers the following advantages: 1. This invention establishes a nonlinear mapping relationship from surface thermal texture to internal temperature field by constructing an intelligent inversion model based on deep neural networks. Unlike traditional infrared thermal imagers that can only detect shallow surface temperature anomalies, this system can deduce the three-dimensional voxel distribution of deep heat sources through surface texture features (such as the continuity and dispersion of temperature gradients), thereby effectively discovering deep spontaneous combustion cores in a latent or concealed state, improving the depth and accuracy of fire source identification.

[0016] 2. This invention employs a UAV terrain-following flight strategy based on a digital elevation model and a specific nighttime / early morning data collection window. Terrain-following flight ensures that the thermal infrared sensor maintains a constant relative altitude on the surface of a rocky hill with significant topographic relief, thereby obtaining a uniform ground resolution (GSD). Combined with the specific time window and atmospheric radiation correction algorithm, the effects of solar shortwave radiation reflection and atmospheric attenuation are effectively eliminated, ensuring that the collected temperature data accurately reflects the surface thermal radiation state and avoiding misjudgments caused by slope changes or ambient lighting.

[0017] 3. This invention generates a visualized three-dimensional voxel risk model from the inversion results and automatically calculates the geometric center coordinates and vertical depth of the high-risk combustion core. Through intuitive red / yellow / green graded early warning rendering and precise depth annotation, the system can directly provide accurate three-dimensional spatial parameters (such as borehole depth and location) for subsequent fire prevention and extinguishing projects such as grouting and drilling. This changes the previous work mode of relying solely on empirical estimations based on two-dimensional planar diagrams, improving the targeting and safety of the control project. Attached Figure Description

[0018] Figure 1 A schematic diagram of the overall structure of the air-ground coordinated fire source detection system inside a coal gangue pile provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the drone's terrain-following flight trajectory and temperature measurement point distribution provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the method for inverting and classifying the spontaneous combustion fire source of a coal gangue pile according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the temperature distribution inversion results on the surface and inside of the gangue hill generated in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Reference Figure 1 This invention provides a system for identifying potentially spontaneously combustible hazardous areas inside a coal gangue pile. The system mainly includes: an unmanned aerial vehicle (UAV) flight platform, a ground control terminal, and a data processing workstation.

[0021] The unmanned aerial vehicle (UAV) flight platform is configured as a transport vehicle for performing aerial data acquisition tasks. In this embodiment, the UAV flight platform is a multi-rotor aircraft, specifically a quadcopter industrial-grade UAV. This UAV flight platform has vertical takeoff and landing (VTOL) and hovering capabilities, enabling it to adapt to complex, unstructured terrain environments formed by coal gangue deposits. The UAV flight platform establishes a two-way data connection with the ground control terminal via a wireless communication link.

[0022] The UAV flight platform carries a multimodal sensor payload, which includes a visible light imaging device and a thermal infrared imaging device. Both the visible light and thermal infrared imaging devices are fixed to the bottom of the UAV flight platform's fuselage via a three-axis stabilization gimbal. The three-axis stabilization gimbal isolates fuselage vibrations and controls the imaging device's shooting angle, ensuring that the imaging line of sight is perpendicular to the horizontal plane or the surface of the target being measured.

[0023] The visible light imaging device is configured with a high-resolution CMOS sensor camera to acquire visible light images of the target coal gangue hill surface. The visible light images contain red, green, and blue (RGB) three-channel color information, which is used for subsequent 3D geometric modeling and texture mapping.

[0024] The thermal infrared imaging device is configured with an uncooled microbolometer infrared camera to acquire thermal infrared images of the target coal waste hill surface. The device operates in the long-wave infrared range of 8 to 14 micrometers. Its thermal sensitivity (NETD) is less than 50 mK, enabling it to detect minute temperature differences on the ground surface.

[0025] The data output by a thermal infrared imaging device is a raw digital image containing radiance information. The thermal infrared imaging device converts the received infrared radiation energy into digital grayscale values ​​(DN values). This conversion process follows the radiative response equation, whose mathematical expression is: ; In the formula, This represents the radiance at the entrance pupil, measured in watts per steradian per square meter. This represents the digital grayscale value of a pixel in a thermal infrared image. This represents the gain coefficient of a thermal infrared imaging device; This represents the bias coefficient of the thermal infrared imaging device. Gain coefficient. and bias coefficient These are the inherent parameters of the equipment determined through blackbody radiation calibration experiments.

