A coal core three-dimensional spatial distribution method and system integrating imaging spectroscopy technology

Abstract

The present invention discloses a coal core three-dimensional spatial distribution method and system integrating imaging spectral technology, belonging to the field of geological exploration technology, obtaining a high-spectral image of a coal core and preprocessing it to obtain a preprocessed high-spectral image; using a spectral angle matching mapping method to map coal and minerals on the high-spectral image to obtain a mapped high-spectral image; based on three-dimensional spatial distribution, the mapped high-spectral image is converted into a three-dimensional model; the three-dimensional point cloud data of the three-dimensional model is input into an improved multi-scale convolution denoising autoencoder network model to detect the coal-rock interface to obtain the coal-rock interface. The three-dimensional spatial distribution technology is used to perform three-dimensional modeling on the coal core, and the coal-rock interface is recognized with the help of the three-dimensional model, thereby accurately evaluating the reserves and quality of coal resources, helping to formulate mining plans, helping to optimize mining methods, improve resource utilization efficiency, and reduce waste.

Description

Technical Field

[0001] The invention belongs to the technical field of geological exploration, and in particular relates to a coal core three-dimensional spatial distribution method and system integrating imaging spectroscopy technology. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] The intelligent development of coal mines is a major trend in the future development of the coal industry. Three-dimensional modeling and visualization technology can vividly and intuitively display geological bodies and geological phenomena, making up for the shortcomings of traditional geological information expression methods. The application of three-dimensional modeling of coal cores is very extensive in the coal industry and is of great significance.

[0004] However, the existing technology cannot perform detailed and transparent exploration when conducting exploration and mineralization analysis based on the three-dimensional model of the coal core. Moreover, coal-rock identification technology is a key technology for realizing intelligent coal mining, and its main purpose is to accurately identify the coal-rock boundary. In recent years, image-based coal-rock identification methods have become a research hotspot and are the most intuitive identification method. However, at present, image recognition can only classify targets in the picture, and cannot clearly define the spatial characteristics of coal and rock, that is, the characteristics of the lithological differences between coal seams and rock layers, and cannot accurately predict the mineralization of coal mines. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a three-dimensional spatial layout method and system of coal core that integrates imaging spectroscopy technology, uses three-dimensional spatial layout technology to perform three-dimensional modeling of the coal core, and uses the three-dimensional model to identify the coal-rock interface, thereby accurately evaluating the reserves and quality of coal resources, helping to formulate mining plans, helping to optimize mining methods, improve resource utilization efficiency, and reduce waste.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0007] In a first aspect, the present invention provides a method for three-dimensional spatial distribution of coal cores by integrating imaging spectroscopy technology, comprising:

[0008] Acquire a hyperspectral image of the coal core and perform preprocessing to obtain a preprocessed hyperspectral image;

[0009] Using a spectral angle matching mapping method to map coal and minerals on the hyperspectral image, to obtain a mapped hyperspectral image;

[0010] Converting the mapped hyperspectral image into a three-dimensional model based on three-dimensional spatial distribution;

[0011] The three-dimensional point cloud data of the three-dimensional model is input into an improved multi-scale convolutional denoising autoencoder network model to detect the coal-rock interface and obtain the coal-rock interface; the improved multi-scale convolutional denoising autoencoder network model includes an input layer, a multi-scale convolutional layer, an encoder module, a bottleneck layer, a decoder module, an output layer, a fusion layer and a fully connected layer which are connected in sequence.

[0012] A further technical solution is to perform image opening operation on the mapped hyperspectral image to form a coal-rock boundary before converting the three-dimensional model.

[0013] A further technical solution is to convert the mapped hyperspectral image into a three-dimensional model based on three-dimensional spatial distribution, specifically:

[0014] Preprocessing the hyperspectral image;

[0015] Extract feature points of hyperspectral images and obtain contour information of hyperspectral images;

[0016] The depth value of each pixel is estimated by depth estimation method, and the hyperspectral image is segmented into different regions to complete the mapping from two-dimensional to three-dimensional.

[0017] The point cloud is created by generating a virtual depth map through the depth estimation method, and then the point cloud data is converted into a triangular network model. The hyperspectral image data is converted into a voxel network using the voxelization method to achieve three-dimensional model reconstruction.

[0018] Optimizing the three-dimensional model and mapping the texture of the hyperspectral image onto the three-dimensional model to complete texture mapping;

[0019] The three-dimensional model is post-processed to render the three-dimensional model.

