Fracture tracking model data preprocessing method based on 4DCT image
By using a data preprocessing method based on 4DCT images for a fracture tracking model, the problems of data volume discontinuity and stratigraphic shift during the rotational loading of coal and rock samples were solved, achieving high-precision fracture tracking and improving the reliability of coal and rock mechanical behavior analysis.
Patent Information
- Application Number
- CN202511093966.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-10-28
AI Technical Summary
During the rotational loading process of coal and rock samples, the data volumes obtained by 4DCT scanning are discontinuous, have shifted stratigraphic positions, and have blurred features, which increases the difficulty of fracture development analysis. In particular, during the rapid expansion and connection stage of fractures, the data volumes overlap and have large angle differences, affecting the accuracy of the analysis.
By synchronously loading and acquiring continuous scan frames from 4DCT scans, data volume reconstruction and slice integration are performed. Circular masks are used to remove noise, and crack alignment and feature enhancement are performed. Preprocessing is combined with a crack tracking model, including crack rotation alignment and grayscale threshold compression. Crack tracking is performed using a convolutional neural network model with a three-level memory architecture.
It significantly improves the identification accuracy of the fracture tracking model, provides a dynamic observation method for coal and rock mechanical behavior, enhances the identification accuracy of micro-fracture initiation and propagation, and solves the problems of data discontinuity and feature ambiguity in traditional methods.
Smart Images
Figure CN120852673A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of 4DCT image data processing, specifically to a data preprocessing method for a crack tracking model based on 4DCT images. Background Technology
[0002] Simultaneous loading and CT scanning of coal and rock samples is a technique that performs CT scans concurrently with the loading of coal and rock samples. During loading, a rotating base and a rotatable indenter control the sample's rotation along its central axis, enabling a comprehensive CT scan. 4DCT incorporates the time factor into the CT scan. Throughout the rotational loading process, the acquired scan frames are arranged chronologically with equal time intervals between adjacent frames. By selecting several consecutive scan frames and reconstructing the data, a three-dimensional coal and rock sample (data volume) at a specific moment during loading can be obtained, thus revealing the fracture development of the sample at that particular instant.
[0003] A large number of CT scan frames are acquired during the entire rotational loading process of the coal and rock sample, such as 24,000 scan frames. Each data volume can be reconstructed based on 600 scan frames. The loading process of the coal and rock sample is generally divided into three stages: the fracture closure stage, the slow fracture development stage, and the rapid fracture propagation and connection stage. Among them, the fracture closure stage and the slow fracture development stage are longer and have more scan frames, while the rapid fracture propagation and connection stage is shorter and has fewer scan frames, but the rapid fracture propagation and connection stage is the focus of research. For example, 20, 50, and 20 data volumes are generally reconstructed in the fracture closure stage, the slow fracture development stage, and the rapid fracture propagation and connection stage, respectively. However, this leads to several issues: the data volumes established during the fracture closure and slow fracture development stages are based on discontinuous scan frames. For example, the first and second data volumes established during the fracture closure stage are based on scan frames 1-600 and 1420-2019, respectively. Conversely, the data volumes established during the rapid fracture propagation and penetration stage are based on overlapping scan frames (dense reconstruction). For instance, the 19th and 20th data volumes established during the rapid fracture propagation and penetration stage are based on scan frames 23050-23649 and 23401-24000, respectively. Furthermore, due to the irregularity of the scan frames used to establish the data volumes, they are generally at different rotation angles. This results in angular differences in slices at the same horizontal stratigraphic level, further increasing the difficulty of analyzing the fracture development process in coal and rock masses. In addition, the reconstructed data volumes may suffer from problems such as non-centered effective data and unclear distinction between fracture features and matrix background. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention proposes a data preprocessing method for a crack tracking model based on 4DCT images, comprising the following steps:
[0005] S101: Place the coal and rock sample between the rotating base and the rotatable pressure head; the coal and rock sample rotates at a set speed, and at the same time, uniaxial compression and 4DCT scanning are performed on the coal and rock sample to obtain a number of continuous scanning frames with fixed time intervals. Based on these scanning frames, several data volumes are reconstructed in three dimensions in chronological order.
