Infrared spectrum inversion method for gas cloud contour of coal mine gas

Abstract

The invention discloses a coal mine gas cloud contour infrared spectrum inversion method. The method comprises the following steps: 1, simulating an underground gas cloud formation environment; 2, acquiring infrared image data containing gas cloud by using an infrared spectrum detection device; step 3, carrying out de-noising processing on the infrared image data; 4, performing temperature drift correction and radiation calibration on the denoised infrared image, and converting a DN value into radiation brightness; 5, based on the multi-frame de-noised image, performing target detection and contour extraction on the gas cloud by using a Methane-YOLO model, performing multi-view geometric matching of feature points, performing adjustment optimization of three-dimensional coordinates in combination with a bundle method, generating a point cloud model, and reconstructing a curved surface; 6, based on the three-layer radiation transmission model and the Lambert-Beer law, inverting the gas concentration; and step 7, based on the CEEMDAN-VMD secondary decomposition and CNN-LSTM hybrid network, carrying out joint training on the space contour and time sequence data, and carrying out three-dimensional space concentration distribution prediction.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety technology, and in particular to a method for inverting the outline of coal mine gas clouds using infrared spectroscopy. Background Technology

[0002] Coal will remain my country's primary energy source for a considerable period. Coal and gas outbursts are among the most serious hazards threatening safe coal mine production. Numerous coal and gas outburst accidents demonstrate that most pre-outburst gas emission exhibits significant anomalies. Current gas monitoring systems can only reflect the gas concentration at a single point in time, failing to provide a true response to the gas distribution across the entire space and timeframe. Infrared imaging technology, as a typical non-contact detection method, offers advantages such as wide monitoring range and intuitive visualization. However, due to the unique characteristics of mines, including high dust content, low contrast, and complex and variable temperature, humidity, and operating conditions, infrared spectral noise reduction, gas profile extraction, and gas concentration prediction are crucial scientific problems that urgently need to be solved. Coal and gas outbursts are characterized by their suddenness and extreme destructiveness, and with the high-yield and efficient development of mines and their continuous expansion and deepening, this hazard is becoming increasingly severe. Comprehensive and effective identification and accurate early warning of outburst risks are effective means to curb outburst accidents, with abnormal gas emission being one of the main warning signs. Therefore, it is essential to leverage the non-contact detection advantages of infrared spectral imaging technology to achieve wide-range and high-dimensional acquisition of mine gas data.

[0003] Existing equipment for gas detection only measures gas at specific points and cannot detect gas concentration in the three-dimensional spatial domain, thus failing to achieve a breakthrough from point-to-surface measurement. Therefore, an infrared spectral inversion method for coal mine gas cloud profiles is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for inverting the outline of coal mine gas clouds using infrared spectroscopy.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The method for inverting the outline of coal mine gas clouds using infrared spectroscopy includes the following steps: Step 1: Construct a high-dust environment simulation experimental platform in the mine to simulate the formation environment of underground gas clouds; Step 2: Use an infrared spectroscopy detection device to acquire infrared image data containing gas clouds; Step 3: Denoise the infrared image data. The denoising process includes using wavelet transform to remove water vapor noise and using median filtering to remove dust noise. Step 4: Perform temperature drift correction and radiometric calibration on the denoised infrared image, and convert the DN value into radiance; Step 5: Based on the denoised images of multiple frames, the Methane-YOLO model is used to perform target detection and contour extraction on the gas cloud. By matching feature points from multiple perspectives and combining bundle adjustment to optimize the three-dimensional coordinates, a point cloud model is generated and the surface is reconstructed. Step 6: Based on the three-layer radiative transfer model and Lambert-Beer's law, invert the methane concentration; Step 7: Based on the CEEMD-VMD quadratic decomposition and CNN-LSTM hybrid network, jointly train the spatial contour and time series data to construct a gas concentration time series prediction model and predict the three-dimensional spatial concentration distribution.

[0006] The above further includes: Furthermore, the high-dust environment simulation experimental platform for mines includes a gas concentration control system, a dust concentration control system, a temperature control system, a humidity control system, an air pressure control system, and a data acquisition system; The gas concentration control system prepares methane standard gas of different concentrations to establish a quantitative relationship model between gas concentration and infrared spectral response, and to test the inversion accuracy of the algorithm at different concentrations. It consists of a CH4 gas source, an N2 gas source, a mass flow controller and a flow display instrument. By adjusting the two mass flow controllers, the flow ratio of CH4 and N2 is controlled to ensure thorough mixing in the mixing chamber, thereby preparing CH4 gas with a preset volume fraction. The dust concentration control system simulates a high-dust environment underground and studies the effects of dust on infrared light shielding, refraction, and scattering. It includes a dust emitter, coal dust raw material, turbine fan, and dust concentration measuring instrument. The temperature control system is used to control the background temperature and gas temperature of the experimental environment. A heating tape is evenly wrapped around the background plate, and the power of the heating tape is precisely controlled by the temperature controller to stabilize the temperature of the background plate at a preset value. The humidity control system simulates the high humidity environment caused by dust suppression spraying in the well, studies the interference of the absorption band of water vapor at 3.2μm on the characteristic absorption peak of methane, generates water mist by an ultrasonic humidifier, monitors the ambient humidity by a humidity recorder, and forms a closed-loop control to stabilize the humidity at a preset value. The pressure control system controls the experimental pressure to standard atmospheric pressure, eliminating the influence of pressure changes on the gas absorption coefficient and inversion results, and monitors them in real time through a barometer. The data acquisition system collects raw infrared spectral image data under all control variables.

[0007] Furthermore, the specific steps for denoising the infrared image data are as follows: Wavelet transform for noise reduction: Two-dimensional wavelet basis construction: The one-dimensional scaling function u(x) and wavelet function v(x) are extended to a two-dimensional form, as follows: ; ; ; ; in, , , These represent wavelet bases in the horizontal, vertical, and diagonal directions, respectively. Image decomposition: Perform discrete wavelet transform on an infrared image f(x,y) of size M×N to obtain approximation coefficients. and detail coefficient , represented as ; Thresholding: A soft thresholding function is used to denoise the detail coefficients, expressed as follows: Where T is the threshold; Image reconstruction: Reconstructing the denoised image using inverse discrete wavelet transform: ; Median filtering: Sliding window selection: Select a 3×3 or 5×5 window based on the noise type; Pixel replacement: Sort the pixel values ​​within the window and replace the center pixel with the median value; Traverse the image: Slide the window in row-major order to cover the entire image.

