Granular material degradation analysis method and system based on deep learning and micro-topology analysis

Abstract

The application discloses a kind of granular material degradation analysis method and system based on deep learning and micro-topology analysis, by obtaining the high-resolution micron CT scanning sequence image of granular material sample experienced environmental action;Multi-component voxel-level segmentation is carried out on CT image by using U-Net deep learning network with spatial attention mechanism introduced, and four types of microstructure element segmentation results are output;Three-dimensional quantitative parameters representing the degree of degradation are extracted from the segmentation results, including key cementing layer connectivity and inter-particle contact type proportion;Based on quantitative parameters and macroscopic mechanical test data, a multi-element mapping model from micro to macro mechanical properties is established using Gaussian process regression;The model is used to analyze and predict the strength of the granular material to be tested.The application realizes the accurate, automatic and three-dimensional quantitative evaluation of the micro-cementing structure degradation of granular material, and solves the problem that traditional methods cannot quantitatively describe the degradation of cementing network and the evolution of contact state.

Description

Technical Field

[0001] This application relates to the field of geotechnical engineering and mechanical property evaluation of granular materials, specifically to a method and system for analyzing the deterioration of granular materials based on deep learning and microscopic topological analysis. Background Technology

[0002] Granular materials, such as Quaternary residual soils, colluvial soils, and alluvial soils, as well as artificially prepared cemented granular materials, are extremely common geological carriers or engineering materials in highway, railway, water conservancy, and mine backfilling projects. A significant characteristic of these materials is the presence of cementing substances (such as free iron oxide, alumina, or artificially added cement and lime) between the particles. These cementing substances form a connection with a certain strength and stiffness between the particles, constituting a "cemented network," which is an important source of the macroscopic mechanical properties of granular materials (such as compressive strength and shear strength).

[0003] However, granular materials are extremely sensitive to environmental effects. Under long-term service conditions such as wet-dry cycles and freeze-thaw cycles, the cementing material between particles will dissolve, be lost, undergo phase transformation, or suffer fatigue damage, leading to the gradual degradation of the cementing network. This microscopic degradation manifests as a thinning of the cementing layer, reduced connectivity, and a degradation of the interparticle contact from a robust "cemented contact" to a weaker "surface contact" or even an extremely unstable "point contact," ultimately resulting in a significant decline in macroscopic mechanical properties, which in turn leads to engineering disasters such as roadbed settlement, slope instability, and infill failure. To assess the degradation state of granular materials, X-ray computed tomography (CT) technology, with its non-destructive and three-dimensional imaging advantages, has become an effective means of observing the internal microstructure of materials. However, existing CT image-based methods for analyzing granular materials have obvious limitations and blind spots, mainly in the following aspects:

[0004] (1) High reliance on subjective observation and lack of quantitative evaluation standards. At present, most engineering researchers are still at the stage of visual observation of CT slice images, and can only give qualitative descriptions such as "increased porosity" and "loosening of structure". This method, which relies on subjective experience, cannot accurately quantify the spatial distribution changes and degradation degree of the cementing layer, resulting in the evaluation results failing to provide objective and quantifiable reference for engineering safety.

[0005] (2) Traditional thresholding cannot handle complex multi-component structures. Existing technologies mostly use traditional algorithms such as bimodal thresholding for image binarization, which can only simply divide the image into solid and pore phases. Since the gray values ​​of cement (such as iron oxide), skeletal particles (such as quartz), clay matrix and pores in granular materials often overlap or have transition regions, traditional algorithms cannot achieve accurate segmentation of multi-components (four or more phases), resulting in distortion of the extracted parameters. For example, Chinese patent CN108256258B uses K-means clustering segmentation, which can only extract pore parameters and ignores the key cementing layer information; in the paper "Quantitative Characterization and Mechanical Property Prediction of Microstructure of Soil-Rock Mixture Based on CT Scan" published by Wang Yu et al. in Volume 40, Issue 3, 2021 of the Chinese Journal of Rock Mechanics and Engineering, only the two phases of boulders and pores are segmented by traditional thresholding methods, and the cementing material and clay matrix cannot be identified.

[0006] (3) The microscopic evaluation indicators are too simplistic to reflect the macroscopic mechanical essence. Existing studies, when attempting quantitative analysis, typically extract only single parameters such as global porosity or fractal dimension. For example, the Chinese patent application with publication number CN121545640A predicts strength solely based on fractal dimension. As a holistic statistical measure, fractal dimension completely ignores the topological characteristics of the cemented network that reflect the mechanical essence. Even if two samples have similar total porosity or fractal dimension, the connectivity of their internal cemented structures and the type of particle contact may be drastically different, leading to significant differences in macroscopic mechanical properties. Summary of the Invention

[0007] The main objective of this application is to provide a method for analyzing the degradation of granular materials based on deep learning and microscopic topological analysis, comprising the following steps:

[0008] Step S1: Obtain granular material samples that have undergone different environmental action stages, and perform computed tomography (CT) tests on them to obtain high-resolution micron-scale CT scan sequence images;

[0009] Step S2: Use a deep learning algorithm to perform multi-component pixel-level image segmentation on the high-resolution micron CT scan sequence image, and output a pixel-level segmentation result image containing four types of microstructure elements: oxide cement layer, skeleton particles, clay matrix and pores.

[0010] Step S3: Use image analysis algorithms to extract quantitative parameters characterizing the micro-deterioration degree of granular materials from the pixel-level segmentation result image. The quantitative parameters include at least the area ratio of oxide cementing layer, the average thickness of oxide cementing layer, the area ratio of pores, the equivalent diameter distribution of pores, the connectivity of cementing layer, and the proportion of inter-particle contact types.

[0011] Step S4: Obtain macroscopic mechanical test data of the same batch of granular material samples corresponding to the different environmental action stages; based on the extracted quantitative parameters and the macroscopic mechanical test data, establish a multivariate mapping model from microstructural parameters to macroscopic mechanical property degradation.