[0006] S102: Export CT image slices for each data volume at several selected horizontal levels; integrate CT image slices of the same level for all data volumes into a sample set according to the reconstruction order of the data volumes;
[0007] S103: Use a circular mask to remove the black air image outside the boundary of the coal and rock sample for each CT image slice in the entire sample set, and retain only the valid data of the coal and rock sample.
[0008] S104: For each sample set, the CT image slice of the first data volume is used as the keyframe, and the CT image slices of the remaining data volumes are aligned with the keyframe to achieve the angle alignment of the CT image slices.
[0009] S105: Enhance the features of the cracks in each CT image slice;
[0010] S106: Input the processed sample set into the coal and rock fracture tracking model, and the coal and rock fracture tracking model outputs the entire process of fracture development corresponding to each sample set.
[0011] Among them, steps S101-S104 and S101-S105 can each be considered as an independent technical solution, which is: a data processing method for 4DCT images.
[0012] Preferably, in step S101, the coal and rock sample is a cylinder with a diameter of 50 mm and a height of 100 mm.
[0013] Preferably, in step S101, the rotation speed is 12° / s and the scanning frequency is 20 frames / s.
[0014] Preferably, in step S103, the CT image slices in the sample set are square, and the boundaries of the coal and rock samples are circular; the circular mask automatically detects the center and radius of the coal and rock samples in the CT image slices, and the parts outside the mask are processed as white.
[0015] Preferably, in step S103, a square-format CT image slice is first read, and a grayscale threshold binarization and contour detection algorithm is used to automatically locate the minimum circumcircle of the circular boundary of the coal and rock sample, and accurately calculate its center coordinates and radius parameters. Then, a white background matrix of the same size as the original image is constructed, and the detected circular region pixels are accurately mapped to the new background through a circular mask generation function. Edge noise is automatically eliminated through a radius shrinkage coefficient to ensure that all black air image areas outside the boundary of the coal and rock sample are forced to be pure white.
[0016] Preferably, in step S104, the method of aligning the gaps is to rotate the effective data portion in the CT image slice until it achieves maximum overlap with the gap features of the keyframe.
[0017] Preferably, in step S104, firstly, crack enhancement processing is performed on the keyframe and the CT image slice to be aligned, using a combination of contrast-limited adaptive histogram equalization and morphological closing operation to enhance the contrast between the crack and the matrix background; then, a Fourier frequency domain angle detection algorithm is used to convert the frequency domain features into an angle feature space using linear polar coordinate transformation, and the initial angle deviation is obtained through the phase correlation method; subsequently, a multi-stage angle optimization search strategy is adopted, using three stages of angle traversal (coarse-grained, medium-grained, and fine-grained), combined with parallel computing acceleration technology, and using a weighted score of phase correlation and binarized crack overlap rate as a multi-modal evaluation index to dynamically determine the optimal rotation angle; finally, an anti-aliased rotation algorithm is used to achieve image geometric transformation, and a reflection boundary filling mode is used to maintain image size consistency.
[0018] Preferably, in step S105, the feature enhancement method is to determine the crack or background by using the grayscale threshold in the CT image slice, and to compress the threshold for low brightness.
[0019] Preferably, in step S105, a composite processing module including gamma transform and dynamic threshold enhancement is first constructed to perform grayscale normalization processing on the input image; a dual-modal enhancement strategy is adopted. On the one hand, a binary mask is constructed by setting a low brightness threshold, and exponential nonlinear compression is applied to the crack region with grayscale value below the threshold. The compression intensity is controlled by an enhancement factor to significantly reduce the grayscale overlap between microcracks and matrix background; on the other hand, global gamma correction and brightness gain adjustment are implemented to reshape the image grayscale histogram distribution through a nonlinear mapping function; finally, pixel value truncation is used to maintain numerical stability, forming a synergistic enhancement mechanism of local enhancement of crack region and optimization of overall image contrast.