[0008] Furthermore, the step of performing temperature drift correction and radiometric calibration on the denoised infrared image to convert the DN value into radiance includes the following steps: Non-imaging area monitoring: Continuously acquire pixel output values ​​of non-imaging areas in infrared images and record pixel jump values ​​of non-imaging areas at different temperatures; Correction factor calculation: Determine the linear correction factor k based on the relationship between the jump value and temperature change; Real-time correction: Applying the formula to each pixel in the imaging area ,in, The corrected pixel value. For the measured pixel values, Here, T represents the temperature coefficient, and T represents the current ambient temperature. For reference temperature; Blackbody calibration experiment: Use a blackbody radiation source to set different temperatures and record the corresponding DN values; Least squares fitting: Substitute the values ​​from the blackbody calibration experiment into the formula. ,in, Let i be the entrance pupil radiance. Here is the corrected DN value for band i. Using the scaling coefficients, solve the system of equations to obtain a and b; Radiance calculation: Radiance is calculated based on the corrected DN value.

[0009] Furthermore, the step of performing target detection and contour extraction on the gas cloud based on multi-frame denoised images using the Methane-YOLO model includes the following steps: Input processing: The denoised infrared image is adjusted to 448×448 pixels and normalized to the [0,1] range; Feature extraction: Multi-scale features are extracted through the C2f-SPPF backbone network, and the feature representation is enhanced by combining the global attention mechanism C2PSA. Detection head prediction: Each grid cell predicts 3 bounding boxes and 6 class probabilities; Output dimension: Output = S × S × (B × 5 + C), where S is the number of grids, B is the number of bounding boxes per grid, and C is the number of target categories in the mine; Non-maximum suppression: Merge overlapping bounding boxes, retain high-confidence detection results, and represent the bounding box coordinate prediction as follows: Where W=448 and H=448 are the input image dimensions; Contour extraction: Based on the detected bounding boxes, extract the contour of the gas cloud and generate a binary mask.

[0010] Furthermore, the process of generating a point cloud model and reconstructing the surface by matching feature points through multi-view geometry and combining bundle adjustment to optimize the three-dimensional coordinates includes the following steps: Feature extraction: Extract SIFT feature points from multi-view images and generate descriptors; Coarse matching: Feature points are matched based on descriptor similarity, and pairs whose nearest neighbor distance is less than 0.8 compared to the second nearest neighbor distance are retained; Geometric verification: Estimate the fundamental matrix F using the RANSAC algorithm and remove outliers; Exact matching: Matching pairs that satisfy the fundamental matrix constraints are retained as input for subsequent 3D reconstruction. These fundamental matrix constraints are expressed as follows: Where F is the basic matrix, and For matching point pairs; Initialization: Initial 3D point coordinates and camera pose are calculated using triangulation methods; Constructing the error term: For each pair of matched points, calculate the reprojection error between the observed point and the predicted point, whereby the reprojection error is expressed as... ,in, For the observation point, For prediction points; Optimization solution: The Levenberg-Marquardt algorithm is used to minimize the total error, and the camera pose and 3D point coordinates are adjusted iteratively. Convergence criterion: Stop when the error change is less than the threshold or when the maximum number of iterations is reached; Point cloud generation: Save the optimized 3D point coordinates as a point cloud file; Normal estimation: The normal direction of each point is calculated using the K-nearest neighbor algorithm; Triangulation: A greedy projection triangulation algorithm is used to connect adjacent points and generate a triangular mesh; Surface Reconstruction: A smooth surface model is generated using the Poisson reconstruction algorithm; Post-processing: Remove isolated points, repair holes, and optimize the model topology.

[0011] Furthermore, the specific steps for inverting the gas concentration based on the three-layer radiative transfer model and the Lambert-Beer law are as follows: Step 1: Data Preparation and Preprocessing: Acquire the data after denoising in Step 3 and temperature drift correction and radiometric calibration in Step 4. At this point, the original infrared image DN values ​​have been converted into entrance pupil radiance values ​​for each pixel in a specific band. ; Step 2: Construct a three-layer radiative transfer physical model: A three-layer model—foreground layer, target layer, and background layer—is established to describe the infrared radiation transfer process. Foreground layer: refers to the air layer between the infrared spectral imager and the gas cloud; Target layer: the gas cloud layer to be measured, the region containing methane gas; Background layer: the coal wall or tunnel background behind the gas cloud; Total Radiance Received It consists of two parts: the attenuation portion of the background radiation after passing through the target gas layer and the thermal radiation portion of the target gas layer itself, denoted as... ,in, This refers to the actual radiance received by the infrared system. It is the transmittance of the target gas layer at wavelength λ. It is the radiance of the background layer. It is the thermal radiation brightness of the target gas layer itself; Step 3: Introduce the blackbody radiation law and gas emissivity model: Solve for the unknowns in the three-layer radiative transfer physical model by introducing Planck's blackbody radiation law and the gas emissivity model: The spectral radiance B(λ,T) of a blackbody is calculated using Planck's formula: ,,in, , These are the first and second radiation constants; The gas layer is equivalent to an emissivity According to Kirchhoff's laws, the emissivity of a blackbody is equal to its absorptivity. The thermal radiation of the gas itself is expressed as: ,in, It is the gas at its own temperature Blackbody radiation brightness; Step 4: Establish the relationship between transmittance and concentration using Lambert-Beer's Law: Lambert-Beer's Law establishes the relationship between gas transmittance and its concentration and absorption characteristics: ,in, λ is the absorption coefficient of methane gas at wavelength λ, n is the column concentration of the gas to be inverted, and L is the effective optical path length of infrared light through the gas. Step 5: Derive and solve the concentration inversion algorithm: Substitute the formulas from Steps 3 and 4 into the physical model from Step 2, perform mathematical derivation and simplification, and obtain the inversion formula for column concentration n: ; Step 6: Pixel-by-pixel calculation and concentration image generation: Traverse each pixel in the infrared image and repeat the calculation process from Step 1 to Step 5 above. Recombine the column concentration values ​​n calculated for each pixel to generate a gas concentration distribution image.