[0012] Step S5: Using the established multivariate mapping model, analyze the degree of degradation and predict the strength of the granular material under test.

[0013] In one embodiment, step S1 specifically includes:

[0014] The granular material samples that have undergone wet-dry or freeze-thaw cycles are subjected to liquid nitrogen quick-freezing and vacuum freeze-drying to preserve their original microstructure.

[0015] The dried sample was vacuum impregnated with low-viscosity epoxy resin. After the resin cured, the embedded sample was scanned with high-resolution micron CT using X-ray computed tomography equipment to obtain a sequence of images containing information on the internal three-dimensional microstructure.

[0016] The reconstructed CT sequence images were preprocessed sequentially with ring artifact correction, beam hardening correction, and three-dimensional nonlocal mean filtering for noise reduction.

[0017] In one embodiment, step S2 specifically includes:

[0018] The acquired high-resolution micron-sized CT scan sequence images were subjected to adaptive contrast enhancement and resampling normalization preprocessing.

[0019] The preprocessed image is input into a pre-trained deep learning semantic segmentation network. The network adopts the U-Net architecture with spatial attention mechanism and is optimized by a combination of multi-class cross-entropy loss and Dice loss to output a pixel-level segmentation result image with accurate boundaries for the four types of micro-structural elements.

[0020] In one embodiment, step S3 specifically includes:

[0021] Quantitative parameters of four basic morphology parameters were extracted from the pixel-level segmentation result image: the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the area ratio of the pores, and the distribution of the equivalent diameter of the pores.

[0022] Extract the total length of the contact surface of the skeleton particles in the pixel-level segmentation result image, and the length of the contact surface continuously covered by the oxide cementing layer. Calculate the ratio between the two to obtain the cementing layer connectivity, which characterizes the three-dimensional topological features of the cemented network.

[0023] Extract and calculate the number of point contacts, surface contacts, and cemented contacts, and calculate the percentage of each contact type to the total number of contacts to obtain the proportion of interparticle contact types that characterize the stability of the skeleton.

[0024] In one embodiment, the process of extracting the quantitative parameter specifically includes:

[0025] The connectivity of the cementing layer is obtained in the following way:

[0026] Extract the total length of all skeleton particle contact surfaces in the pixel-level segmentation result image, and the length of the contact surface continuously covered by the oxide cementing layer. Calculate the ratio of the covered contact surface length to the total length, and use this as the cementing layer connectivity rate to characterize the three-dimensional topological features of the cemented network.

[0027] The ratio of interparticle contact types is obtained in the following way:

[0028] A morphological dilation operation with a specified radius is performed on the discrete skeleton particle region in the pixel-level segmentation result image to obtain the dilated overlapping feature region; based on the geometric features of the dilated overlapping feature region and its coverage by the oxide cementing layer, the contact type between the skeleton particles is determined to be point contact, surface contact, or cementing contact; the number of point contact, surface contact, and cementing contact are counted respectively, and the percentage of each contact type in the total number of contacts is calculated.

[0029] The average thickness of the oxide cement layer is obtained by the following method:

[0030] Extract the central skeleton line of the oxide cement layer region and perform a three-dimensional Euclidean distance transformation on the region to obtain the distance transformation matrix; take the corresponding value of each voxel point on the central skeleton line in the distance transformation matrix as the half thickness at that point, and calculate twice the mean of all half thicknesses as the average thickness of the oxide cement layer.

[0031] The equivalent pore diameter distribution is obtained through the following method:

[0032] For the pore region in the pixel-level segmentation result image, local extrema are found based on three-dimensional distance transformation as seed points and a three-dimensional watershed segmentation algorithm is run to segment the interconnected pore network into independent pore unit connected domains; the volume of each connected domain is counted and the equivalent diameter of each connected domain is calculated using the sphere volume inverse formula, thereby obtaining the pore equivalent diameter distribution.

[0033] In one embodiment, the extraction process of the proportion of interparticle contact types further includes:

[0034] A morphological dilation operation with a specified radius is performed on the discrete skeleton particle region in the pixel-level segmentation result image to obtain the dilated overlapping feature region.

[0035] If the overlapping feature area is extremely small and not covered by any adhesive, it is determined to be point contact; if the overlapping feature area is distributed as a line segment and not covered by any adhesive, it is determined to be surface contact; if the overlapping feature area is mainly covered by oxide adhesive layer pixels, it is determined to be adhesive contact.

[0036] In one embodiment, step S4 specifically includes:

[0037] Macroscopic mechanical tests were conducted on the same batch of granular material samples corresponding to different environmental action stages to obtain macroscopic mechanical test data including unconfined compressive strength and compressive modulus.

[0038] The extracted quantitative parameters are used as the independent variable matrix, and the macroscopic mechanical test data are used as the dependent variable. The Gaussian process regression algorithm is used to eliminate multicollinearity and establish the multivariate mapping model.

[0039] In one embodiment, step S5 specifically includes:

[0040] Acquire CT scan images of the granular material to be tested, and extract quantitative parameters of the soil to be tested according to steps S2 to S3, including the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the pore area ratio, the pore equivalent diameter distribution, the cement layer connectivity rate, and the proportion of interparticle contact types.

[0041] The six quantitative parameters of the soil under test, namely the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the pore area ratio, the pore equivalent diameter distribution, the cement layer connectivity rate, and the proportion of interparticle contact types, are input into the multivariate mapping model to quickly invert and calculate and output the quantitative prediction result of the degree of mechanical property degradation of the soil under test.

[0042] A degradation analysis system for granular materials based on deep learning and microscopic topological analysis includes:

[0043] The image acquisition and preprocessing module is used to acquire high-resolution micron-scale CT scan sequence images of in-situ processed granular material samples and perform standardized preprocessing.