[0020] A further technical solution is to denoise the hyperspectral image based on an improved low-rank sparse deep network when converting the three-dimensional model. The improved low-rank sparse deep network includes an input layer, a convolution layer, a low-rank sparse layer, a multi-scale convolution layer, and a reconstruction layer connected in sequence. The specific steps are:

[0021] The input layer receives the hyperspectral image, uses multiple convolutional layers to extract the features of the hyperspectral image layer by layer, and uses the Swish activation function and batch normalization after the convolutional layer;

[0022] Entering the low-rank sparse layer, the convolution output is first processed through the low-rank constraint layer to extract low-rank features; then matrix decomposition is performed to decompose the high-dimensional feature matrix into a low-rank matrix and a sparse matrix;

[0023] Multi-scale convolution operations are used in the multi-scale convolution layer to capture image features of different scales, and then the features after low-rank sparse processing are remapped back to the original image space through the reconstruction layer to output the denoised hyperspectral image.

[0024] A further technical solution is to achieve three-dimensional model reconstruction as follows: first, a virtual depth map is generated by a depth estimation method to create a point cloud; a triangulated network is generated from the point cloud by a triangulation method; the image grayscale value is compared with a threshold and converted into a binary voxel grid, and then the hyperspectral image data is mapped to a three-dimensional voxel grid, a three-dimensional array is created to represent the voxel grid, and voxels are set according to the threshold to complete voxel modeling.

[0025] According to a further technical solution, each layer in the encoder module and the decoder module introduces a skip connection through an addition operation.

[0026] A further technical solution is that the decoder module in the improved multi-scale convolutional denoising autoencoder network model includes an upsampling layer and a deconvolution layer connected in sequence, reconstructs the input data through the deconvolution layer, and outputs the predicted coal-rock interface, which is expressed as follows:

[0027]

[0028]

[0029] in, is the reconstructed data block, and Z is the encoded latent feature representation.

[0030] In a second aspect, the present invention provides a coal core three-dimensional spatial distribution system integrating imaging spectroscopy technology, comprising:

[0031] An image acquisition module is configured to: acquire a hyperspectral image of the coal core and perform preprocessing to obtain a preprocessed hyperspectral image;

[0032] An image mapping module is configured to: map the hyperspectral image for coal and minerals using a spectral angle matching mapping method to obtain a mapped hyperspectral image;

[0033] A three-dimensional model building module is configured to: convert the mapped hyperspectral image into a three-dimensional model based on a three-dimensional spatial layout;

[0034] The coal-rock detection module is configured to: input the three-dimensional point cloud data of the three-dimensional model into an improved multi-scale convolutional denoising autoencoder network model to detect the coal-rock interface and obtain the coal-rock interface; the improved multi-scale convolutional denoising autoencoder network model includes an input layer, a multi-scale convolutional layer, an encoder module, a bottleneck layer, a decoder module, an output layer, a fusion layer and a fully connected layer which are sequentially connected.

[0035] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a method for three-dimensional spatial layout of a coal core using a fusion imaging spectroscopy technique as described in the first aspect.

[0036] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a method for three-dimensional spatial layout of a coal core using fusion imaging spectroscopy technology as described in the first aspect are implemented.

[0037] One or more of the above technical solutions have the following beneficial effects:

[0038] After acquiring the image and spectral data, the present invention performs image opening operation and interface detection based on a multi-scale convolutional denoising autoencoder network model after three-dimensional layout, which clearly divides the interface between coal and rock and defines the interface between coal and rock, solving the problem in the prior art that it is impossible to perform refined and transparent exploration during exploration and mineralization analysis.

[0039] The present invention can obtain an intuitive three-dimensional distribution diagram, i.e., a three-dimensional model, based on the hyperspectral image of the coal core. When constructing the three-dimensional model, an improved low-rank sparse deep network is used to denoise the hyperspectral image, thereby improving the denoising effect and restoring the details and quality of the hyperspectral image, and having better adaptability to high-dimensional data; it helps the staff to understand the spatial characteristics of deep coal and the three-dimensional spatial relationship more intuitively and comprehensively, and can conduct analysis and research based on the spatial characteristics of deep coal and the three-dimensional spatial relationship.

[0040] The present invention uses an improved multi-scale convolution denoising autoencoder network model to detect the coal-rock interface of the three-dimensional model of the coal core to obtain the coal-rock interface. The model performs better when processing complex and blurred images, reduces background interference through the denoising mechanism, has strong robustness, and solves the integrity problem of the model at the same time, with good overall performance. According to the clearly divided interface between coal and rock, the reserves and quality of coal resources can be accurately evaluated, which helps to formulate mining plans, optimize mining methods, improve resource utilization efficiency, and reduce waste.