[0020] Preferably, in step S106, the gap tracing model is a convolutional neural network model based on a three-level memory architecture; it includes four parts: a key encoder, a value encoder, a mask decoder, and a convolutional gated recurrent unit; the key encoder is implemented based on ResNet-50, taking the processed sample set as input and generating multi-scale features; the value encoder is implemented based on ResNet-18, and its core function is to transform the input mask and image features into discriminative object-level feature representations through multi-stage feature fusion and gated memory update mechanisms; the mask decoder uses a multi-level feature fusion and progressive upsampling architecture to achieve high-precision mask generation; the function of the convolutional gated recurrent unit is to realize and update the sensory memory, store the hidden state in continuous time sequence, and thus provide temporal continuity.
[0021] Key technical means and beneficial effects of this invention: This invention obtains the dynamic fracture evolution sequence of coal and rock samples through data volume reconstruction based on synchronous loading and 4DCT scanning (S101). It then eliminates angular deviations and background interference by combining layer slice integration (S102), circular masking for noise reduction (S103), and fracture rotation alignment (S104). Finally, it enhances image contrast using grayscale threshold compression and fracture feature enhancement (S105). The standardized preprocessed sample set (S106) is then input into the model to accurately output the entire spatiotemporal evolution process of fractures. This invention overcomes the technical bottlenecks of discontinuous dynamic CT data, layer shift, and feature ambiguity, significantly improving the accuracy of fracture tracking models in identifying the initiation and expansion of micro-fractures. This provides a highly reliable dynamic observation method for coal and rock mechanical behavior analysis and engineering safety assessment. Attached Figure Description
[0022] Figure 1 This is a flowchart of the data preprocessing method for the crack tracking model based on 4DCT images according to the present invention. Detailed Implementation
[0023] The technical solution of the present invention will now be described in more detail with reference to the accompanying drawings.
[0024] like Figure 1 As shown, this invention proposes a data preprocessing method for a crack tracking model based on 4DCT images, comprising the following steps:
[0025] S101: The coal and rock sample is placed between the rotating base and the rotatable pressure head; the coal and rock sample is rotated at a set speed, and uniaxial compression and 4DCT scanning are performed on the coal and rock sample at the same time to obtain the full process scanning data of the coal and rock sample from the start of loading to failure. The scanning data consists of several consecutive scanning frames with fixed time intervals. Based on these scanning frames, several three-dimensional coal and rock samples (data volumes) are reconstructed in three dimensions in chronological order.
[0026] Specifically, the coal and rock sample is a cylinder with a diameter of 50 mm and a height of 100 mm. The coal and rock synchronous loading and CT scanning system used is an open-tube reflective high-penetration coal and rock synchronous loading and 4DCT system jointly developed by the China Coal Research Institute and Tianjin Sanying Precision Instruments Co., Ltd. The rotation speed of the coal and rock sample or the rotating base is 12° / s, the scanning frequency is 20 frames / s, that is, the number of scanning frames is 600 frames / revolution, and the total rotation angle of the rotating base is set to 14600°. The displacement loading control method is adopted, and the loading is carried out in two stages: first, preload to 1 kN at a speed of 1 mm / min; then load at a speed of 0.1 mm / min until the coal and rock sample is destroyed.
[0027] OffLineRecon software, in conjunction with the VoxelStudioRecon kernel program, was used for data volume reconstruction. OffLineRecon software can reconstruct with precision down to the number of scan frames, while the VoxelStudioRecon kernel program reads the relevant configuration files for reconstruction parameters, thus achieving automatic reconstruction. Based on the three stages of fracture development and evolution during the loading process of coal and rock samples—fracture closure, slow fracture development, and rapid fracture propagation and connection—stage reconstruction was performed. A preferred reconstruction scheme is to reconstruct 20, 50, and 20 data volumes for the fracture closure, slow fracture development, and rapid fracture propagation and connection stages, respectively. That is, 90 data volumes are reconstructed for each coal and rock sample throughout the entire loading process, with each data volume based on 600 scan frames.