[0044] The image segmentation module is used to run the built-in deep learning semantic segmentation network to perform pixel-level classification of the image and output a pixel-level segmentation result image containing four types of micro-structure elements: oxide cement layer, skeleton particles, clay matrix and pores.

[0045] The morphological parameter extraction module is used to automatically extract multiple quantitative parameters from the pixel-level segmentation result image using morphological and topological algorithms. The quantitative parameters include at least the area ratio of the oxide cementing layer, the average thickness of the oxide cementing layer, the pore area ratio, the pore equivalent diameter distribution, the cementing layer connectivity, and the proportion of interparticle contact types.

[0046] The model building module is used to establish a multivariate mapping model between the quantitative parameters and the measured macroscopic mechanical test data during the model training phase. It also establishes a predictive model between the oxide cement layer area ratio, average thickness of oxide cement layer, pore area ratio, pore equivalent diameter distribution, cement layer connectivity, and particle contact type ratio of granular material samples and the unconfined compressive strength and compressive modulus of granular materials.

[0047] The evaluation output module is used to input the quantitative parameters of the granular material to be tested into the multivariate mapping model during the practical application stage, and output a degradation degree analysis and strength prediction report.

[0048] Therefore, this application has the following beneficial effects:

[0049] First, it breaks through the limitations of traditional single-indicator evaluation. Existing technologies generally use global parameters such as porosity or fractal dimension, neglecting the key impact of local cementation failure. This application, for the first time, incorporates the connectivity of the cemented layer and the ratio of inter-particle contact types as core topological parameters into the evaluation system. Even with similar total porosity, it can keenly identify the microscopic damage nature of particles degrading from cemented contact to unstable point contact, significantly improving the physical scientific nature and accuracy of mechanical prediction.

[0050] Second, it overcomes the shortcomings of traditional image segmentation methods. This application employs a 3D U-Net deep learning network with an introduced spatial attention mechanism, which successfully achieves accurate voxel-level segmentation of the four-phase structure of oxide cement layer, skeletal particles, clay matrix, and pores, completely eliminating the subjective blindness of traditional threshold segmentation and manual discrimination.

[0051] Third, it solves the problem of undisturbed sampling and significantly reduces the cost of exploration. This application transforms the traditional macroscopic mechanical test, which requires large-volume undisturbed samples and takes dozens of days, into an image calculation process based on CT scans of tiny samples. The evaluation cycle is shortened to a few hours, which greatly reduces the manpower and equipment costs of engineering exploration while ensuring the accuracy of the prediction.

[0052] Fourth, it boasts high prediction accuracy and robustness. The multivariate mapping model, based on a six-parameter system and Gaussian process regression, can control the intensity prediction error under both wet-dry and freeze-thaw cycles to within three percent, far superior to traditional single-parameter models. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a system flowchart of a method for analyzing the degradation of granular materials based on deep learning and microscopic topological analysis.

[0055] Figure 2 This is a structural diagram of the internal modules of the degradation analysis system of this application;

[0056] Figure 3 This is a correlation analysis diagram showing the changes in microscopic characteristic parameters and macroscopic unconfined compressive strength and compressive modulus at different degradation stages in the embodiments of this application;

[0057] Figure 4 These are CT scan images of granular materials with different numbers of dry-wet or freeze-thaw cycles in the embodiments of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0060] To address the shortcomings of existing technologies, this application provides a method for analyzing the degradation of granular materials based on deep learning and microscopic topological analysis, including steps S1 to S5, as described above. Figure 1 , Figure 1 This is a system flowchart of a method for analyzing the degradation of granular materials based on deep learning and microscopic topological analysis.

[0061] Step S1: Obtain granular material samples that have undergone different environmental action stages, and perform computed tomography (CT) tests on them to obtain high-resolution micron-scale CT scan sequence images;

[0062] Step S2: Use a deep learning algorithm to perform multi-component pixel-level image segmentation on the high-resolution micron CT scan sequence image, and output a pixel-level segmentation result image containing four types of microstructure elements: oxide cement layer, skeleton particles, clay matrix and pores.

[0063] Step S3: Use image analysis algorithms to extract quantitative parameters characterizing the micro-deterioration degree of granular materials from the pixel-level segmentation result image. The quantitative parameters include at least the area ratio of oxide cementing layer, the average thickness of oxide cementing layer, the area ratio of pores, the equivalent diameter distribution of pores, the connectivity of cementing layer, and the proportion of inter-particle contact types.

[0064] Step S4: Obtain macroscopic mechanical test data of the same batch of granular material samples corresponding to the different environmental action stages; based on the extracted quantitative parameters and the macroscopic mechanical test data, establish a multivariate mapping model from microstructural parameters to macroscopic mechanical property degradation.

[0065] Step S5: Using the established multivariate mapping model, analyze the degree of degradation and predict the strength of the granular material under test.

[0066] Specifically, this embodiment provides a method for analyzing the degradation of granular materials based on deep learning and microscopic topological analysis, referring to... Figure 1 The system flowchart shown specifically includes the following steps S1 to S5:

[0067] Step S1: Sample preparation and CT scan image acquisition

[0068] First, samples of iron oxide-rich granular material undergoing different stages of degradation were obtained. Taking wet-dry cycle degradation as an example, soil samples were prepared in the laboratory after undergoing 0, 5, 10, and 15 wet-dry cycles, respectively. Each wet-dry cycle included 24 hours of full water saturation at room temperature followed by 24 hours of drying in a 40°C oven. To avoid shrinkage and damage to the pore structure caused by direct drying, samples with a size of approximately 1 cm were prepared. 3 Small samples were rapidly frozen in liquid nitrogen, instantly transforming pore water into amorphous ice. They were then continuously dried in a vacuum freeze dryer at -50°C and a vacuum level below 10 Pa for 48 hours, completely preserving the original three-dimensional microporous morphology of the samples. After drying, the samples were impregnated and embedded in low-viscosity epoxy resin under a vacuum of -0.09 MPa. Several vacuum-to-normal-pressure cycles were performed to ensure complete resin filling of the micropores. After curing at room temperature for 24 hours, a hard embedded block was formed. The embedded samples were then placed in an X-ray micron-CT scanner with an accelerating voltage of 15 kV and a power of 7 W. At least 1000 projection images were acquired, and a three-dimensional grayscale image sequence with a voxel resolution better than 2 μm was reconstructed using a filtered back-projection reconstruction algorithm. The reconstructed images were then subjected to ring artifact correction, beam hardening correction, and three-dimensional nonlocal mean filtering for noise reduction, resulting in high-quality CT sequence images.