[0041] The present invention can provide intuitive geological information support for mine design and engineering construction, reduce construction risks and improve safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0043] Figure 1 It is a flow chart of the three-dimensional spatial layout method of coal core according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0045] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0046] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0047] Terminology explanation:

[0048] The coal core is obtained by sampling with advanced drilling equipment. It is a columnar coal core several meters or even more than ten meters deep. The three-dimensional model constructed based on the coal core is used to evaluate the coal reserves and calculate the volume of the coal mine.

[0049] Embodiment 1

[0050] like Figure 1 As shown, this embodiment discloses a coal core three-dimensional spatial distribution method integrating imaging spectroscopy technology, the method comprising the following steps:

[0051] S1: Acquire a hyperspectral image of the coal core and perform preprocessing to obtain a preprocessed hyperspectral image;

[0052] In this embodiment, the coal core is collected and the drilling coordinates, orientation, and depth of the coal core are recorded, and the coal core is placed on an imaging hyperspectral scanning platform for spectral scanning to obtain a hyperspectral image of the coal core. Since the scanning of the coal core is a planar scanning, the image is cropped to retain only the area where the coal core is located.

[0053] The hyperspectral image of the coal core is preprocessed, specifically: denoising through reflectivity reconstruction and reflectivity spectrum curve. Spectral reflectivity reconstruction is to obtain the standard calibration plate data while spectral scanning, and use the brightness value (DN value) and reflectivity value of the standard calibration plate image pixel to calculate the coefficients k and b required for calculation, establish the relationship between the hyperspectral image DN value and reflectivity, and use the empirical linear method to calculate the reflectivity. The processed spectrum curve is further processed for noise removal.

[0054] S2: mapping coal and minerals on the hyperspectral image using a spectral angle matching mapping method to obtain a mapped hyperspectral image;

[0055] In this embodiment, first, a coal and mineral extraction benchmark library is constructed, and certain possible mineral types are added according to research needs.

[0056] Then, the spectral angle matching method is used to map the coal and minerals of the coal core. The spectral angle matching method is to compare each pixel with each spectrum in the reference spectral library. The calculation formula is as follows:

[0057]

[0058] Where S is the spectrum to be matched, R is the reference spectrum, and · represents the dot product. Represents the Euclidean norm (i.e., the modulus of a vector). The smaller the spectral angle value, the higher the matching degree, indicating that the two are more similar.

[0059] Before converting the three-dimensional model, the hyperspectral image after mapping is opened to form the coal-rock boundary, which facilitates the subsequent detection of the coal-rock interface. The image opening operation is to first perform an erosion operation on the image to remove all substances except the substance with the largest proportion in the pixel, and then perform an expansion operation to expand the substance with the largest proportion in the pixel and fill it into the entire pixel to form the coal-rock boundary.

[0060] Specifically, the erosion operation can be used to eliminate small and meaningless objects and shrink or refine the image. The erosion operator is , A is expressed by corrosion of B as , as follows:

[0061] Among them, A represents the input image, B represents the structural element, and x represents the coordinates of the output image.

[0062] The dilation operation can be used to expand some components of the image and fill the holes in the image that are smaller than the target structural element. The dilation operator is , A is expanded by B , as follows:

[0063]

[0064] Among them, A represents the input image, B represents the structural element, and x represents the coordinates of the output image.

[0065] S3: converting the mapped hyperspectral image into a three-dimensional model based on three-dimensional spatial distribution;

[0066] In this embodiment, the three-dimensional space layout is specifically performed as follows:

[0067] S301: Preprocess the hyperspectral image, that is, denoise the hyperspectral image based on an improved low-rank sparse deep network. This method can improve the denoising effect, restore the details and quality of the hyperspectral image, and has better adaptability to high-dimensional data.

[0068] The improved low-rank sparse deep network includes an input layer, a convolutional layer, a low-rank sparse layer, a multi-scale convolutional layer, and a reconstruction layer connected in sequence. The specific steps are:

[0069] 1) Constructing a dataset: First, standardize and normalize the hyperspectral image, and add noise to the image by simulating real noise data to construct a dataset as training data.

[0070] 2) Design the network structure: The input layer receives the hyperspectral image, which is represented in three-dimensional form, two of which are spatial dimensions (width and height) and the third dimension is the spectral dimension (number of bands).