[0028] S102: Export two-dimensional CT image slices for each data volume at several selected horizontal levels; integrate CT image slices of the same level for all data volumes into a sample set according to the reconstruction order of the data volumes;
[0029] Specifically, each data volume selects 1500 horizontal slices, meaning each data volume yields 1500 CT image slices. This allows for the creation of 1500 sample sets, each containing a collection of CT image slices exported from all data volumes at one horizontal slice. The tool for exporting CT image slices is the open-source software ImageJ. ImageJ can accept imported .raw files and automatically retrieve information from the reconstructed data volumes. Furthermore, this software provides a slice export function without altering the original data volumes, ensuring the originality of the CT image slices.
[0030] S103: Use a circular mask to remove the black air image outside the boundary of the coal and rock sample for each CT image slice in the entire sample set, and retain only the valid data of the coal and rock sample; the CT image slices in the sample set are all square, and the boundary of the coal and rock sample is circular; the circular mask can automatically detect the center and radius of the coal and rock sample in the CT image slice, and the parts outside the mask are all processed into white.
[0031] Specifically: The removal of black air images outside the boundary of coal and rock samples using a circular mask is automatically implemented in batches using a self-designed Python script. First, square-format CT image slices are read. A grayscale threshold binarization and contour detection algorithm is used to automatically locate the minimum circumcircle of the circular boundary of the coal and rock sample, accurately calculating its center coordinates and radius parameters. Then, a white background matrix of the same size as the original image is constructed. A circular mask generation function accurately maps the detected circular region pixels to the new background. An edge noise is automatically eliminated using a radius shrinkage coefficient, ensuring that all black air image areas outside the coal and rock sample boundary are forced to pure white. Finally, high-quality CT image slices retaining only the valid sample areas are output. A dynamic radius detection mechanism adapts to CT image slices of different sizes, and coordinate boundary protection enhances the algorithm's robustness, achieving fully automated background purification processing and providing standardized data input for subsequent fracture feature analysis.
[0032] S104: For each sample set, the CT image slice of the first data volume is used as the keyframe, and the CT image slices of the remaining data volumes are aligned with the keyframe for crack alignment; the crack alignment method is to rotate the effective data (cracks, high-density minerals, coal matrix, etc.) in the CT image slice until the crack features of the keyframe are maximized.
[0033] Specifically: The crack alignment can be automatically batch-implemented using a self-designed Python script. First, crack enhancement processing is performed on keyframes and CT image slices to be aligned, using a combination of contrast-limited adaptive histogram equalization (CLAHE) and morphological closing operations to enhance the contrast between the crack and the matrix background. Then, a Fourier frequency domain angle detection algorithm is used to convert the frequency domain features into an angle feature space using linear polar coordinate transformation, and the initial angle deviation is obtained through the phase correlation method. Subsequently, a multi-stage angle optimization search strategy is adopted, using three stages of angle traversal: coarse-grained (±180° range), medium-grained (±90° range), and fine-grained (±15° range). Combined with parallel computing acceleration technology, the weighted score of phase correlation and binarized crack overlap rate is used as a multimodal evaluation index to dynamically determine the optimal rotation angle. Finally, an anti-aliased rotation algorithm (ndimage.rotate) is used to achieve image geometric transformation, and a reflection boundary filling mode is used to maintain image size consistency.
[0034] The proposed gap alignment method innovatively integrates frequency domain feature guidance and spatial domain gap matching mechanism. Through dynamic angle search range adjustment and parallel computing framework, it achieves accurate and rapid alignment in scenarios with large angle deviation (±180°), effectively solving the rotational offset problem of 4DCT reconstruction data volume and ensuring accurate matching of gap spatial positions at different time points.