[0069] Step S2: Multi-component voxel-level intelligent segmentation

[0070] To overcome the limitation of traditional threshold segmentation methods in accurately distinguishing overlapping grayscale components, this embodiment employs a 3D U-Net deep learning network with a spatial attention mechanism to perform voxel-level semantic segmentation of CT sequence images. Before network training, geotechnical engineering experts manually annotated hundreds of CT slices at the voxel level using professional annotation software, representing four types of structures: oxide cementing layer, skeletal particles, clay matrix, and pores. During training, given that the proportion of voxels in the cementing layer is typically less than 15%, a combination of multi-class cross-entropy loss and Dice loss was used for optimization to effectively alleviate the class imbalance problem. The model converged after 200 training epochs on a workstation, achieving an average Dice coefficient of over 0.90 for the four types of structures on the validation set. After training, the network can automatically classify the input 3D CT images voxel by voxel, outputting voxel-level segmentation results images containing precise boundaries of the four types of microstructural elements.

[0071] Step S3: Extraction of microscopic topological parameters

[0072] Based on the segmentation results obtained in step S2, six quantitative parameters characterizing the microscopic degradation degree of granular materials are extracted using a three-dimensional image analysis algorithm. The first parameter is the volume ratio of the oxide cementing layer, i.e., the ratio of the number of voxels in the cementing layer to the total number of voxels, reflecting the loss of total cementing material. The second parameter is the average thickness of the oxide cementing layer, obtained by extracting the three-dimensional central skeleton surface of the cementing layer and multiplying the mean half-thickness at each skeleton point by two using Euclidean distance transformation, reflecting the thinning trend of the cementing layer. The third parameter is the pore volume ratio. The fourth parameter is the pore equivalent diameter distribution, obtained by segmenting the pore connected domain using a three-dimensional watershed algorithm, then using the volume of a sphere to infer the equivalent diameter of each pore and statistically analyzing its mean and coefficient of variation. The fifth parameter is the cementing layer connectivity rate, defined as the ratio of the total length of the particle contact surface continuously covered by the oxide cementing layer to the length of the contact surface between all skeleton particles. This parameter directly reflects the connectivity effectiveness of the cementing network at key positions in the force chain. The sixth item is the proportion of interparticle contact types. By performing a three-dimensional morphological expansion operation with a specified radius on the discrete skeleton particle region, the expanded overlapping region is obtained. Based on the geometric characteristics of the overlapping region and the proportion of cementing layer voxels in it, the three types of contact are automatically determined: point contact, surface contact, and cementing contact. The number of each type of contact is counted and the percentage of each type of contact in the total number of contacts is calculated, thereby quantitatively characterizing the evolution process of particle contact state from cementing contact to frictional contact.

[0073] Step S4: Macroscopic mechanical experiments and model establishment

[0074] Unconfined compressive strength tests and one-dimensional consolidated compression tests were conducted on standard cylindrical specimens prepared in the same batch as the CT scan specimens to obtain the peak compressive strength and compression modulus corresponding to each degradation stage as macroscopic mechanical test data. Using the six quantitative parameters extracted in step S3 as the independent variable matrix and the macroscopic mechanical data as the dependent variable, a multivariate mapping model was established using the Gaussian process regression algorithm. Gaussian process regression can effectively handle multicollinearity among parameters and provide the uncertainty range of predicted values, making it more suitable for modeling small-sample geotechnical engineering data.

[0075] Step S5: Deterioration Analysis and Intensity Prediction

[0076] In practical engineering applications, steps S1 to S3 are performed on the granular material sample to be tested to obtain its six quantitative parameters. These parameters are then input into the multivariate mapping model established in step S4. The model rapidly inverts and calculates, outputting the predicted values ​​of the unconfined compressive strength and compressive modulus of the material in its current deteriorated state, along with the percentage strength reduction relative to the undeteriorated state. This method achieves automated and quantitative cross-scale prediction from three-dimensional images of the microstructure to macroscopic mechanical properties, providing an efficient and reliable technical means for the engineering safety assessment of granular materials.

[0077] In one embodiment, step S1 specifically includes:

[0078] The granular material samples that have undergone wet-dry or freeze-thaw cycles are subjected to liquid nitrogen quick-freezing and vacuum freeze-drying to preserve their original microstructure.

[0079] The dried sample was vacuum impregnated with low-viscosity epoxy resin. After the resin cured, the embedded sample was scanned with high-resolution micron CT using X-ray computed tomography equipment to obtain a sequence of images containing information on the internal three-dimensional microstructure.

[0080] The reconstructed CT sequence images were preprocessed sequentially with ring artifact correction, beam hardening correction, and three-dimensional nonlocal mean filtering for noise reduction.

[0081] Specifically, in this embodiment, the granular material sample size after undergoing environmental effects such as wet-dry cycles or freeze-thaw cycles is approximately 1 cm. 3 The samples were rapidly immersed in liquid nitrogen for quick-freezing, instantly transforming pore water into amorphous ice to prevent crystal expansion from damaging the original microstructure. Subsequently, the quick-frozen samples were placed in a vacuum freeze dryer and continuously dried for 48 hours at -50°C and a vacuum degree below 10 Pa, allowing the ice to sublimate directly and completely preserving the initial three-dimensional pore morphology and particle arrangement characteristics of the samples.