[0071] The convolution layer uses multiple convolution layers to extract image features layer by layer. Each convolution layer performs convolution operations on the input hyperspectral data through different convolution kernels to extract spatial and spectral features. After using the convolution layer, the Swish activation function is used to introduce nonlinearity. The specific formula is as follows:

[0072]

[0073] in, is the input, is the sigmoid function, and the formula is: .

[0074] The Swish activation function shows higher accuracy than the traditional ReLU in deep learning tasks; because Swish is a continuous and smooth function, it will not cause the gradient corresponding to negative input values ​​to be zero like ReLU, reducing the "dead neuron" phenomenon; Swish has better gradient information during forward and backward propagation, especially when training deep neural networks, which helps to improve the stability and convergence speed of training.

[0075] Batch normalization is applied after each convolution layer to improve training stability and accelerate convergence. The formula is as follows:

[0076]

[0077] in, is the output of a data point in the batch; and is the mean and variance of the feature in the batch; is a small constant used to avoid division by zero errors, is a standardized value.

[0078] The next step is to enter the low-rank sparse layer. First, the convolution output is processed through the low-rank constraint layer (SVD layer) to extract low-rank features; then matrix decomposition is performed to decompose the high-dimensional feature matrix X into a low-rank matrix L and a sparse matrix S to enhance the denoising ability. The formula is as follows:

[0079]

[0080] Among them, L represents the low-rank part and S represents the sparse part.

[0081] Then use sparse representation technology to process sparse items, and process S through regularization method. The formula is as follows:

[0082]

[0083] in, is the L1 norm of S.

[0084] The next step is to conduct deep residual learning. By learning the residual between the input and output, the network can more effectively learn the noise part of the image, alleviate the gradient vanishing problem, reduce the difficulty of network training, and improve the denoising performance.

[0085] Further considering the spatial and spectral characteristics of hyperspectral images, multi-scale feature fusion is performed in the multi-scale convolution layer, and multi-scale convolution operations can be used to capture image features of different scales, thereby better processing local and global structures in the image.

[0086] Then it is processed through the reconstruction layer: the reconstruction layer remaps the features after low-rank sparse processing back to the original image space, and then outputs the denoised hyperspectral image.

[0087] Loss function: Use image reconstruction error as the loss function to evaluate the denoising effect, and use an optimization algorithm to minimize the loss function.

[0088] The present invention improves low-rank sparse deep networks: 1) Use low-rank decomposition (SVD) to replace the fully connected layer or convolution layer in the deep network. Traditional fully connected layers and convolution layers usually have a large number of parameters, while low-rank decomposition decomposes a large matrix into several small matrices, thereby reducing the number of network parameters, computational complexity and memory consumption. 2) Add a deep residual learning module to alleviate the gradient vanishing problem and improve training efficiency. By learning the residual instead of directly learning the target output, the network can be more easily optimized, thereby accelerating the convergence process and improving the denoising effect. For hyperspectral images, which have complex information and high dimensions, the residual module helps the network focus on removing noise while retaining important structural features. 3) Add multi-scale feature fusion to help the network capture noise features and image details at different scales at the same time, thereby more effectively denoising. Enhance the synergy of local and global features: By fusing information at different scales, the network can capture local details and global structures at the same time, thereby better balancing noise removal and retaining image details during denoising. 4) The Swish activation function is used when introducing nonlinearity.

[0089] 3) Network training: Divide the data set into a training set and a test set. Use the training set to train the network, and use the test set to evaluate the model. Use the gradient descent method to update the network weights, and improve the denoising performance by minimizing the loss function. Repeat multiple times until the model converges.

[0090] Using Canny edge detection, the specific steps are as follows: convert the image into a grayscale image, apply a Gaussian filter to smooth the image, use the Sobel operator to calculate the gradient amplitude and direction of the pixel, suppress non-edge points in the gradient amplitude image, apply high and low double thresholds for edge connection, retain the real edges and remove isolated weak edges, and finally achieve edge detection, which will help to extract the contour and features of the image later.

[0091] S302: extracting feature points of the hyperspectral image and obtaining contour information of the hyperspectral image.