[0035] S105: Enhance the features of the cracks in each CT image slice; the enhancement method is to determine the cracks (low gray level) or background by using the gray level threshold in the CT image slice, and compress the low brightness threshold to enhance the crack effect.
[0036] Specifically, crack feature enhancement can be achieved through the following technical process: First, a composite processing module (GammaAdjuster class) containing gamma transform and dynamic threshold enhancement is constructed to perform grayscale normalization processing on the input image; a dual-modal enhancement strategy is adopted. On the one hand, a binary mask is constructed by setting a low brightness threshold (black_threshold = 0.09), and exponential nonlinear compression is applied to the crack region with grayscale values below the threshold. The compression intensity is controlled by an enhancement factor, which significantly reduces the grayscale overlap between microcracks and the matrix background; on the other hand, global gamma correction and brightness gain adjustment are implemented, and the image grayscale histogram distribution is reshaped through a nonlinear mapping function; finally, pixel value truncation is used to maintain numerical stability, forming a synergistic enhancement mechanism that combines local enhancement of crack regions and optimization of overall image contrast.
[0037] The above-mentioned fracture feature enhancement method innovatively combines threshold-guided regional compression with global grayscale transformation. By dynamically configuring parameters to adapt to the grayscale characteristics of different coal and rock matrices, it effectively solves the problem of difficulty in distinguishing micro-cracks from noise in traditional methods, and improves the clarity of fracture edges by more than 42%.
[0038] S106: The processed sample set is input into the coal and rock fracture tracking model, which outputs the entire process of fracture development for each sample set. The fracture tracking model can be a convolutional neural network model based on a three-level memory architecture, comprising a key encoder, a value encoder, a mask decoder, and a convolutionally gated recurrent unit. The key encoder can be implemented based on ResNet-50, taking the processed sample set as input and generating multi-scale features. The value encoder can be implemented based on ResNet-18, and its core function is to transform the input mask and image features into discriminative object-level feature representations through multi-stage feature fusion and a gated memory update mechanism. The mask decoder, as the core prediction module of the coal and rock fracture tracking model, uses a multi-level feature fusion and progressive upsampling architecture to achieve high-precision mask generation. The function of the convolutionally gated recurrent unit is to realize and update the sensory memory, storing the hidden state in continuous time sequence, thereby providing temporal continuity. The three-level memory storage mechanism works in concert with each component, balancing short-term accuracy, long-term consistency, and memory efficiency, enabling accurate fracture segmentation and propagation.
[0039] This invention is not limited to the preferred embodiments described above. Anyone can derive other methods in various forms under the guidance of this invention. Any technical solution that is the same as or similar to this application falls within the protection scope of this invention.
Claims
1. A data preprocessing method for a crack tracking model based on 4DCT images, characterized in that, Includes the following steps: S101: Place the coal and rock sample between the rotating base and the rotatable pressure head; the coal and rock sample rotates at a set speed, and at the same time, uniaxial compression and 4DCT scanning are performed on the coal and rock sample to obtain a number of continuous scanning frames with fixed time intervals. Based on these scanning frames, several data volumes are reconstructed in three dimensions in chronological order. S102: Export CT image slices for each data volume at several selected horizontal levels; integrate CT image slices of the same level for all data volumes into a sample set according to the reconstruction order of the data volumes; S103: Use a circular mask to remove the black air image outside the boundary of the coal and rock sample for each CT image slice in the entire sample set, and retain only the valid data of the coal and rock sample. S104: For each sample set, the CT image slice of the first data volume is used as the keyframe, and the CT image slices of the remaining data volumes are aligned with the keyframe to achieve the angle alignment of the CT image slices. S105: Enhance the features of the cracks in each CT image slice; S106: Input the processed sample set into the coal and rock fracture tracking model, and the coal and rock fracture tracking model outputs the entire process of fracture development corresponding to each sample set.
2. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 1, characterized in that, In step S101, the coal and rock sample is a cylinder with a diameter of 50 mm and a height of 100 mm.
3. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 1, characterized in that, The rotation speed is 12° / s, and the scanning frequency is 20 frames / s.
4. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 1, characterized in that, In step S103, the CT image slices in the sample set are square, and the boundaries of the coal and rock samples are circular; the circular mask automatically detects the center and radius of the coal and rock samples in the CT image slices, and the parts outside the mask are processed as white.
5. The data preprocessing method for crack tracking model based on 4DCT images according to claim 4, characterized in that, In step S103, square-format CT image slices are first read, and grayscale threshold binarization and contour detection algorithms are used to automatically locate the minimum circumcircle of the circular boundary of the coal and rock sample, and accurately calculate its center coordinates and radius parameters. Then, a white background matrix of the same size as the original image is constructed, and the detected circular region pixels are accurately mapped to the new background through a circular mask generation function. Edge noise is automatically eliminated through a radius shrinkage coefficient to ensure that all black air image areas outside the boundary of the coal and rock sample are forced to be pure white.
6. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 1, characterized in that, In step S104, the method of aligning the cracks is to rotate the effective data portion in the CT image slice until it achieves maximum overlap with the crack features of the keyframe.
7. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 6, characterized in that, In step S104, firstly, the keyframe and the CT image slice to be aligned are subjected to gap enhancement processing. The contrast between the gap and the matrix background is enhanced by combining contrast-limited adaptive histogram equalization and morphological closing operation. Subsequently, the Fourier frequency domain angle detection algorithm is used to convert the frequency domain features into an angle feature space using linear polar coordinate transformation, and the initial angle deviation is obtained through the phase correlation method. Subsequently, a multi-stage angle optimization search strategy was adopted, which involves three stages of angle traversal: coarse-grained, medium-grained, and fine-grained. Combined with parallel computing acceleration technology, the weighted score of phase correlation and binarized crack overlap rate was used as a multi-modal evaluation index to dynamically determine the optimal rotation angle. Finally, an anti-aliased rotation algorithm was used to achieve image geometric transformation, and a reflection boundary filling mode was used to maintain image size consistency.
8. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 7, characterized in that, In step S105, the feature enhancement method is to determine the crack or background by using the grayscale threshold in the CT image slice, and to compress the threshold for low brightness.
9. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 8, characterized in that, In step S105, a composite processing module including gamma transform and dynamic threshold enhancement is first constructed to perform grayscale normalization processing on the input image. A dual-modal enhancement strategy is adopted. On the one hand, a binary mask is constructed by setting a low brightness threshold, and exponential nonlinear compression is applied to the crack region with grayscale value below the threshold. The compression intensity is controlled by an enhancement factor to significantly reduce the grayscale overlap between microcracks and matrix background. On the other hand, global gamma correction and brightness gain adjustment are implemented to reshape the image grayscale histogram distribution through a nonlinear mapping function. Finally, pixel value truncation is used to maintain numerical stability, forming a synergistic enhancement mechanism that combines local enhancement of crack region and optimization of overall image contrast.
10. The data preprocessing method for the crack tracking model based on 4DCT images according to claim 9, characterized in that, In step S106, the crack tracking model is a convolutional neural network model based on a three-level memory architecture; It consists of four parts: a key encoder, a value encoder, a mask decoder, and a convolutional gated recurrent unit. The key encoder is based on ResNet-50 and takes the processed sample set as input to generate multi-scale features. The value encoder is based on ResNet-18. Its core function is to transform the input mask and image features into discriminative object-level feature representations through multi-stage feature fusion and gated memory update mechanism. The mask decoder adopts a multi-stage feature fusion and progressive upsampling architecture to achieve high-precision mask generation. The function of the convolutional gated recurrent unit is to realize and update the sensory memory, store the hidden state in continuous time sequence, and thus provide temporal continuity.