[0082] Next, the dried sample was placed in a vacuum impregnation device, and low-viscosity epoxy resin was injected under a vacuum of -0.09 MPa. Multiple vacuum-to-normal-pressure cycles were performed to ensure the resin fully penetrated and filled every tiny pore inside the sample. After the resin was allowed to cure at room temperature for 24 hours, a hard embedded mass was formed. Then, the cured embedded sample was fixed on the sample stage of an X-ray micro-CT scanning device. The X-ray source acceleration voltage was set to 10-100 kV and the power to 1-10 W, and 500-3200 projection images were acquired. A three-dimensional grayscale image sequence with a voxel resolution of 1-5 μm was reconstructed using a filtered back-projection reconstruction algorithm to obtain complete three-dimensional microstructural information of the sample.

[0083] Finally, the reconstructed CT sequence images were preprocessed sequentially. Frequency domain filtering was used to correct ring artifacts, polynomial fitting was used to correct beam hardening, and three-dimensional nonlocal mean filtering was applied to suppress noise, thereby obtaining high-quality, high signal-to-noise ratio CT scan sequence images, laying the foundation for subsequent multi-component segmentation and parameter extraction.

[0084] In one embodiment, step S2 specifically includes:

[0085] The acquired high-resolution micron-sized CT scan sequence images were subjected to adaptive contrast enhancement and resampling normalization preprocessing.

[0086] The preprocessed image is input into a pre-trained deep learning semantic segmentation network. The network adopts the U-Net architecture with spatial attention mechanism and is optimized by a combination of multi-class cross-entropy loss and Dice loss to output a pixel-level segmentation result image with accurate boundaries for the four types of micro-structural elements.

[0087] Specifically, in this embodiment, the high-resolution micron-sized CT scan sequence images obtained in step S1 are first preprocessed and enhanced. Since inconsistencies in voxel resolution may exist during CT scanning, all sequence images need to be resampled to a uniform voxel resolution based on the scale information in the image metadata to ensure comparability between different batches of images. Subsequently, a three-dimensional confined contrast adaptive histogram equalization algorithm is applied to adjust the grayscale distribution in local image regions, significantly enhancing the detail clarity and contrast of dark pore boundaries and bright iron oxide cement layer structures.

[0088] Secondly, the preprocessed 3D CT image data is input into a pre-trained deep learning semantic segmentation network. This network adopts a 3D U-Net architecture with a spatial attention gating module. Its encoder extracts multi-scale features step by step through multi-layer 3D convolution and pooling operations, and the decoder restores spatial resolution through deconvolution and skip connections. The spatial attention module enables the network to automatically focus on the sparsely distributed but mechanically important oxide cementation layer region during training and inference.

[0089] During the network training phase, voxel-level ground truth values ​​of four types of structural elements—oxide cement layer, skeletal particles, clay matrix, and pores—annotated by geotechnical engineering experts, were used as supervision signals. Given that cement layer voxels typically account for less than 15% of the total voxels, resulting in a severe class imbalance, the model training employed a combined loss function of multi-class cross-entropy loss and Dice loss. By adjusting the weight coefficients of these two losses (e.g., cross-entropy weight 0.3, Dice weight 0.7), the model focused more on the segmentation accuracy of a minority of classes during optimization. After sufficient training and convergence, the model can perform voxel-by-voxel classification on input CT images, outputting voxel-level segmentation results containing precise boundaries of the four microstructural elements. The combined loss function is the core loss function used to train the deep learning semantic segmentation network, specifically expressed as follows:

[0090]

[0091] Where α and β are the weights balancing the two loss terms, used to automatically adjust the attention given to pixel predictions for the four components; the multi-class cross-entropy loss is:

[0092] ,

[0093] Among them, y c P represents the vector labels for the oxide cementing layer, quartz framework particles, clay matrix, and pores. c The model predicts the probability that an image pixel belongs to one of the four corresponding labels; the Dice loss is:

[0094]

[0095] Where X represents the total number of pixels predicted by the model to belong to a certain label, and Y represents the total number of pixels manually labeled to belong to a certain label. Dice loss measures the degree of overlap between the predicted mask and the ground truth mask. The closer the Dice coefficient is to 1, the better the overlap. Dice loss is insensitive to the number of foreground pixels, effectively preventing the model from ignoring small foreground targets due to a large background proportion.

[0096] In one embodiment, step S3 specifically includes:

[0097] Quantitative parameters of four basic morphology parameters were extracted from the pixel-level segmentation result image: the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the area ratio of the pores, and the distribution of the equivalent diameter of the pores.

[0098] Extract the total length of the contact surface of the skeleton particles in the pixel-level segmentation result image, and the length of the contact surface continuously covered by the oxide cementing layer. Calculate the ratio between the two to obtain the cementing layer connectivity, which characterizes the three-dimensional topological features of the cemented network.

[0099] Extract and calculate the number of point contacts, surface contacts, and cemented contacts, and calculate the percentage of each contact type to the total number of contacts to obtain the proportion of interparticle contact types that characterize the stability of the skeleton.

[0100] Specifically, in this embodiment, four basic morphological quantitative parameters are first extracted from the voxel-level segmentation result image obtained in step S2. The first parameter is the volume percentage A of the oxide cement layer. C The number of voxels belonging to the cementing layer category is counted by traversing the three-dimensional label field and divided by the total number of voxels in the effective volume of the sample, reflecting the total loss rate of cementing material under environmental influence.

[0101] The second item is the average thickness T of the oxide cement layer. C The central skeleton surface of the cementing layer region is extracted using a three-dimensional skeletonization algorithm. At the same time, a three-dimensional Euclidean distance transformation is performed on the cementing layer region. The corresponding distance value of each voxel point on the skeleton surface in the distance transformation matrix is ​​taken as the half thickness at that point. Twice the mean of all half thicknesses is calculated as the average thickness of the cementing layer.