[0092] Harris corner detection is used to detect feature points. Specifically, the image is grayed, the horizontal and vertical gradients are calculated, the covariance matrix of the gradient matrix is ​​constructed, and then the Harris response function is used to measure the intensity of the focus, and then the image is non-maximum suppressed, and a threshold is applied to select corners whose response function values ​​are greater than a certain threshold to filter out weak response corners. Finally, the detected corners are marked in the image; then contours are extracted through Hough transform and OpenCV's findContours function. Harris corner detection is to identify and extract key point features in an image and capture local feature points.

[0093] S303: Estimate the depth value of each pixel by using the depth estimation method, and segment the hyperspectral image into different regions to complete the mapping from two-dimensional to three-dimensional. Specifically, the depth value of each pixel is estimated based on the depth estimation method, and the graph cut method is used to segment different regions in the image to provide a basis for three-dimensional modeling.

[0094] S304: Generate a virtual depth map through a depth estimation method to create a point cloud, then convert the point cloud data into a triangular network model, and use a voxelization method to convert the image data into a voxel network to achieve three-dimensional model reconstruction.

[0095] Specifically, first, a virtual depth map is generated by the depth estimation method to create a point cloud; the point cloud is used to generate a triangular network through the Delaunay triangulation method; the image grayscale value is compared with the threshold and converted into a binary voxel grid. After mapping the hyperspectral image data to a three-dimensional voxel grid, a three-dimensional array is created to represent the voxel grid, and the voxel is set to "activated" or "inactivated" according to the threshold to complete the voxel modeling.

[0096] S305: Optimizing the three-dimensional model, and mapping the texture of the hyperspectral image onto the three-dimensional model to complete texture mapping.

[0097] Model optimization is as follows: smoothing the surface of the model to reduce noise and irregularities in the model. Rebuilding and patching the model through tools such as hole filling and mesh repair to repair missing parts or errors in the model. Texture mapping is as follows: mapping the texture of a two-dimensional image to a three-dimensional model. First, unfold the surface of the three-dimensional model onto a two-dimensional plane, generate UV coordinates, ensure that the coordinates do not overlap and that the texture image correctly covers the model surface, use Photoshop to create and edit texture images and ensure that their resolution is appropriate and not distorted, map the texture image to the UV coordinates, assign the texture image to the correct material and apply it to the model, check the texture, and adjust and optimize the texture image or UV coordinates.

[0098] S306: Post-process the three-dimensional model and render the three-dimensional model, such as shadows, light sources, etc.

[0099] S4: inputting the three-dimensional point cloud data of the three-dimensional model into an improved multi-scale convolutional denoising autoencoder network model to detect the coal-rock interface, thereby obtaining the coal-rock interface.

[0100] In this embodiment, by analyzing the three-dimensional model, the coal-rock interface of the three-dimensional model is detected based on the improved multi-scale convolutional denoising autoencoder network model (MS-CDAE), the coal-rock interface is found, and the mineralization prediction of the coal mine is carried out.

[0101] The specific steps are:

[0102] S401: Construct a data set: Preprocess the 3D point cloud data into a form suitable for network input. Usually, it is divided into multiple small blocks and labels are generated for the data to indicate the location and category of the interface (the category is the coal-rock interface). The data processed and labeled data are constructed into a data set and divided into a training set and a test set.

[0103] S402: Design network structure: The improved multi-scale convolutional denoising autoencoder network model includes an input layer, a multi-scale convolutional layer, an encoder module, a bottleneck layer, a decoder module, an output layer, a fusion layer and a fully connected layer connected in sequence.

[0104] 1) After the input layer receives the 3D volume data block (3D point cloud data), it is input into the multi-scale convolution layer. The introduction of learnable multi-scale convolution kernels means that in the convolution operation, a single fixed-scale convolution kernel is no longer used. Instead, the network is allowed to adaptively select different convolution kernel scales based on the input data to capture features at different scales. The specific method is as follows:

[0105] Design convolution kernels of multiple scales, such as 3*3*3, 5*5*5, 7*7*7, etc. For each scale, define a convolution kernel and apply it to the input, and then fuse the feature maps of different scales by weighted summation; for each convolution kernel, design a weight vector W, which determines the importance of the convolution kernel of each scale. The larger the weight value, the more important the convolution kernel of this scale occupies in the final feature representation. In this way, the network can adaptively adjust the contribution of convolution kernels of different scales according to the input data.

[0106] Assume that the input feature map is M, and there are N convolution kernels of different scales, and the size of each convolution kernel is (i=1,2,3...N), and then perform the following steps. For each scale of convolution kernel , perform convolution operation:

[0107]

[0108] Among them, * represents the convolution operation, It is the output feature map of the corresponding scale convolution.