[0026] The UAV flight platform also integrates a high-precision positioning module. This module employs real-time dynamic carrier phase differential (RTK) technology. It receives satellite navigation signals and combines them with differential data from a ground reference station to calculate the UAV flight platform's three-dimensional spatial coordinates at the time of exposure. These coordinates include longitude, latitude, and altitude, with a positioning accuracy better than 10 centimeters. These coordinates are then incorporated into the metadata of the visible light and thermal infrared images, serving as a positional reference for subsequent data processing.

[0027] The ground control terminal is configured as a mobile computing device with a display interface. It is equipped with flight control software. The ground control terminal is used to plan the flight path of the UAV flight platform, set flight altitude, flight speed, and overlap parameters, and upload the generated flight commands to the UAV flight platform. The ground control terminal is also used to receive flight status data and image transmission signals transmitted back from the UAV flight platform in real time.

[0028] The data processing workstation is configured as a high-performance desktop computer or server. It reads visible light and thermal infrared images from the UAV flight platform's memory via a data interface. The workstation is equipped with a central processing unit (CPU) and a graphics processing unit (GPU). The GPU features parallel computing units supporting the CUDA architecture, used for executing 3D reconstruction algorithms and inference operations on deep neural network models.

[0029] The data processing workstation stores and runs a 3D modeling module and an intelligent inversion module. The 3D modeling module is configured to perform aerial triangulation and multi-view stereo matching algorithms to process visible light images to construct a 3D geometric model, and to map thermal infrared images onto the 3D geometric model to generate a 3D thermal texture model. The intelligent inversion module is configured to load a pre-trained deep neural network model. This deep neural network model, based on the 3DU-Net network architecture, is used to read the 3D thermal texture model data and output voxel distribution data for the high-temperature areas inside the waste rock pile.

[0030] The system in this embodiment also includes environmental monitoring equipment. This equipment is located near the ground control terminal and is used to measure the ambient temperature and humidity at the work site. The environmental monitoring equipment transmits the measurement data to a data processing workstation via wired or wireless means for atmospheric transmission correction calculations on the thermal infrared images.

[0031] In this embodiment, the core algorithm running the intelligent inversion module is a deep neural network model. This model is specifically constructed as a three-dimensional U-Net (3DU-Net) network architecture. This architecture is dedicated to processing three-dimensional data with voxel features and can establish a three-dimensional nonlinear mapping relationship from the surface temperature field of the waste rock pile to the internal temperature field.

[0032] Before inputting the data into the deep neural network model, the system first performs voxelization discretization on the generated 3D thermal texture model. The system defines a 3D bounding box encompassing the entire target waste rock pile space and divides this bounding box into segments of size [missing information]. The four-dimensional tensor, in which For batch size, , , These correspond to voxel resolutions for height, width, and depth, respectively. For voxels located on the surface of the waste rock pile, their values ​​are initialized to the temperature values ​​mapped from the thermal infrared image; for voxels located inside the waste rock pile and in the air domain, their values ​​are initialized to zero or a preset background value. This four-dimensional tensor constitutes the input layer of the deep neural network model.

[0033] The 3DU-Net network architecture consists of two parts: an encoder path and a decoder path. The encoder path is configured as a feature extraction network to progressively reduce the spatial resolution of the data and extract high-dimensional abstract features; the decoder path is configured as a feature reconstruction network to progressively restore the spatial resolution and infer the internal temperature distribution.

[0034] The encoder path contains four cascaded downsampling modules. Each downsampling module consists of two consecutive 3D convolutional layers and one 3D max-pooling layer. The 3D convolutional layers are configured with several 3×3×3 convolutional kernels. The convolutional kernels perform sliding operations along the length, width, and height dimensions of the input data to capture the local correlation features and gradient changes of the surface temperature field in space. The mathematical expression for the 3D convolution operation is: ; In the formula, Indicates the first Convolutional layers in spatial coordinates Output feature value at, Indicates the first The input feature values ​​of the layer Indicates the first The relative positions of the convolutional kernels The weight parameters at that location, Indicates the size of the convolution kernel. Indicates the first Layer bias terms. This represents a non-linear activation function.