[0012] Furthermore, in step six, the input feature set of the CNN-LSTM hybrid network is constructed based on the CEEMD-VMD quadratic decomposition, including the following steps: Constructing a four-dimensional spatiotemporal dataset: From the preprocessed and contour-extracted infrared image sequence, a spatiotemporal dataset is obtained. Assuming that each time point t yields a gas concentration distribution image of size [H, W], then after T time points, a four-dimensional dataset is constructed. This represents a dataset containing T time steps, each time step being a single-channel image of size H×W; CEEMDAN decomposition: For the concentration time series X(i,j) at position (i,j), the CEEMDAN algorithm is used to decompose it into a series of intrinsic mode functions (C-IMFs) and a residual; Calculate sample entropy: Calculate the sample entropy of each C-IMF component. The lower the sample entropy value, the more regular the sequence. Sequence Reconstruction: Based on the sample entropy values, all C-IMF components and residuals are reconstructed into three new sequences: High-Frequency Sequence: composed of the C-IMF components with the largest sample entropy values, representing the random noise and drastic fluctuations in the sequence; Low-Frequency Sequence: composed of the C-IMF components with intermediate sample entropy values, representing the short-term regular changes in the sequence; Trend Sequence: composed of the C-IMF component with the smallest sample entropy value and residuals, representing the long-term trend of the sequence. VMD secondary decomposition: The reconstructed high-frequency and low-frequency sequences are decomposed in a secondary manner. For the high-frequency and low-frequency sequences at each spatial location, the variational mode decomposition algorithm is used to further decompose them to obtain q intrinsic mode sequences, denoted as V-IMF. Feature concatenation: For each time point t and each spatial location (i, j), its original concentration value is concatenated with the values ​​of all q V-IMFs at that time to form a multi-feature vector; Constructing spatiotemporal samples: Training samples are constructed using the sliding time window method. The time window length Δt and the sliding step size S are set. For a sample, the input is the set of feature vectors of all locations in the entire space within the past Δt time steps, and the output is the image of the entire gas concentration distribution in the next time step.

[0013] Furthermore, the specific steps for constructing the time-series prediction model for gas concentration are as follows: CNN Module: Input: The feature map at each time step t is a 3D tensor [H, W, (1+q)], where each channel represents a feature; Operation: Slide the convolutional layers in the spatial dimension using 3D or 2D convolutions. The convolutional layers automatically extract local spatial features, and the pooling layers perform feature dimensionality reduction; Output: For each time step, the CNN outputs a spatial feature vector or feature map. LSTM Module: Input: Spatial feature vectors extracted by CNN at Δt time steps are input into the LSTM layer in chronological order; Operation: The gating mechanism of the LSTM unit learns long-term dependencies in the time series, memorizes and forgets relevant information, and captures the dynamic changes in gas outbursts; Output: The output state of the last time step of the LSTM encodes the spatiotemporal evolution information of the entire spatial concentration field over the past Δt time period; Fully connected layers and prediction: The output of the LSTM module is mapped through one or more fully connected layers. Deconvolution or upsampling layers are used to reconstruct the feature vector into a predicted image consistent with the original spatial dimensions [H, W], i.e., a three-dimensional spatial concentration distribution prediction map for future time steps. Model training and validation: Mean squared error or mean absolute error is used as the loss function to minimize the difference between the predicted concentration field and the true concentration field. An automated hyperparameter search method is used for optimization. Historical infrared spectral monitoring data is used to divide the training set, validation set and test set in chronological order to train and validate the model.

[0014] Furthermore, it also includes a visualization process for the inversion results, where the three-dimensional gas cloud outline is overlaid as a pseudo-color layer on the underground digital twin model of the coal mine.

[0015] The present invention has the following beneficial effects: 10. In this invention, passive infrared spectral imaging technology is used as the core detection method. This method can obtain gas information in the entire field of view non-contactly by receiving infrared radiation in a specific band and generate an infrared image containing spatial distribution data. This enables real-time and visualized monitoring of gas outbursts in coal mine working faces in two dimensions and even in three dimensions through inversion. It completes the fundamental transformation from "seeing only a part of the whole" to "global control" and provides an unprecedented data dimension and information foundation for accurate early warning.

[0016] 11. In this invention, wavelet transform and median filtering are combined to selectively remove water vapor and dust. For contour extraction, an improved Methane-YOLO deep learning model is further proposed. By introducing a global attention mechanism and a small target detection layer, the accuracy of identifying and segmenting low-contrast, blurred-edge gas clouds is greatly enhanced.

[0017] 12. In this invention, advanced secondary mode decomposition (CEEMDAN-VMD) is combined with CNN-LSTM hybrid neural network to make full use of the high-dimensional characteristics of infrared spectral data. It not only predicts the concentration time series, but also potentially contains the evolution law of spatial distribution characteristics, realizing the upgrade from static surface monitoring to dynamic spatiotemporal evolution prediction. Attached Figure Description

[0018] Figure 1 This is a step diagram of the infrared spectral inversion method for the contour of coal mine gas clouds proposed in this invention; Figure 2 This is a framework diagram of the infrared spectral methane concentration time-series prediction model in this invention; Figure 3 This is a diagram showing the composition of the high-dust environment simulation experimental platform in the mine according to the present invention; Figure 4 The infrared image used in this invention; Figure 5 This is a diagram showing the effect of background frame difference extraction and display of infrared images in this invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1-5 As shown, this invention provides a method for inverting the infrared spectroscopic profile of coal mine gas clouds, comprising the following steps: Step 1: Construct a high-dust environment simulation experimental platform in the mine to simulate the formation environment of underground gas clouds; Step 2: Use an infrared spectroscopy detection device (passive infrared spectroscopy imager) to collect infrared image data containing gas clouds (coal mine roadways or working faces). Step 3: Denoise the infrared image data. The denoising process includes using wavelet transform to remove water vapor noise and using median filtering to remove dust noise. Step 4: Perform temperature drift correction and radiometric calibration on the denoised infrared image, and convert the DN value into radiance; Step 5: Based on the denoised images of multiple frames, the Methane-YOLO model is used to perform target detection and contour extraction on the gas cloud. By matching feature points from multiple perspectives and combining bundle adjustment to optimize the three-dimensional coordinates, a point cloud model is generated and the surface is reconstructed. Step 6: Based on the three-layer radiative transfer model and Lambert-Beer's law, invert the methane concentration; Step 7: Based on the CEEMD-VMD quadratic decomposition and CNN-LSTM hybrid network, jointly train the spatial contour and time series data to construct a gas concentration time series prediction model and predict the three-dimensional spatial concentration distribution.