[0109] All feature maps of different scales are weighted fused to finally obtain the feature map:

[0110]

[0111] in, is the learned weight parameter, which indicates the relative importance of convolution kernels of different scales. It is learned through network training, and usually softmax or other normalization methods are used to ensure the effectiveness of the weights.

[0112] 2) After the multi-scale convolutional layer outputs, the convolutional layer of the encoder module (including the convolutional layer and pooling layer connected in sequence) is connected to further extract features. In the encoder, each layer introduces skip connections through addition operations to avoid gradient vanishing and accelerate training. The details are as follows:

[0113] The encoder input is ; Perform convolution operation to generate feature map ; Then perform residual connection to input And the convolution output Add together and get ; Add a pooling layer to reduce the feature map size to reduce noise and repeat the encoder module process.

[0114] 3) After multiple convolutions and pooling, the output is sent to a bottleneck layer, representing the compressed feature space.

[0115] 4) Next, connect to the decoder module (including the upsampling layer and deconvolution layer connected in sequence). When introducing residual connections, each layer in the decoder introduces skip connections through addition operations to avoid gradient vanishing and accelerate training. The specific operations are as follows:

[0116] The upsampling layer can restore the image scale at a lower computational cost, and the deconvolution layer can be refined on this basis to improve the image quality. The input to a layer of the decoder is added to the layer of the deconvolution output (similar to the operation in the encoder, which will not be repeated here), and the residual connection ensures that the information flows more smoothly in the network. The bottleneck layer is connected to the upsampling layer to restore the size of the feature map, and the deconvolution layer is used to restore the size of the input data.

[0117] Furthermore, the encoder module of this embodiment is a denoising autoencoder, and the decoder module includes a deconvolution layer, which mainly extracts multi-scale features by using multiple convolution kernels of different sizes, reduces the dimension based on the encoder-decoder architecture to obtain high-level feature representation, and reconstructs the original input data for denoising.

[0118] The denoising autoencoder extracts features through convolutional layers and pooling layers to reduce noise; the decoder reconstructs the input data block through the deconvolution layer and outputs the predicted coal-rock interface. The formula is as follows:

[0119]

[0120]

[0121] in, is the reconstructed data block, and Z is the encoded latent feature representation.

[0122] 5) The decoder module is connected to the output layer and outputs the reconstructed image.

[0123] 6) Feature fusion and interface detection: The fusion layer fuses the output feature maps of different scales to integrate multi-scale information. The formula is as follows:

[0124]

[0125] in, is the fused feature map, Represented as feature maps of different scales, Indicates splicing.

[0126] Then, the interface detection is performed, and the final classification is performed through the fully connected layer to detect the interface in the 3D model. The specific formula is as follows:

[0127]

[0128] Where P is the predicted probability of each voxel, Represented as a fully connected layer for classification.

[0129] S403: Loss function and optimization: The cross entropy loss function is used to evaluate the accuracy of interface detection. The formula is as follows:

[0130]

[0131] in, is the i-th true label, is the predicted probability, is a network parameter.

[0132] Then optimize and use the Adam optimization algorithm to adjust the network parameters to minimize the loss function.

[0133] The present invention aims at improving the multi-scale convolution denoising autoencoder network model: 1) When performing multi-scale convolution operations, a learnable multi-scale convolution kernel is introduced instead of a fixed-scale convolution kernel. The network can not only use convolution kernels of different scales in each convolution layer, but also adaptively adjust the use of convolution kernels of different scales according to the characteristics of the task and input data, thereby improving the model's perception of multi-scale features. Ultimately, the learning ability of the network is enhanced, and complex patterns and features can be better captured. 2) Introducing residual connections in the encoder and decoder. Each layer in the encoder and decoder can introduce jump connections through addition operations to avoid gradient vanishing and accelerate training. The basic idea of ​​the residual connection is to add the input directly to the output through a "shortcut" connection, so that the network can learn the residual instead of the complete mapping. The network can propagate gradients more effectively during training, thereby alleviating the gradient vanishing problem in deep networks and improving training stability and the learnability of network depth.

[0134] Through the above steps, the improved multi-scale convolutional denoising autoencoder network model can effectively detect the coal-rock interface in three-dimensional data, and provide refined guidance for coal reserve estimation based on the restoration ratio of the three-dimensional model.