[0035] In this embodiment, the nonlinear activation function The Modified Linear Unit (ReLU) is chosen to introduce nonlinearity and suppress the gradient vanishing problem; its expression is: Each convolutional layer is followed by a batch normalization layer to normalize the feature distribution and accelerate model convergence. The stride of the 3D max pooling layer is set to 2 to adjust the spatial size of the feature map. , , The network is reduced to half its original size in each of the three directions, while the number of feature channels is increased. The number of feature channels increases by doubling layer by layer with increasing network depth in the sequence of 64, 128, 256, and 512.

[0036] The decoder path contains four cascaded upsampling modules. Each upsampling module consists of a 3D transposed convolutional layer and two consecutive 3D convolutional layers. The 3D transposed convolutional layer is configured with a stride of 2 to increase the spatial resolution of the feature map by a factor of 2.

[0037] At each level of the decoder path, a skip connection is set. Skip connections are used to concatenate the shallow feature map with high spatial resolution generated at the corresponding level in the encoder path with the feature map with deep semantic information generated after upsampling in the decoder path. This concatenation operation is performed along the feature channel dimension, and the fused feature map simultaneously contains detailed texture features of the surface temperature field and semantic features of the location of deep heat sources.

[0038] The output layer of the deep neural network model is configured as a 1×1×1 three-dimensional convolutional layer. This layer maps multi-channel feature vectors to continuous single-channel values, which represent the predicted temperature value at the corresponding voxel space coordinates. The final output data is a three-dimensional temperature field matrix with the same size as the input voxel grid, where each element directly corresponds to the temperature at a specific depth inside the waste rock pile.

[0039] Since it is impossible to directly obtain a large amount of real data on the distribution of spontaneous combustion sources inside coal gangue piles as training samples for deep neural network models, this embodiment uses a multiphysics numerical simulation method to construct a virtual simulation dataset. This dataset is generated using numerical calculation software based on finite element analysis (FEA) and aims to establish a physical causal relationship between a defined internal heat source and the surface temperature field.

[0040] The construction process begins with establishing a thermophysical mathematical model of the waste rock pile based on heat conduction theory. In a three-dimensional Cartesian coordinate system, the heat transfer within the waste rock pile follows a heat conduction differential equation with an internal heat source. For the solution domain... For any point within the range, its governing equation is expressed as: ; In the formula, This indicates the density of coal gangue material, expressed in kilograms per cubic meter. This indicates the specific heat capacity of coal gangue material, expressed in joules per kilogram (Kelvin). This represents the temperature at various points in space, expressed in Kelvin. Indicates time, The thermal conductivity of coal gangue material is expressed in watts per meter Kelvin. This represents the Hamiltonian operator, corresponding to the calculation of the spatial gradient. This indicates the heat generation rate of the internal heat source, measured in watts per cubic meter.

[0041] To simulate a realistic physical environment, the system operates at the surface boundary of the solution domain. A third type of boundary condition (Robin boundary condition) is set to describe the convective heat transfer process between the surface of the waste rock pile and the surrounding ambient air. The mathematical expression for this boundary condition is: ; In the formula, This represents the outward normal direction vector of the boundary surface; This represents the convective heat transfer coefficient, measured in watts per square kelvin. This coefficient is dynamically set based on natural wind speed and surface roughness, with a range of 5 to 25. This indicates the calculated temperature of the surface of the waste rock pile; Represents ambient air temperature, set as a constant value or a function that changes over time.

[0042] When constructing the training dataset, the system generates a large number of virtual 3D mesh models of waste rock piles with different geometries. For each virtual model, the system randomly assigns physical property parameters, including density. Specific heat capacity and thermal conductivity These parameters follow a normal distribution within the preset range of coal gangue properties to simulate the heterogeneity of actual geological conditions.