[0102] The third item is the pore volume ratio A. p The fourth term represents the ratio of the number of pore classification label voxels to the total number of effective voxels. eq Based on the three-dimensional distance transformation, local extreme points are found as seed points and a three-dimensional watershed segmentation algorithm is run to segment the interconnected pore network into independent connected domains. After the volume of each connected domain is counted, the equivalent diameter is calculated using the sphere volume inverse formula, thereby obtaining the statistical distribution of the pore equivalent diameter.

[0103] Secondly, the connectivity C of the cemented layer was extracted. r This is a core topological parameter. Extracting this parameter is an innovative method for evaluating the integrity of soil cementation. First, the boundary edges of all particles are extracted, and the total length of the contact surface of the outer perimeter of all particles within the image is calculated and denoted as L. total Then, the total length of the boundary that intersects with both the particle edge and the oxide cement layer is extracted and set as L. cemented ,calculate A decrease in cemented connectivity indicates a break in the load-bearing force chain network. This parameter directly characterizes the spatial connectivity effectiveness of the cemented network at critical locations in the force chain.

[0104] Finally, the proportion R of interparticle contact types was extracted. cem A three-dimensional morphological dilation operation with a specified radius is performed on the discrete skeleton particle region. Based on the geometric characteristics of the dilated overlapping region and the proportion of voxels covered by the cementing layer, three types of contact are automatically determined: point contact, surface contact, and cementing contact. The number of each type of contact is counted and the percentage of each type of contact in the total number of contacts is calculated, thereby quantitatively characterizing the degradation process of particle contact from cementing contact to frictional contact.

[0105] In one embodiment, the process of extracting the quantitative parameter specifically includes:

[0106] The connectivity of the cementing layer is obtained in the following way:

[0107] Extract the total length of all skeleton particle contact surfaces in the pixel-level segmentation result image, and the length of the contact surface continuously covered by the oxide cementing layer. Calculate the ratio of the covered contact surface length to the total length, and use this as the cementing layer connectivity rate to characterize the three-dimensional topological features of the cemented network.

[0108] The ratio of interparticle contact types is obtained in the following way:

[0109] A morphological dilation operation with a specified radius is performed on the discrete skeleton particle region in the pixel-level segmentation result image to obtain the dilated overlapping feature region; based on the geometric features of the dilated overlapping feature region and its coverage by the oxide cementing layer, the contact type between the skeleton particles is determined to be point contact, surface contact, or cementing contact; the number of point contact, surface contact, and cementing contact are counted respectively, and the percentage of each contact type in the total number of contacts is calculated.

[0110] The average thickness of the oxide cement layer is obtained by the following method:

[0111] Extract the central skeleton line of the oxide cement layer region and perform a three-dimensional Euclidean distance transformation on the region to obtain the distance transformation matrix; take the corresponding value of each voxel point on the central skeleton line in the distance transformation matrix as the half thickness at that point, and calculate twice the mean of all half thicknesses as the average thickness of the oxide cement layer.

[0112] The equivalent pore diameter distribution is obtained through the following method:

[0113] For the pore region in the pixel-level segmentation result image, local extrema are found based on three-dimensional distance transformation as seed points and a three-dimensional watershed segmentation algorithm is run to segment the interconnected pore network into independent pore unit connected domains; the volume of each connected domain is counted and the equivalent diameter of each connected domain is calculated using the sphere volume inverse formula, thereby obtaining the pore equivalent diameter distribution.

[0114] Specifically, in this embodiment, the process of extracting the quantitative parameters includes:

[0115] (a) Extraction of the connectivity of the cementing layer

[0116] Cemented layer connectivity is a core parameter characterizing the three-dimensional topological connectivity of the cemented network within granular materials. Its extraction process is as follows: First, based on the voxel-level segmentation image obtained in step S2, a three-dimensional traveling cube algorithm is used to extract the triangular mesh model of the outer surface of all skeletal particles. The total surface length of the contact surfaces between all skeletal particles is calculated using the mesh model. Second, the mesh patches corresponding to all contact surfaces are traversed to determine whether their corresponding positions in the original three-dimensional label field are continuously covered by oxide cemented layer labels. To avoid noise interference, a minimum continuous coverage length threshold of 25 voxels is set. The lengths of patches that meet the continuous coverage condition are summed to obtain the length of the contact surface effectively covered by the oxide cemented layer. Finally, the cemented layer connectivity C is calculated. This ratio directly reflects the spatial connectivity effectiveness of the cemented network at key positions of the force chains between particles; its decrease indicates fracture degradation of the load-bearing structure.

[0117] (ii) Extraction of the proportion of interparticle contact types

[0118] The proportion of contact types between particles was automatically determined and statistically analyzed using the following three-dimensional morphological method. First, connected component analysis was performed on the discrete skeleton particle regions in the segmentation result image, assigning a unique identifier to each independent particle. Then, a three-dimensional spherical structural element morphological dilation operation with a radius of three voxels was performed on the mask of each particle. When the dilated regions of two particles overlapped in three-dimensional space, geometric contact was determined to exist between the two particles. Further analysis of the geometric characteristics of the dilated overlapping region and its composition in the original label field was conducted: if the volume of the overlapping region was less than twenty-seven voxels and the proportion of cementing layer voxels was less than 10%, it was determined to be point contact; if the volume of the overlapping region was greater than or equal to twenty-seven voxels, the aspect ratio of the smallest circumscribed cuboid was greater than three, and the proportion of cementing layer voxels was less than 10%, it was determined to be surface contact; if the proportion of cementing layer voxels in the overlapping region was greater than or equal to fifty percent, it was determined to be robust cemented contact. The number of each type of contact was counted separately, and the total number of contacts was the sum of the three. The percentage of each contact type in the total number of contacts was calculated, with the proportion of cemented contact being the core indicator for evaluating skeleton stability.