[0135] Embodiment 2

[0136] This embodiment provides a coal core three-dimensional spatial distribution system integrating imaging spectroscopy technology, including:

[0137] An image acquisition module is configured to: acquire a hyperspectral image of the coal core and perform preprocessing to obtain a preprocessed hyperspectral image;

[0138] An image mapping module is configured to: map the hyperspectral image for coal and minerals using a spectral angle matching mapping method to obtain a mapped hyperspectral image;

[0139] A three-dimensional model building module is configured to: convert the mapped hyperspectral image into a three-dimensional model based on a three-dimensional spatial layout;

[0140] The coal-rock detection module is configured to: input the three-dimensional point cloud data of the three-dimensional model into an improved multi-scale convolutional denoising autoencoder network model to detect the coal-rock interface and obtain the coal-rock interface; the improved multi-scale convolutional denoising autoencoder network model includes an input layer, a multi-scale convolutional layer, an encoder module, a bottleneck layer, a decoder module, an output layer, a fusion layer and a fully connected layer which are sequentially connected.

[0141] In this embodiment, the image acquisition module includes a hyperspectral sensor, a coal core scanning platform, and a dedicated light source. The object to be measured (coal core) is placed on the coal core scanning platform, and data of the coal core is collected by the hyperspectral sensor.

[0142] The system also includes a result display and analysis module, which displays the final three-dimensional model, analyzes it, queries the coal-rock interface, and makes coal mine mineralization predictions; at the same time, it provides detailed guidance on coal mine reserve estimation based on the restoration ratio of the three-dimensional model.

[0143] Embodiment 3

[0144] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of embodiment 1 when executing the program.

[0145] Embodiment 4

[0146] The purpose of this embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the steps of the method of embodiment 1 are performed.

[0147] The steps involved in the apparatus of the above embodiments 3 and 4 correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0148] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0150] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A coal core three-dimensional spatial distribution method integrating imaging spectroscopy technology, characterized in that: include: Acquire a hyperspectral image of the coal core and perform preprocessing to obtain a preprocessed hyperspectral image; Using a spectral angle matching mapping method to map coal and minerals on the hyperspectral image, to obtain a mapped hyperspectral image; Converting the mapped hyperspectral image into a three-dimensional model based on three-dimensional spatial distribution; Inputting the three-dimensional point cloud data of the three-dimensional model into an improved multi-scale convolutional denoising autoencoder network model to detect the coal-rock interface, thereby obtaining the coal-rock interface; the improved multi-scale convolutional denoising autoencoder network model comprises an input layer, a multi-scale convolutional layer, an encoder module, a bottleneck layer, a decoder module, an output layer, a fusion layer and a fully connected layer connected in sequence; The step of converting the mapped hyperspectral image into a three-dimensional model based on the three-dimensional spatial distribution is specifically as follows: Preprocessing the hyperspectral image; Extract feature points of hyperspectral images and obtain contour information of hyperspectral images; The depth value of each pixel is estimated by depth estimation method, and the hyperspectral image is segmented into different regions to complete the mapping from two-dimensional to three-dimensional. The point cloud is created by generating a virtual depth map through the depth estimation method, and then the point cloud data is converted into a triangular network model. The hyperspectral image data is converted into a voxel network using the voxelization method to achieve three-dimensional model reconstruction. Optimizing the three-dimensional model and mapping the texture of the hyperspectral image onto the three-dimensional model to complete texture mapping; Post-processing the three-dimensional model to render the three-dimensional model; When converting the three-dimensional model, the hyperspectral image is denoised based on an improved low-rank sparse deep network, wherein the improved low-rank sparse deep network includes an input layer, a convolution layer, a low-rank sparse layer, a multi-scale convolution layer, and a reconstruction layer connected in sequence, and the specific steps are as follows: The input layer receives the hyperspectral image, uses multiple convolutional layers to extract the features of the hyperspectral image layer by layer, and uses the Swish activation function and batch normalization after the convolutional layer; Entering the low-rank sparse layer, the convolution output is first processed through the low-rank constraint layer to extract low-rank features; then matrix decomposition is performed to decompose the high-dimensional feature matrix into a low-rank matrix and a sparse matrix; In the multi-scale convolution layer, a multi-scale convolution operation is used to capture image features of different scales. Then, the features after low-rank sparse processing are remapped back to the original image space through the reconstruction layer, and the denoised hyperspectral image is output; The three-dimensional model reconstruction is specifically implemented as follows: first, a virtual depth map is generated by a depth estimation method to create a point cloud; a triangulated network is generated from the point cloud by a triangulation method; the image grayscale value is compared with a threshold and converted into a binary voxel grid, and then the hyperspectral image data is mapped to a three-dimensional voxel grid, a three-dimensional array is created to represent the voxel grid, and voxels are set according to the threshold to complete voxel modeling.