[0043] More importantly, the system sets up various types of internal heat source conditions within the virtual model to correspond to... The spatial distribution of the values. The types of internal heat sources are configured as follows: point heat sources, used to simulate local high-temperature points in the early stages of spontaneous combustion, with a spatial distribution radius set from 0.5 meters to 2 meters; strip heat sources, used to simulate combustion zones developing along fissures or oxygen supply channels, set as columnar regions with specific lengths and orientations; and irregular block heat sources, used to simulate large-scale deep combustion zones. Simultaneously, the system randomly sets the burial depth of the heat sources, covering a range from 0.5 meters to 20 meters below the surface, and discretely samples the central temperature intensity of the heat sources between 300 degrees Celsius and 1000 degrees Celsius.

[0044] By solving the above governing equations, the system calculates the full-field temperature distribution of each virtual model under steady-state or transient conditions. The system extracts two parts from the calculation results as training data pairs: The first part consists of temperature distribution data on the outer surface of the model, which is mapped to two-dimensional or three-dimensional surface temperature textures and used as input features for the deep neural network model. ); The second part is the true value of the heat source distribution within the model, that is, the distribution corresponding to the spatial location. Value or internal temperature The voxel matrix serves as the label data for the deep neural network model. ).

[0045] To match the input-output format of the deep neural network model, the system performs voxel-based resampling of the unstructured temperature field data obtained from finite element mesh calculations. Using a trilinear interpolation algorithm, the continuous temperature field is mapped to a resolution of [resolution missing]. The simulation training dataset, constructed in a regular voxel mesh, contains no fewer than 100,000 pairs of surface temperature to internal heat source samples, and is divided into training, validation, and test sets in an 8:1:1 ratio to drive supervised learning training of the 3D U-Net model.

[0046] After constructing the simulation training dataset, the system performs supervised training of the deep neural network model. This process is configured to run on the graphics processing unit (GPU) of the data processing workstation, adjusting the weight parameters in the 3DU-Net network through an iterative optimization algorithm. and bias parameters This is to minimize the difference between the predicted results and the true labels.

[0047] This embodiment uses mean squared error (MSE) as the primary loss function. The MSE loss function is used to quantify the predicted internal temperature field output by the network. The actual internal temperature field obtained from simulation calculations The voxel-level numerical deviation between them. For a given... The formula for calculating the loss function for batch data of individual voxel samples is: ; In the formula, This indicates the total number of valid voxels in the current training batch; The model represents the first The predicted temperature value output by the individual units; Indicates the first The actual temperature label value corresponding to the individual element.

[0048] To address the class imbalance caused by the relatively small proportion of high-temperature anomaly regions within the total volume of the waste rock pile, this embodiment introduces a weighting term into the loss function. High-temperature voxels with temperatures exceeding a preset threshold (e.g., 300 degrees Celsius) are assigned a higher weighting coefficient. (For example =5), while assigning weight coefficients to low-temperature background voxels. (For example =1). The modified weighted loss function forces the network to pay more attention to the accurate reconstruction of high-temperature danger zones.

[0049] The optimizer is configured as an adaptive moment estimation optimizer (Adam). The Adam optimizer combines the advantages of the momentum method and the RMSProp algorithm, dynamically adjusting the learning rate of each parameter based on the first and second moment estimates of the gradient. The initial learning rate is set to 1 × 10⁻⁶. -4 Momentum decay parameter Set to 0.9, Set to 0.999.

[0050] The training strategy employs mini-batch gradient descent. The batch size is set to 8 or 16 to balance memory usage and gradient estimation stability. After each training epoch, the system evaluates the model's generalization performance using validation set data. If the loss value on the validation set does not decrease within five consecutive epochs, the system triggers a learning rate decay mechanism, reducing the current learning rate to 0.1 times its original value.

[0051] To prevent overfitting, a dropout regularization strategy is introduced after the fully connected or convolutional layers, with a dropout probability set to 0.5. Furthermore, the system performs data augmentation during training, including random rotation (90 degrees, 180 degrees, 270 degrees), random flipping, and adding Gaussian noise to the input voxel grid. This expands the diversity of the training data and improves the model's robustness to real-world environmental noise.

[0052] The training process terminates when the loss value on the validation set converges below a preset threshold or reaches the maximum number of training epochs (e.g., 200 epochs). The system saves all weight and bias parameters in the network at this point and generates the final deep neural network model file, which is then deployed to the intelligent inversion module for subsequent practical inference.