[0021] In one embodiment, the high-dust environment simulation experimental platform in the mine includes a gas concentration control system, a dust concentration control system, a temperature control system, a humidity control system, an air pressure control system, and a data acquisition system. The gas concentration control system prepares standard methane (CH4) gas at different concentrations to establish a quantitative relationship model (calibration curve) between gas concentration and infrared spectral response, and to test the inversion accuracy of the algorithm at different concentrations. It consists of a high-purity CH4 gas source (99.999% purity), a high-purity N2 gas source (99.999% purity, used as dilution and background gas), a mass flow controller (MFC), and a flow display. By adjusting the two mass flow controllers, the flow ratio of CH4 and N2 is controlled, and they are thoroughly mixed in the mixing chamber to prepare CH4 gas with a preset volume fraction. It can accurately generate CH4 gas with six gradient concentrations of 0.5%, 1.0%, 1.5%, 2.0%, 2.5%, and 3.0%, covering the common gas concentration range in underground coal mines. The dust concentration control system simulates a high-dust environment underground to study the effects of dust on infrared light shielding, refraction, and scattering. It includes a dust emitter (powered by an air compressor), coal dust raw materials (different particle sizes), turbine fans (≥2, to ensure uniform dust distribution within the experimental space), and a dust concentration measuring instrument (placed in the center of the experimental space for real-time monitoring and feedback). The dust emission rate is controlled by adjusting the conveyor belt speed knob of the dust emitter, with six concentration gradients set: 100, 200, 300, 400, 500, and 600 mg / m³. Coal powder of different mesh sizes is used to simulate dust interference of different particle sizes, including four levels: 1000 mesh (1.3 μm), 2000 mesh (2.6 μm), 5000 mesh (6.5 μm), and 10000 mesh (13 μm). The system studies particle size effects. To ensure stable experimental conditions, the dust emission concentration must be maintained for 15 minutes before data acquisition. The temperature control system is used to control the background temperature and gas temperature of the experimental environment. Infrared radiation brightness is directly related to temperature and is a key parameter for radiation calibration and concentration inversion. A heating tape is evenly wrapped around the background plate, and the power of the heating tape is precisely controlled by the temperature controller to stabilize the background plate temperature at a preset value. Under the action of the turbine fan, the temperature inside the entire experimental chamber is kept constant. The background plate temperature can be set to six gradients: 30℃, 35℃, 40℃, 45℃, 50℃, and 55℃, and the temperature fluctuation is stabilized within ±1℃. The humidity control system simulates the high humidity environment caused by dust suppression spraying underground. It studies the interference of the absorption band of water vapor (H2O) at 3.2μm on the characteristic absorption peak of methane (3.1-3.5μm). Water mist is generated by an ultrasonic humidifier, and the ambient humidity is monitored by a humidity recorder to form a closed-loop control to stabilize the humidity at a preset value. Based on the empirical value of mine air humidity, the relative humidity can be set to five gradients: 50%, 60%, 70%, 80%, and 90%. The pressure control system controls the experimental pressure to standard atmospheric pressure, eliminating the influence of pressure changes on the gas absorption coefficient and inversion results, ensuring the repeatability of the experiment and the comparability of the data. The pressure is monitored in real time by a barometer. The experiment is conducted in a closed space and the pressure is maintained at one standard atmosphere. The data acquisition system collects raw infrared spectral image data under all control variables. It is the data sensing terminal of the entire platform and provides raw materials for subsequent algorithm processing.

[0022] In one embodiment, the specific steps for denoising the infrared image data are as follows: Wavelet transform for noise reduction: Two-dimensional wavelet basis construction: The one-dimensional scaling function u(x) and wavelet function v(x) are extended to a two-dimensional form, as follows: ; ; ; ; in, , , These represent wavelet bases in the horizontal, vertical, and diagonal directions, respectively. Image decomposition: Perform Discrete Wavelet Transform (DWT) on an infrared image f(x,y) of size M×N to obtain approximation coefficients. and detail coefficient , represented as ; Thresholding: A soft thresholding function is used to denoise the detail coefficients, expressed as follows: , where T is the threshold (adaptively determined by the Bayesian criterion); Image reconstruction: Reconstructing the denoised image using Inverse Discrete Wavelet Transform (IDWT): ; Median filtering: Sliding window selection: Select a 3×3 or 5×5 window based on the noise type (3×3 is recommended for salt and pepper noise); Pixel replacement: Sort the pixel values ​​within the window and replace the center pixel with the median value; Traverse the image: Slide the window in row-major order to cover the entire image.

[0023] In one embodiment, performing temperature drift correction and radiometric calibration on the denoised infrared image to convert the DN value into radiance includes the following steps: Non-imaging area monitoring: Continuously acquire pixel output values ​​of non-imaging areas (such as blackbody reference sources) in infrared images. The values ​​of this area are only affected by ambient temperature. Record the pixel jump values ​​of non-imaging areas at different temperatures (e.g., the jump value is 5DN when the temperature rises from 25℃ to 30℃). Correction factor calculation: Determine the linear correction factor k based on the relationship between the jump value and temperature change; For example: k = ΔDN / ΔT = 5 DN / 5℃ = 1 DN / ℃.