2. A method for three-dimensional spatial distribution of coal cores using fusion imaging spectroscopy technology as claimed in claim 1, characterized in that: Before converting the three-dimensional model, an image opening operation is performed on the mapped hyperspectral image to form a coal-rock boundary line.

3. The method for three-dimensional spatial distribution of coal cores by integrating imaging spectroscopy technology as claimed in claim 1, characterized in that: Each layer in the encoder module and the decoder module introduces skip connections through addition operations.

4. The method for three-dimensional spatial distribution of coal cores by integrating imaging spectroscopy technology as claimed in claim 1, characterized in that: The decoder module in the improved multi-scale convolutional denoising autoencoder network model includes an upsampling layer and a deconvolution layer connected in sequence, reconstructs the input data through the deconvolution layer, and outputs the predicted coal-rock interface, which is expressed as follows: in, is the reconstructed data block, and Z is the encoded latent feature representation.

5. A coal core three-dimensional spatial distribution system integrating imaging spectroscopy technology, characterized in that: include: An image acquisition module is configured to: acquire a hyperspectral image of the coal core and perform preprocessing to obtain a preprocessed hyperspectral image; An image mapping module is configured to: map the hyperspectral image for coal and minerals using a spectral angle matching mapping method to obtain a mapped hyperspectral image; A three-dimensional model building module is configured to: convert the mapped hyperspectral image into a three-dimensional model based on a three-dimensional spatial layout; The coal-rock detection module is configured to: input the three-dimensional point cloud data of the three-dimensional model into an improved multi-scale convolutional denoising autoencoder network model to detect the coal-rock interface, and obtain the coal-rock interface; the improved multi-scale convolutional denoising autoencoder network model includes an input layer, a multi-scale convolutional layer, an encoder module, a bottleneck layer, a decoder module, an output layer, a fusion layer and a fully connected layer connected in sequence; The step of converting the mapped hyperspectral image into a three-dimensional model based on the three-dimensional spatial distribution is specifically as follows: Preprocessing the hyperspectral image; Extract feature points of hyperspectral images and obtain contour information of hyperspectral images; The depth value of each pixel is estimated by depth estimation method, and the hyperspectral image is segmented into different regions to complete the mapping from two-dimensional to three-dimensional. The point cloud is created by generating a virtual depth map through the depth estimation method, and then the point cloud data is converted into a triangular network model. The hyperspectral image data is converted into a voxel network using the voxelization method to achieve three-dimensional model reconstruction. Optimizing the three-dimensional model and mapping the texture of the hyperspectral image onto the three-dimensional model to complete texture mapping; Post-processing the three-dimensional model to render the three-dimensional model; When converting the three-dimensional model, the hyperspectral image is denoised based on an improved low-rank sparse deep network, wherein the improved low-rank sparse deep network includes an input layer, a convolution layer, a low-rank sparse layer, a multi-scale convolution layer, and a reconstruction layer connected in sequence, and the specific steps are as follows: The input layer receives the hyperspectral image, uses multiple convolutional layers to extract the features of the hyperspectral image layer by layer, and uses the Swish activation function and batch normalization after the convolutional layer; Entering the low-rank sparse layer, the convolution output is first processed through the low-rank constraint layer to extract low-rank features; then matrix decomposition is performed to decompose the high-dimensional feature matrix into a low-rank matrix and a sparse matrix; In the multi-scale convolution layer, a multi-scale convolution operation is used to capture image features of different scales. Then, the features after low-rank sparse processing are remapped back to the original image space through the reconstruction layer, and the denoised hyperspectral image is output; The three-dimensional model reconstruction is specifically implemented as follows: first, a virtual depth map is generated by a depth estimation method to create a point cloud; a triangulated network is generated from the point cloud by a triangulation method; the image grayscale value is compared with a threshold and converted into a binary voxel grid, and then the hyperspectral image data is mapped to a three-dimensional voxel grid, a three-dimensional array is created to represent the voxel grid, and voxels are set according to the threshold to complete voxel modeling.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a three-dimensional spatial layout method of a coal core using a fusion imaging spectroscopy technology as described in any one of claims 1 to 4 are implemented.

7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the three-dimensional spatial layout method of coal cores fused with imaging spectroscopy technology as described in any one of claims 1-4 are implemented.

Images