[0053] Reference Figure 2 and attached Figure 3As shown in the figure, step S1 is the air-ground collaborative data acquisition stage, which is executed autonomously by the UAV flight platform under the command of the ground control terminal.

[0054] Before conducting flight operations, the ground control terminal first imports the digital elevation model (DEM) or rough 3D terrain data of the target survey area. Based on this terrain data, the ground control terminal generates a terrain-following flight path. Unlike traditional fixed-altitude flight paths, the elevation coordinates of waypoints in a terrain-following flight path dynamically adjust according to the undulations of the target waste rock hill surface. Specifically, the system calculates the horizontal coordinates of each point along the flight path. Corresponding terrain elevation And set a constant relative flight altitude. (For example, 100 meters), then the absolute flight altitude of that waypoint The calculation is as follows: ; This variable-altitude flight strategy ensures that the imaging equipment on the UAV platform maintains a consistent object distance from the surface of the coal gangue hill throughout the entire operation, thereby guaranteeing that all acquired images have a uniform ground resolution (GSD) and avoiding problems such as low resolution in the foot of the hill or insufficient overlap in the top of the hill due to slope changes.

[0055] The flight path is planned as a bow-shaped reciprocating scanning trajectory. To meet the high requirements of feature matching in subsequent 3D reconstruction algorithms, the system strictly limits the overlap parameters of image acquisition. The forward overlap is set to no less than 80%, which is the overlap ratio of the ground cover area between two adjacent images along the flight direction; the lateral overlap is set to no less than 70%, which is the overlap ratio of the ground cover area between two adjacent parallel flight paths.

[0056] The data acquisition window was strictly limited to nighttime (2 hours after sunset to before sunrise) or early morning (1 hour before sunrise). The physical rationale for choosing this time period was to eliminate direct reflection interference from shortwave solar radiation and the non-uniform temperature rise on the surface caused by sunlight. During this period, the temperature distribution on the surface of the coal gangue hill is mainly determined by internal heat conduction, rather than external solar radiation heating, thereby improving the signal-to-noise ratio of thermal infrared images and enabling surface thermal anomalies to accurately reflect the internal ignition source status.

[0057] While the drone flight platform performs its data collection mission, ground-based environmental monitoring equipment records atmospheric parameters at the work site in real time. These parameters include ambient air temperature. relative humidity and atmospheric pressure These data are simultaneously timestamped for subsequent calculations to correct the radiation temperature of the thermal infrared images, thus eliminating the effects of atmospheric absorption and scattering attenuation on infrared radiation transmission. Finally, the UAV flight platform stores the acquired visible light image sequences and thermal infrared image sequences with high-precision POS (Position and Orientation System) information on the onboard solid-state drive, completing the data acquisition step.

[0058] Step S2 is the true 3D thermal texture reconstruction stage, which runs automatically in the 3D modeling module of the data processing workstation. Its core purpose is to construct a digital coal waste pile model that includes both accurate geometric structure and real temperature information.

[0059] First, the system performs high-precision geometric reconstruction using visible light imagery. The 3D modeling module employs the Structure from Motion (SfM) algorithm to process the visible light image sequence. The system extracts SIFT (Scale-Invariant Feature Transform) feature points from each visible light image and uses descriptor matching to find corresponding feature points between adjacent images. Based on these corresponding points and the POS information in the image metadata, the system uses bundle adjustment to calculate the precise exterior orientation elements (position and pose) of each image, as well as the 3D sparse point cloud coordinates of feature points in the scene.

[0060] Subsequently, based on the sparse point cloud and camera parameters, the system employs a Multi-View Stereo (MVS) algorithm for dense matching, generating high-density 3D point cloud data. Using Poisson reconstruction or Deloni triangulation algorithms, the dense point cloud is constructed into a continuous triangular irregular network (TIN) mesh model, i.e., the 3D geometric model of the target waste rock pile. This model accurately reproduces the macroscopic morphology, slope variations, and local fracture characteristics of the waste rock pile.