[0024] Real-time correction: Applying the formula to each pixel in the imaging area ,in, The corrected pixel value. For the measured pixel values, Here, T represents the temperature coefficient (calibrated experimentally), and T represents the current ambient temperature (°C). This is a reference temperature (usually 25°C). Blackbody calibration experiment: Use a high-precision blackbody radiation source (such as BXC-15 / BXL-500) to set different temperatures (such as 30℃, 40℃, 50℃), and record the corresponding DN values; Example data: Least squares fitting: Substitute the values ​​from the blackbody calibration experiment into the formula. ,in, Let be the entrance pupil radiance of band i (W / (m²·sr·μm)). Here is the corrected DN value for band i. Given the scaling coefficients (obtained through boldbody calibration), solve the system of equations to obtain a and b; Radiance calculation: Radiance is calculated based on the corrected DN value.

[0025] In one embodiment, the target detection and contour extraction of the gas cloud based on the multi-frame denoised images using the Methane-YOLO model includes the following steps: Input processing: The denoised infrared image is adjusted to 448×448 pixels and normalized to the [0,1] range; Feature extraction: Multi-scale features are extracted through the C2f-SPPF backbone network, and the feature representation is enhanced by combining the global attention mechanism C2PSA. Detection head prediction: Each grid cell predicts 3 bounding boxes (containing x, y, w, h, confidence) and 6 class probabilities; Output dimension: Output = S × S × (B × 5 + C), where S = 7 is the number of grids, B = 3 is the number of bounding boxes for each grid, and C = 6 is the number of target categories in the mine; Non-maximum suppression (NMS): Merges overlapping bounding boxes, retaining high-confidence detection results (e.g., confidence threshold of 0.5). The bounding box coordinate prediction is expressed as... Where W=448 and H=448 are the input image dimensions; Contour extraction: Based on the detected bounding boxes, extract the contour of the gas cloud and generate a binary mask.

[0026] In one embodiment, the step of generating a point cloud model and reconstructing a surface by matching feature points through multi-view geometry and combining bundle adjustment to optimize three-dimensional coordinates includes the following steps: Feature extraction: Extract SIFT feature points from multi-view images and generate descriptors; Coarse matching: Feature points are matched based on descriptor similarity, and pairs whose nearest neighbor distance is less than 0.8 compared to the second nearest neighbor distance are retained; Geometric verification: Estimate the fundamental matrix F using the RANSAC algorithm and remove outliers; Exact matching: Matching pairs that satisfy the fundamental matrix constraints are retained as input for subsequent 3D reconstruction. These fundamental matrix constraints are expressed as follows: Where F is the basic matrix, and For matching point pairs; Initialization: Initial 3D point coordinates and camera pose are calculated using triangulation methods; Constructing the error term: For each pair of matched points, calculate the reprojection error between the observed point and the predicted point, whereby the reprojection error is expressed as... ,in, For the observation point, For prediction points; Optimization solution: The Levenberg-Marquardt algorithm is used to minimize the total error, and the camera pose and 3D point coordinates are adjusted iteratively. Convergence criterion: Stop when the error change is less than the threshold (e.g., 1e-5) or when the maximum number of iterations (e.g., 100) is reached; Point cloud generation: Save the optimized 3D point coordinates as a point cloud file (e.g., .ply format); Normal estimation: The normal direction of each point is calculated using the K-nearest neighbor algorithm; Triangulation: A greedy projection triangulation algorithm is used to connect adjacent points and generate a triangular mesh; Surface Reconstruction: A smooth surface model is generated using the Poisson reconstruction algorithm; Post-processing: Remove isolated points, repair holes, and optimize the model topology.

[0027] In one embodiment, the specific steps for inverting the gas concentration based on the three-layer radiative transfer model and the Lambert-Beer law are as follows: Step 1: Data Preparation and Preprocessing: Acquire the data after denoising in Step 3 and temperature drift correction and radiometric calibration in Step 4. At this point, the original infrared image DN values ​​have been converted into entrance pupil radiance values ​​for each pixel in a specific wavelength band (3.1-3.5μm). (Unit: W / (m²·sr·μm)); Step 2: Construct a three-layer radiative transfer physical model: A three-layer model—foreground layer, target layer, and background layer—is established to describe the infrared radiation transfer process. Foreground layer: refers to the air layer between the infrared spectral imager and the gas cloud; Target layer: the gas cloud layer to be measured, the region containing methane gas; Background layer: the coal wall or tunnel background behind the gas cloud; Total Radiance Received It consists of two parts: the attenuation portion of the background radiation after passing through the target gas layer and the thermal radiation portion of the target gas layer itself, denoted as... ,in, It is the actual radiance received by the infrared system (known, obtained from calibration in step four). It is the transmittance (unknown) of the target gas layer (methane) at wavelength λ. It is the radiance of the background layer. It is the thermal radiation brightness of the target gas layer itself; Step 3: Introduce the blackbody radiation law and gas emissivity model: Solve for the unknowns in the three-layer radiative transfer physical model by introducing Planck's blackbody radiation law and the gas emissivity model: The spectral radiance B(λ,T) of a blackbody is calculated using Planck's formula: ,,in, , These are the first and second radiation constants; The gas layer is equivalent to an emissivity According to Kirchhoff's laws, the emissivity of a blackbody is equal to its absorptivity. The thermal radiation of the gas itself is expressed as: ,in, It is the gas at its own temperature Blackbody radiation brightness; Step 4: Establish the relationship between transmittance and concentration using Lambert-Beer's Law: Lambert-Beer's Law establishes the relationship between gas transmittance and its concentration and absorption characteristics: ,in, λ is the absorption coefficient of methane gas at wavelength λ (unit: cm⁻¹), which is a physical constant related to temperature and pressure. This invention obtains this coefficient by querying the international standard HITRAN database. n is the column concentration of the gas to be inverted (unit: ppm·m), which is the core parameter ultimately required by this invention. L is the effective optical path length of infrared light through the gas (unit: m), which can be obtained through measurement or estimation in experimental and application scenarios.

[0028] Step 5: Derive and solve the concentration inversion algorithm: Substitute the formulas from Steps 3 and 4 into the physical model from Step 2, perform mathematical derivation and simplification, and obtain the inversion formula for column concentration n: In practical applications, background radiance This can be obtained by calibrating the same background area under conditions of no gas leakage, or by measuring the background temperature and substituting it into Planck's formula. Gas temperature It can be approximated by ambient temperature or measured by other sensors;

[0029] Step 6: Pixel-by-pixel calculation and concentration image generation: Traverse each pixel in the infrared image and repeat the calculation process from Step 1 to Step 5 above. Recombine the column concentration values ​​n calculated for each pixel to generate a gas concentration distribution image.