[0119] (III) Extraction of the average thickness of the oxide cement layer

[0120] The average thickness of the oxide cementing layer was calculated using a combination of 3D skeletonization and distance transformation. First, a 3D skeletonization algorithm was used to extract the central skeleton surface of the oxide cementing layer region. This skeleton surface consists of a series of interconnected 3D voxel points, accurately reflecting the spatial orientation of the irregular cementing layer. Simultaneously, a 3D Euclidean distance transformation was performed on the cementing layer region, generating a distance transformation matrix, where each voxel value represents the shortest distance from that point to the nearest non-cementing layer background boundary. The spatial coordinates of each voxel point on the skeleton surface were superimposed with the distance transformation matrix, and the distance value at each skeleton point was read as the half-thickness of the cementing layer at that point. The arithmetic mean of the half-thicknesses of all skeleton points was calculated and multiplied by two to obtain the mathematical average thickness of the irregularly distributed cementing layer in the entire 3D space. This parameter accurately reflects the trend of gradual thinning of the cementing layer during environmental degradation.

[0121] (iv) Extraction of equivalent pore diameter distribution

[0122] For the complex and interconnected pore network in the segmentation result image, a three-dimensional watershed segmentation algorithm is used for individualization. First, a three-dimensional distance transformation is performed on the pore region to obtain the distance value from each pore voxel to the nearest solid boundary. Local maxima in the distance transformation field are found as seed points for watershed segmentation, with each seed point representing the center position of a potential pore cell. Then, the three-dimensional watershed algorithm is run, growing outward from the seed point until the boundaries of different pore cells meet, thus segmenting the connected pore network into several independent pore cell connected regions. The volume of each connected region is obtained by counting the number of voxels contained in it, and the equivalent diameter of each connected region is calculated using the inverse formula of sphere volume. Finally, the equivalent diameters of all pore cells are counted, a distribution histogram is plotted, and the mean and coefficient of variation are calculated as quantitative indicators characterizing the evolution of the pore structure.

[0123] In one embodiment, the extraction process of the proportion of interparticle contact types further includes:

[0124] A morphological dilation operation with a specified radius is performed on the discrete skeleton particle region in the pixel-level segmentation result image to obtain the dilated overlapping feature region.

[0125] If the overlapping feature area is extremely small and not covered by any adhesive, it is determined to be point contact; if the overlapping feature area is distributed as a line segment and not covered by any adhesive, it is determined to be surface contact; if the overlapping feature area is mainly covered by oxide adhesive layer pixels, it is determined to be adhesive contact.

[0126] Specifically, in this embodiment, based on the voxel-level segmentation result image obtained in step S2, three-dimensional connected component analysis is performed on the discretely distributed skeleton particle regions, and a unique identifier is assigned to each independent particle. Subsequently, a three-dimensional morphological dilation operation with a specified radius is performed on the mask region of each skeleton particle. In this embodiment, the dilation operation uses a spherical structural element with a radius of three voxels. This radius value is determined comprehensively based on the voxel resolution of the CT image and the typical particle size range of granular materials, which can effectively identify the real contact relationship between particles. After the dilation operation is completed, it is detected whether the dilated regions of different particles overlap in three-dimensional space. If the dilated regions of particle A and particle B intersect, it is determined that there is geometric contact between the pair of particles, and the intersection region is the dilation overlap feature region. Next, geometric morphological analysis and component determination are performed on each dilation overlap feature region:

[0127] Firstly, if the volume of the overlapping feature region is less than twenty-seven voxels, and the proportion of voxels identified as oxide cementing layers in the original three-dimensional tag field corresponding to the overlapping region is less than ten percent, then the contact is classified as a point contact. Mechanically, point contact is characterized by a tiny contact area and highly unstable force transmission.

[0128] Secondly, if the volume of the expansion overlap region is greater than or equal to twenty-seven voxels, the aspect ratio of its smallest circumscribed cuboid is greater than three, it exhibits a clear linear distribution, and the proportion of oxide cement layer voxels within the overlap region is still less than ten percent, then the contact is determined to be a surface contact. Surface contact represents direct frictional contact between particles over a certain area.

[0129] Third, if the proportion of voxels identified as oxide cementing layers in the original three-dimensional tag field corresponding to the expansion and overlap feature region reaches or exceeds 50%, then regardless of the geometry of the overlapping region, the contact is classified as cemented contact. Cemented contact indicates that the particles are strongly bonded together by cementing substances such as iron oxides, which is the main source of the macroscopic strength of granular materials.

[0130] In one embodiment, step S4 specifically includes:

[0131] Macroscopic mechanical tests were conducted on the same batch of granular material samples corresponding to different environmental action stages to obtain macroscopic mechanical test data including unconfined compressive strength and compressive modulus.

[0132] The extracted quantitative parameters are used as the independent variable matrix, and the macroscopic mechanical test data are used as the dependent variable. The Gaussian process regression algorithm is used to eliminate multicollinearity and establish the multivariate mapping model.

[0133] Specifically, in this embodiment, a macroscopic mechanical test was first performed on a standard granular material sample prepared in the same batch as the CT scan sample and subjected to the same environmental action phase. A microcomputer-controlled electronic universal testing machine was used to perform the unconfined compressive strength test according to the "Standard for Geotechnical Testing Methods". The loading rate was controlled at one percent of the axial strain per minute. The stress-strain curve was recorded, and the peak stress was taken as the unconfined compressive strength σ. c Simultaneously, one-dimensional consolidated compression tests were used to obtain the compression deformation of the specimens under various load levels, and the soil element compression modulus Es was calculated. At least three sets of parallel tests were conducted on specimens for each type of degradation stage, and the average value was taken as the representative macroscopic mechanical value for that stage.