[0061] While generating the geometric model, the system performs physical model-based radiometric correction on the original thermal infrared image. The ambient temperature recorded in step S1 is used as a reference. and relative humidity By combining Planck's blackbody radiation law and atmospheric transport models (such as the simplified MODTRAN model), the radiance recorded by the thermal infrared sensor is corrected. The correction formula is: ; In the formula, This represents the true radiance of a surface target after atmospheric correction. This represents the total radiance received by the airborne thermal infrared sensor; Indicates the atmospheric transmittance in the infrared band at the operating altitude, calculated using humidity and distance; Indicates atmospheric path radiance; Indicates the brightness of reflected radiation from the surrounding environment. This represents the emissivity of the coal gangue surface, set as a constant between 0.93 and 0.96 based on the geological material properties. The system will calculate the... The inverse transformation is used to obtain the true surface dynamic temperature, generating a corrected temperature image.

[0062] Finally, thermal infrared texture mapping is performed. Since the visible light camera and the thermal infrared camera have a fixed relative position (lever arm vector) in their physical installation, and their field of view (FOV) differs, the system first spatially registers the thermal infrared image with the visible light image. Using the constructed 3D geometric model as a reference plane, the system calculates the projected coordinates of each triangular facet on the 3D model surface in the thermal infrared image using back projection technology. The system extracts the temperature values ​​at the projected coordinates and assigns them as the texture attributes of the triangular facet. For areas covered by multiple thermal infrared images, the system uses a weighted average method to fuse the temperature values, with the weights depending on the shooting angle (higher weight for more vertical images) and distance. After panoramic mapping processing, a three-dimensional thermal texture model is finally generated.

[0063] Step S3 is the intelligent inversion and diagnosis stage, which is automatically executed by the intelligent inversion module deployed on the data processing workstation. The core task of this step is to transform the 3D thermal texture model containing only surface information into a voxel distribution of high-temperature regions that includes internal state information.

[0064] First, the system preprocesses the 3D thermal texture model generated in step S2 to remove environmental background noise and locate potential anomaly areas. The intelligent inversion module reads the temperature values ​​of all nodes on the surface of the 3D thermal texture model and calculates the average background temperature of the surface using a statistical analysis algorithm. During the calculation process, the system uses the histogram statistical method to remove the 5% of data with the highest temperature (considered as potential heat sources) and the 5% of data with the lowest temperature (considered as measurement noise or water obstruction), and calculates the arithmetic mean of the temperature values ​​of the remaining data to obtain a stable background average temperature.

[0065] Subsequently, the system sets a temperature anomaly threshold. According to the judgment logic set in this embodiment, this threshold has a specific proportional relationship with the average background temperature, calculated using the following formula: ; In the formula, This is the temperature anomaly threshold. The background average temperature, The anomaly coefficient is set to 0.46 in this embodiment, which means that the temperature anomaly threshold is set to a value that is 46% higher than the average background temperature.

[0066] The system iterates through every node on the surface of the 3D thermal texture model, identifying nodes with temperature values ​​greater than a certain threshold. The continuous regions are marked as regions of interest. The system extracts only the topological structure and temperature texture data within these regions of interest and their surrounding preset range (e.g., extending outward by 5 meters), clips and reassembles them into standardized voxel input blocks, which serve as input data for the deep neural network model. This filtering step reduces the computational load of invalid background data and improves the targeting of subsequent inversion.

[0067] Next, the system invokes a pre-trained deep neural network model to perform inference operations on the input region of interest data. During inference, the encoder path of the deep neural network model uses 3D convolutional kernels to extract features from the input data. At this point, the convolutional kernels respond to the spatial distribution characteristics of the surface temperature field, specifically identifying the following geometric features: Continuity feature: Identify the spatial connectivity of high-temperature pixels and distinguish between sporadic noise and large areas of thermal anomalies; Discreteness characteristics: Calculate the spatial distribution density of high-temperature regions; Edge gradient characteristics: Calculate the rate of temperature decrease (temperature gradient) at the edge of a high-temperature region. ).

[0068] Deep neural network models execute pattern recognition logic based on physical laws through their deep network weights: if the input surface temperature field features show a clustered distribution (i.e., low dispersion and high continuity), and the edge temperature gradient is higher than a preset gradient threshold (reflecting the diffusion characteristics of heat conduction), the model determines that there is a concentrated heat source in the deep part corresponding to that location; conversely, if it shows high dispersion or extremely low edge gradient (gradual change), it is determined to be a shallow hot spot caused by solar radiation residue or surface biomass combustion.