[0030] In one embodiment, step six involves constructing the input feature set of a CNN-LSTM hybrid network based on CEEMD-VMD quadratic decomposition, including the following steps: Constructing a four-dimensional spatiotemporal dataset: From the preprocessed and contour-extracted infrared image sequence, a spatiotemporal dataset is obtained. Assuming that a gas concentration distribution image is obtained at each time point t, with image size [H, W] (height × width), then after T time points, a four-dimensional dataset is constructed. This represents a dataset containing T time steps, each time step being a single-channel image (concentration field) of size H×W (each pixel has a concentration value); CEEMDAN decomposition: For the concentration time series X(i,j) at position (i,j), the CEEMDAN algorithm is used to decompose it into a series of intrinsic mode functions (C-IMF) and a residual. This adaptively decomposes the original non-stationary and nonlinear concentration series into subsequences at different frequency scales, removes noise, and separates features with different trends. Calculate sample entropy: Calculate the sample entropy of each C-IMF component to quantify the complexity and regularity of that component. The lower the sample entropy value, the more regular the sequence. Sequence Reconstruction: Based on the sample entropy values, all C-IMF components and residuals are reconstructed into three new sequences to simplify subsequent processing: High-frequency sequence: composed of the superposition of several C-IMF components with the largest sample entropy values, representing the random noise and drastic fluctuations in the sequence; Low-frequency sequence: composed of the superposition of C-IMF components with intermediate sample entropy values, representing the short-term regular changes in the sequence; Trend sequence: composed of the superposition of the C-IMF component with the smallest sample entropy value and the residuals, representing the long-term trend of the sequence. VMD secondary decomposition: The reconstructed high-frequency and low-frequency sequences are decomposed in two steps. For the high-frequency and low-frequency sequences at each spatial location, the variational mode decomposition algorithm is used to further decompose them to obtain q new intrinsic mode sequences with more accurate frequency bandwidth, denoted as V-IMF. Feature concatenation: For each time point t and each spatial location (i, j), its original concentration value is concatenated with the values ​​of all q V-IMFs at that time to form a multi-feature vector; Constructing spatiotemporal samples: Training samples are constructed using the sliding time window method. The time window length Δt and the sliding step size S are set. For a sample, the input is the set of feature vectors (a four-dimensional tensor) of all locations in the entire space within the past Δt time steps, and the output is the image of the entire gas concentration distribution in the next one (or several) time steps.

[0031] In one embodiment, the specific steps for constructing the time-series prediction model for gas concentration are as follows: CNN Module (Capturing Spatial Dependencies): Input: The feature map at each time step t is a 3D tensor [H, W, (1+q)], which can be viewed as a "multi-channel" image, where each channel represents a feature (such as the original concentration, V-IMF1, V-IMF2...); Operation: Slide the convolutional layers in the spatial dimension using 3D convolution (3D CNN) or 2D convolution (processed separately for each time step). The convolutional layers automatically extract local spatial features, and the pooling layers perform feature dimensionality reduction; Output: For each time step, the CNN outputs a highly abstract spatial feature vector or feature map; LSTM Module (Capturing Temporal Dependencies): Input: Spatial feature vectors extracted by CNN at Δt time steps are input into the LSTM layer in chronological order; Operation: The gating mechanism of the LSTM unit (input gate, forget gate, output gate) can learn long-term dependencies in the time series, remember and forget relevant information, and capture the dynamic changes in gas outbursts; Output: The output state of the last time step of the LSTM encodes the spatiotemporal evolution information of the entire spatial concentration field over the past Δt time period; Fully connected layers and prediction: The output of the LSTM module is mapped through one or more fully connected layers. Deconvolution or upsampling layers are used to reconstruct the feature vector into a predicted image consistent with the original spatial dimensions [H, W], i.e., a three-dimensional spatial concentration distribution prediction map for future time steps. Model training and validation: Mean squared error (MSE) or mean absolute error (MAE) is used as the loss function to minimize the difference between the predicted concentration field and the true concentration field. Automated hyperparameter search methods such as Bayesian optimization are used to optimize the network structure (such as the number of convolutional kernels, the number of LSTM units), learning rate, window size Δt, etc., to improve model performance. Historical infrared spectral monitoring data is used to divide the training set, validation set, and test set in chronological order to train and validate the model.

[0032] In one embodiment, the method further includes a visualization step of the inversion results, which involves overlaying the three-dimensional gas cloud contour as a pseudo-color layer onto the underground digital twin model of the coal mine.

[0033] 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 variations 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 inverting the outline of coal mine gas clouds using infrared spectroscopy, characterized in that, Includes the following steps: Step 1: Construct a high-dust environment simulation experimental platform in the mine to simulate the formation environment of underground gas clouds; Step 2: Use an infrared spectroscopy detection device to acquire infrared image data containing gas clouds; Step 3: Denoise the infrared image data. The denoising process includes using wavelet transform to remove water vapor noise and using median filtering to remove dust noise. Step 4: Perform temperature drift correction and radiometric calibration on the denoised infrared image, and convert the DN value into radiance; Step 5: Based on the denoised images of multiple frames, the Methane-YOLO model is used to perform target detection and contour extraction on the gas cloud. By matching feature points from multiple perspectives and combining bundle adjustment to optimize the three-dimensional coordinates, a point cloud model is generated and the surface is reconstructed. Step 6: Based on the three-layer radiative transfer model and Lambert-Beer's law, invert the methane concentration; Step 7: Based on the CEEMD-VMD quadratic decomposition and CNN-LSTM hybrid network, jointly train the spatial contour and time series data to construct a gas concentration time series prediction model and predict the three-dimensional spatial concentration distribution.

2. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 1, characterized in that, The mine high-dust environment simulation experimental platform includes a gas concentration control system, a dust concentration control system, a temperature control system, a humidity control system, an air pressure control system, and a data acquisition system. The gas concentration control system prepares methane standard gas of different concentrations to establish a quantitative relationship model between gas concentration and infrared spectral response, and to test the inversion accuracy of the algorithm at different concentrations. It consists of a CH4 gas source, an N2 gas source, a mass flow controller and a flow display instrument. By adjusting the two mass flow controllers, the flow ratio of CH4 and N2 is controlled to ensure thorough mixing in the mixing chamber, thereby preparing CH4 gas with a preset volume fraction. The dust concentration control system simulates a high-dust environment underground and studies the effects of dust on infrared light shielding, refraction, and scattering. It includes a dust emitter, coal dust raw material, turbine fan, and dust concentration measuring instrument. The temperature control system is used to control the background temperature and gas temperature of the experimental environment. A heating tape is evenly wrapped around the background plate, and the power of the heating tape is precisely controlled by the temperature controller to stabilize the temperature of the background plate at a preset value. The humidity control system simulates the high humidity environment caused by dust suppression spraying in the well, studies the interference of the absorption band of water vapor at 3.2μm on the characteristic absorption peak of methane, generates water mist by an ultrasonic humidifier, monitors the ambient humidity by a humidity recorder, and forms a closed-loop control to stabilize the humidity at a preset value. The pressure control system controls the experimental pressure to standard atmospheric pressure, eliminating the influence of pressure changes on the gas absorption coefficient and inversion results, and monitors them in real time through a barometer. The data acquisition system collects raw infrared spectral image data under all control variables.

3. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 1, characterized in that, The specific steps for denoising the infrared image data are as follows: Wavelet transform for noise reduction: Two-dimensional wavelet basis construction: The one-dimensional scaling function u(x) and wavelet function v(x) are extended to a two-dimensional form, as follows: ； ； ； ； in, , , These represent wavelet bases in the horizontal, vertical, and diagonal directions, respectively. Image decomposition: Perform discrete wavelet transform on an infrared image f(x,y) of size M×N to obtain approximation coefficients. and detail coefficient , represented as ; Thresholding: A soft thresholding function is used to denoise the detail coefficients, expressed as follows: Where T is the threshold; Image reconstruction: Reconstructing the denoised image using inverse discrete wavelet transform: ; Median filtering: Sliding window selection: Select a 3×3 or 5×5 window based on the noise type; Pixel replacement: Sort the pixel values ​​within the window and replace the center pixel with the median value; Traverse the image: Slide the window in row-major order to cover the entire image.

4. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 1, characterized in that, The process of performing temperature drift correction and radiometric calibration on the denoised infrared image, converting the DN value into radiance, includes the following steps: Non-imaging area monitoring: Continuously acquire pixel output values ​​of non-imaging areas in infrared images and record pixel jump values ​​of non-imaging areas at different temperatures; Correction factor calculation: Determine the linear correction factor k based on the relationship between the jump value and temperature change; Real-time correction: Applying the formula to each pixel in the imaging area ,in, The corrected pixel value. For the measured pixel values, Here, T represents the temperature coefficient, and T represents the current ambient temperature. For reference temperature; Blackbody calibration experiment: Use a blackbody radiation source to set different temperatures and record the corresponding DN values; Least squares fitting: Substitute the values ​​from the blackbody calibration experiment into the formula. ,in, Let i be the entrance pupil radiance. Here is the corrected DN value for band i. Using the scaling coefficients, solve the system of equations to obtain a and b; Radiance calculation: Radiance is calculated based on the corrected DN value.

5. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 1, characterized in that, The method for target detection and contour extraction of gas clouds based on multi-frame denoised images using the Methane-YOLO model includes the following steps: Input processing: The denoised infrared image is adjusted to 448×448 pixels and normalized to the [0,1] range; Feature extraction: Multi-scale features are extracted through the C2f-SPPF backbone network, and the feature representation is enhanced by combining the global attention mechanism C2PSA. Detection head prediction: Each grid cell predicts 3 bounding boxes and 6 class probabilities; Output dimension: Output = S × S × (B × 5 + C), where S is the number of grids, B is the number of bounding boxes per grid, and C is the number of target categories in the mine; Non-maximum suppression: Merge overlapping bounding boxes, retain high-confidence detection results, and represent the bounding box coordinate prediction as follows: Where W=448 and H=448 are the input image dimensions; Contour extraction: Based on the detected bounding boxes, extract the contour of the gas cloud and generate a binary mask.

6. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 5, characterized in that, The process of generating a point cloud model and reconstructing a surface by matching feature points from multiple perspectives using geometric means, combined with bundle adjustment to optimize 3D coordinates, includes the following steps: Feature extraction: Extract SIFT feature points from multi-view images and generate descriptors; Coarse matching: Feature points are matched based on descriptor similarity, and pairs whose nearest neighbor distance is less than 0.8 compared to the second nearest neighbor distance are retained; Geometric verification: Estimate the fundamental matrix F using the RANSAC algorithm and remove outliers; Exact matching: Matching pairs that satisfy the fundamental matrix constraints are retained as input for subsequent 3D reconstruction. These fundamental matrix constraints are expressed as follows: Where F is the basic matrix, and For matching point pairs; Initialization: Initial 3D point coordinates and camera pose are calculated using triangulation methods; Constructing the error term: For each pair of matched points, calculate the reprojection error between the observed point and the predicted point, whereby the reprojection error is expressed as... ,in, For the observation point, For prediction points; Optimization solution: The Levenberg-Marquardt algorithm is used to minimize the total error, and the camera pose and 3D point coordinates are adjusted iteratively. Convergence criterion: Stop when the error change is less than the threshold or when the maximum number of iterations is reached; Point cloud generation: Save the optimized 3D point coordinates as a point cloud file; Normal estimation: The normal direction of each point is calculated using the K-nearest neighbor algorithm; Triangulation: A greedy projection triangulation algorithm is used to connect adjacent points and generate a triangular mesh; Surface Reconstruction: A smooth surface model is generated using the Poisson reconstruction algorithm; Post-processing: Remove isolated points, repair holes, and optimize the model topology.

7. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 6, characterized in that, The specific steps for inverting methane concentration based on the three-layer radiative transfer model and Beer-Lambert law are as follows: Step 1: Data Preparation and Preprocessing: Acquire the data after denoising in Step 3 and temperature drift correction and radiometric calibration in Step 4. At this point, the original infrared image DN values ​​have been converted into entrance pupil radiance values ​​for each pixel in a specific band. ; Step 2: Construct a three-layer radiative transfer physical model: A three-layer model—foreground layer, target layer, and background layer—is established to describe the infrared radiation transfer process. Foreground layer: refers to the air layer between the infrared spectral imager and the gas cloud; Target layer: the gas cloud layer to be measured, the region containing methane gas; Background layer: the coal wall or tunnel background behind the gas cloud; Total Radiance Received It consists of two parts: the attenuation portion of the background radiation after passing through the target gas layer and the thermal radiation portion of the target gas layer itself, denoted as... ,in, This is the actual radiance received by the infrared system. It is the transmittance of the target gas layer at wavelength λ. It is the radiance of the background layer. It is the thermal radiation brightness of the target gas layer itself; Step 3: Introduce the blackbody radiation law and gas emissivity model: Solve for the unknowns in the three-layer radiative transfer physical model by introducing Planck's blackbody radiation law and the gas emissivity model: The spectral radiance B(λ,T) of a blackbody is calculated using Planck's formula: ,,in, , These are the first and second radiation constants; The gas layer is equivalent to an emissivity According to Kirchhoff's laws, the emissivity of a blackbody is equal to its absorptivity. The thermal radiation of the gas itself is expressed as: ,in, It is the gas at its own temperature Blackbody radiation brightness; Step 4: Establish the relationship between transmittance and concentration using Lambert-Beer's Law: Lambert-Beer's Law establishes the relationship between gas transmittance and its concentration and absorption characteristics: ,in, λ is the absorption coefficient of methane gas at wavelength λ, n is the column concentration of the gas to be inverted, and L is the effective optical path length of infrared light through the gas. Step 5: Derive and solve the concentration inversion algorithm: Substitute the formulas from Steps 3 and 4 into the physical model from Step 2, perform mathematical derivation and simplification, and obtain the inversion formula for column concentration n: ; Step 6: Pixel-by-pixel calculation and concentration image generation: Traverse each pixel in the infrared image and repeat the calculation process from Step 1 to Step 5 above. Recombine the column concentration values ​​n calculated for each pixel to generate a gas concentration distribution image.

8. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 1, characterized in that, In step six, the input feature set of the CNN-LSTM hybrid network is constructed based on the CEEMD-VMD quadratic decomposition, including the following steps: Constructing a four-dimensional spatiotemporal dataset: From the preprocessed and contour-extracted infrared image sequence, a spatiotemporal dataset is obtained. Assuming that each time point t yields a gas concentration distribution image of size [H, W], then after T time points, a four-dimensional dataset is constructed. This represents a dataset containing T time steps, each time step being a single-channel image of size H×W; CEEMDAN decomposition: For the concentration time series X(i,j) at position (i,j), the CEEMDAN algorithm is used to decompose it into a series of intrinsic mode functions (C-IMFs) and a residual; Calculate sample entropy: Calculate the sample entropy of each C-IMF component. The lower the sample entropy value, the more regular the sequence. Sequence Reconstruction: Based on the sample entropy values, all C-IMF components and residuals are reconstructed into three new sequences: High-Frequency Sequence: composed of the C-IMF components with the largest sample entropy values, representing the random noise and drastic fluctuations in the sequence; Low-Frequency Sequence: composed of the C-IMF components with intermediate sample entropy values, representing the short-term regular changes in the sequence; Trend Sequence: composed of the C-IMF component with the smallest sample entropy value and residuals, representing the long-term trend of the sequence. VMD secondary decomposition: The reconstructed high-frequency and low-frequency sequences are decomposed in a secondary manner. For the high-frequency and low-frequency sequences at each spatial location, the variational mode decomposition algorithm is used to further decompose them to obtain q intrinsic mode sequences, denoted as V-IMF. Feature concatenation: For each time point t and each spatial location (i, j), its original concentration value is concatenated with the values ​​of all q V-IMFs at that time to form a multi-feature vector; Constructing spatiotemporal samples: Training samples are constructed using the sliding time window method. The time window length Δt and the sliding step size S are set. For a sample, the input is the set of feature vectors of all locations in the entire space within the past Δt time steps, and the output is the image of the entire gas concentration distribution in the next time step.

9. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 8, characterized in that, The specific steps for constructing the time-series prediction model for gas concentration are as follows: CNN module: Input: The feature map at each time step t is a three-dimensional tensor of [H, W, (1+q)], where each channel represents a feature; Operation: Slide the convolutional layers in the spatial dimension using three-dimensional or two-dimensional convolutions. The convolutional layers automatically extract local spatial features, and the pooling layers perform feature dimensionality reduction. Output: For each time step, the CNN outputs a spatial feature vector or feature map; LSTM Module: Input: Spatial feature vectors extracted by CNN at Δt time steps are input into the LSTM layer in chronological order; Operation: The gating mechanism of the LSTM unit learns long-term dependencies in the time series, memorizes and forgets relevant information, and captures the dynamic changes in gas outbursts; Output: The output state of the last time step of the LSTM encodes the spatiotemporal evolution information of the entire spatial concentration field over the past Δt time period; Fully connected layers and prediction: The output of the LSTM module is mapped through one or more fully connected layers. Deconvolution or upsampling layers are used to reconstruct the feature vector into a predicted image consistent with the original spatial dimensions [H, W], i.e., a three-dimensional spatial concentration distribution prediction map for future time steps. Model training and validation: Mean squared error or mean absolute error is used as the loss function to minimize the difference between the predicted concentration field and the true concentration field. An automated hyperparameter search method is used for optimization. Historical infrared spectral monitoring data is used to divide the training set, validation set and test set in chronological order to train and validate the model.

10. The method for inverting the outline of coal mine gas clouds using infrared spectroscopy according to claim 1, characterized in that, It also includes a visualization process for the inversion results, where the 3D gas cloud outline is overlaid as a pseudo-color layer on the underground digital twin model of the coal mine.

Images