[0069] After feature compression of the encoder path and nonlinear mapping of the decoder path, the deep neural network model finally outputs a three-dimensional tensor corresponding to the spatial dimensions of the input ROI. Each element in this tensor represents the spatial coordinates of the corresponding location inside the target waste rock pile. The system calculates the predicted temperature value at the location. It then stitches and merges the output results of all ROI regions according to their positions in the original coordinate system, filling uncalculated areas with ambient ground temperature, thereby reconstructing a complete voxel distribution matrix of the high-temperature region inside the target waste rock pile. This matrix digitally represents the three-dimensional location, depth, and intensity information of the heat source within the waste rock pile.

[0070] Step S4 is the 3D visualization and hierarchical early warning stage. This stage aims to transform the abstract digital field data output in step S3 into intuitive engineering early warning information. The data processing workstation reads the voxel distribution matrix of the high-temperature region inside the target waste rock pile, output by the deep neural network model. Each voxel in this matrix contains spatial location coordinates. and predicted internal temperature values The system first parses the matrix data, extracts all valid voxels with non-zero temperature values, and classifies the risk level based on the physicochemical characteristics of spontaneous combustion of coal gangue.

[0071] Risk level classification follows preset temperature threshold standards. The system sets two critical temperature thresholds: critical oxidation temperature. (e.g., 80 degrees Celsius) and auto-ignition temperature (For example, 260 degrees Celsius). Based on these two thresholds, the system divides the internal area into three specific risk levels: When voxel temperature When the area is identified as a low-risk zone, it indicates that the area is in a slow oxidation stage and has not yet formed thermal accumulation. when When the area is identified as a medium-risk area, it indicates that the region is in a phase of accelerated oxidation and heating, requiring close monitoring. when When a high-risk area is identified, it indicates that a violent oxidation or combustion reaction has occurred within the area, and there is an open flame or a high-temperature smoldering fire.

[0072] See attached document Figure 4 After completing the risk classification, the system generates a 3D visual early warning map. Using ray casting algorithms in volume rendering technology, the system renders the 3D voxel matrix into a visualized 3D image. The system establishes color transfer functions and opacity transfer functions to map different risk levels to specific visual attributes.

[0073] For low-risk areas, the system renders them green or sets the opacity to 0 (completely transparent) to reduce visual occlusion; For medium-risk areas, the system renders them as yellow and sets a low opacity (e.g., 0.3) to make them appear as semi-transparent clouds. For high-risk areas, the system renders them as highly saturated red with a high opacity (e.g., 0.8 to 1.0) to make them stand out clearly in three-dimensional space. Through this differentiated rendering process, the three-dimensional visualization early warning map can intuitively display the shape and distribution of internal heat sources through surface texture.

[0074] For areas identified as high-risk, the system further performs a connected component analysis algorithm. The system identifies spatially adjacent (sharing vertices, edges, or faces) red high-risk voxels as independent combustion core clusters. For each identified combustion core cluster, the system uses a moment estimation algorithm to calculate the three-dimensional coordinates of its geometric center (Centroid). .

[0075] The system calculates the vertical distance from the geometric center to the ground surface directly above it, i.e., the center depth. The system automatically generates leader line labels in the 3D visualization early warning map, indicating the center depth. The core temperature peak is marked above the corresponding high-risk area, providing precise grouting depth or drilling depth parameters for subsequent fire prevention and extinguishing projects. The final generated warning map supports rotation, scaling, and cross-sectional operations at any angle, allowing operators to view the temperature field profile at any depth inside the waste rock pile.

Claims

1. A method for three-dimensional inversion of internal fire source of coal gangue dump based on unmanned aerial vehicle remote sensing, characterized in that, The method comprises the following steps: Step S1: controlling a UAV to carry a visible light imaging device and a thermal infrared imaging device, and collecting data of a target gangue hill along a preset flight route to obtain visible light images and thermal infrared images; Step S2: constructing a three-dimensional geometric model of the target gangue hill based on the visible light images, and mapping the thermal infrared images to the surface of the three-dimensional geometric model through feature matching to generate a three-dimensional thermal texture model; Step S3: inputting the three-dimensional thermal texture model into a pre-trained deep neural network model, extracting temperature field spatial distribution features of the surface of the three-dimensional thermal texture model by the deep neural network model, and outputting voxel distribution of a high-temperature region inside the target gangue hill through nonlinear mapping; Step S4: generating a three-dimensional visual warning map according to the voxel distribution of the high-temperature region, and marking a self-ignition danger zone.

2. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, In the step S1, the preset flight route is a ground-following flight route; the UAV carries the visible light imaging device and the thermal infrared imaging device, The data collection along the preset flight route specifically comprises: controlling the UAV to keep a constant relative height with the surface of the target gangue hill and fly along an arch-shaped reciprocating path; and in the flying process, the heading overlap and the lateral overlap between adjacent visible light images and adjacent thermal infrared images are both greater than 70%.

3. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, In the step S1, the execution time period of the data collection is set as night or early morning, and the environmental temperature and the environmental humidity are monitored in real time during the collection process; In the step S2, before the three-dimensional thermal texture model is generated, the temperature data of the thermal infrared images are corrected by an atmospheric transmission model based on the environmental temperature and the environmental humidity.

4. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, In the step S3, the deep neural network model adopts a 3DU-Net network architecture; The 3DU-Net network architecture comprises an encoder path and a decoder path; the encoder path is used for down-sampling and feature extraction of surface temperature features of the three-dimensional thermal texture model; the decoder path is used for up-sampling of the extracted features, and finally reconstructs a three-dimensional matrix of the temperature field inside the target gangue hill by fusing features of different scales through a skip connection.

5. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 4, characterized in that, The training process of the deep neural network model comprises: constructing a virtual gangue hill three-dimensional model, setting internal heat source conditions of multiple different depths and shapes based on a heat conduction equation, and calculating and generating corresponding virtual surface temperature fields by using a multi-physical field numerical simulation software; constructing a training data set, taking the virtual surface temperature fields as input data, and taking the real distribution of the corresponding internal heat source conditions as label data; supervising and training the 3DU-Net network architecture by using the training data set until the loss function converges, and obtaining the deep neural network model.

6. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, The step S3 further comprises a preprocessing sub-step: calculating the background average temperature of the surface of the three-dimensional thermal texture model; setting a temperature anomaly threshold, which is set as a value higher than 46% of the background average temperature; based on the temperature anomaly threshold, extracting a surface temperature anomaly region from the three-dimensional thermal texture model as a region of interest; Only the data of the region of interest is input into the deep neural network model for internal inversion.

7. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, In the step S3, the extracting the spatial distribution characteristics of the temperature field of the surface of the three-dimensional thermal texture model specifically includes: identifying geometric morphological characteristics of the high-temperature region, the geometric morphological characteristics including continuity, discreteness, and edge gradient of the high-temperature region; if the geometric morphological characteristics present a cluster distribution and the edge gradient is higher than a preset gradient threshold, the deep neural network model determines that a concentrated heat source exists in the corresponding position in the deep part. 8.The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, In the step S4, the marking the self-ignition risk area includes: obtaining temperature values and depth values in the voxel distribution of the high-temperature region; dividing the risk levels according to the temperature values into high risk, medium risk, and low risk; in the three-dimensional visualization early warning map, different colors are used to render regions of different risk levels, and the center depth coordinates corresponding to the high-risk region are calibrated. 9.The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, The method further includes a dynamic correction step: after grouting treatment or drilling exploration is performed on the self-ignition risk area, actual internal temperature data is obtained; the actual internal temperature data is compared with the voxel distribution of the high-temperature region output by the deep neural network model, and an error value is calculated; the actual internal temperature data and the corresponding three-dimensional thermal texture model data are added to the training database, and the deep neural network model is incrementally trained.

10. The coal gangue dump internal fire source three-dimensional inversion method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, In the step S3, a multi-time dynamic verification step is further included: repeated flight operations and intelligent inversion are performed on the target gangue hill at different time points to obtain multi-period voxel distributions of the high-temperature region; by comparing the multi-period voxel distributions of the high-temperature region, regions with overlapping spatial coordinate positions and stable temperature values are identified, and it is confirmed that the regions are real risk areas with stable heat sources, and transient pseudo-anomaly areas affected by the environment are excluded.

Images