[0134] Secondly, the six quantitative parameters extracted in step S3—oxide cement layer volume ratio, average thickness, pore volume ratio, mean distribution of pore equivalent diameter, cement layer connectivity, and interparticle cementation contact ratio—form a six-dimensional independent variable matrix. The corresponding macroscopic mechanical test data—unconfined compressive strength and compressive modulus—are used as dependent variables to establish a training dataset. Given the potential for multicollinearity among the various microscopic parameters, such as the correlation between cement layer volume ratio and connectivity, traditional least squares methods are prone to overfitting. Therefore, this embodiment employs a Gaussian process regression algorithm to construct a multivariate mapping model. A radial basis function is selected as the kernel function, and the kernel function hyperparameters are automatically optimized by maximizing the logarithmic marginal likelihood function. After model training, the weighted ranking of the influence of each microscopic parameter on macroscopic mechanical properties can be quantitatively output, along with predicted values ​​and confidence intervals, providing a reliable basis for subsequent degradation assessment and strength prediction.

[0135] It is particularly important to note that after extracting a large number of microscopic parameters from samples at each stage of degradation, this embodiment uses a computer-controlled electronic universal testing machine to perform unconfined compressive strength tests and one-dimensional consolidated compression tests on standard cylindrical soil samples from the same batch as the electron microscopy samples. The peak compressive strength σ of the samples at each stage is obtained. c and compressive modulus E S At this point, the model training set has a set consisting of (A) C T C A p D eq C r R cem The six-dimensional set of microscopic independent variables, and the set of mechanical test data (σ) c E S ).

[0136] Considering the potential for multicollinearity among the quantitative parameters—for example, thinner sample thickness often leads to a smaller area, causing overfitting with conventional least squares fitting—the above system employs a Gaussian process machine learning algorithm to establish a quantitative mapping model, thereby determining the fitting σ for each microscopic parameter variable. c The linear weights ω of Es i and ω 、 i To construct unconfined compressive strength The multivariate prediction equation for the soil element compression modulus Es is expressed mathematically as follows:

[0137]

[0138]

[0139] The regression coefficients determined through model training clarify the contribution of each micro-parameter to the macro-intensity. After the model is built, it is saved as a calculation module file that can be used for prediction.

[0140] In one embodiment, step S5 specifically includes:

[0141] Acquire CT scan images of the granular material to be tested, and extract quantitative parameters of the soil to be tested according to steps S2 to S3, including the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the pore area ratio, the pore equivalent diameter distribution, the cement layer connectivity rate, and the proportion of interparticle contact types.

[0142] The six quantitative parameters of the soil under test, namely the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the pore area ratio, the pore equivalent diameter distribution, the cement layer connectivity rate, and the proportion of interparticle contact types, are input into the multivariate mapping model to quickly invert and calculate and output the quantitative prediction result of the degree of mechanical property degradation of the soil under test.

[0143] Specifically, in this embodiment, a sample of the granular material to be tested is first obtained. In practical engineering applications, since it is difficult to obtain completely undisturbed, large-volume, uncirculated samples from granular material formations using conventional sampling methods, the testing personnel only need to scrape a sample of approximately 1 cm³ from borehole core samples or slope profiles. 3 Small, undisturbed soil clumps are sufficient for testing. The sample is subjected to liquid nitrogen flash freezing, vacuum freeze-drying, epoxy resin embedding, and high-resolution micron-scale CT scanning as described in step S1 to obtain three-dimensional CT scan sequence images of the material under test.

[0144] Next, the acquired CT scan images are sequentially input into the analysis system of this application, automatically executing the deep learning multi-component segmentation process in step S2 and the microscopic topological parameter extraction process in step S3. Specifically, the system calls the pre-trained 3DU-Net semantic segmentation network to perform voxel-level classification on the CT images, outputting segmentation results containing four types of components: oxide cement layer, skeletal particles, clay matrix, and pores. Subsequently, the morphological parameter extraction module automatically calculates six quantitative parameters from the segmentation results: oxide cement layer volume ratio, average oxide cement layer thickness, pore volume ratio, mean distribution of pore equivalent diameter, cement layer connectivity, and interparticle cementation contact ratio.

[0145] Finally, the six quantitative parameters mentioned above are used as input feature vectors and loaded into the Gaussian process regression multivariate mapping model trained in step S4. The model calculates the similarity between the test sample and the training samples based on a kernel function, quickly inverts and calculates, and outputs the predicted values ​​of the unconfined compressive strength and compressive modulus of the tested granular material in the current deterioration state. It also provides the percentage strength degradation relative to the undeteriorated baseline state and the prediction confidence interval. The entire process does not require time-consuming and labor-intensive macroscopic mechanical destructive testing.

[0146] A degradation analysis system for granular materials based on deep learning and microscopic topological analysis includes:

[0147] The image acquisition and preprocessing module is used to acquire high-resolution micron-scale CT scan sequence images of in-situ processed granular material samples and perform standardized preprocessing.

[0148] The image segmentation module is used to run the built-in deep learning semantic segmentation network to perform pixel-level classification of the image and output a pixel-level segmentation result image containing four types of micro-structure elements: oxide cement layer, skeleton particles, clay matrix and pores.

[0149] The morphological parameter extraction module is used to automatically extract multiple quantitative parameters from the pixel-level segmentation result image using morphological and topological algorithms. The quantitative parameters include at least the area ratio of the oxide cementing layer, the average thickness of the oxide cementing layer, the pore area ratio, the pore equivalent diameter distribution, the cementing layer connectivity, and the proportion of interparticle contact types.

[0150] The model building module is used to establish a multivariate mapping model between the quantitative parameters and the measured macroscopic mechanical test data during the model training phase. It also establishes a predictive model between the oxide cement layer area ratio, average thickness of oxide cement layer, pore area ratio, pore equivalent diameter distribution, cement layer connectivity, and particle contact type ratio of granular material samples and the unconfined compressive strength and compressive modulus of granular materials.

[0151] The evaluation output module is used to input the quantitative parameters of the granular material to be tested into the multivariate mapping model during the practical application stage, and output a degradation degree analysis and strength prediction report.

[0152] Specifically, this embodiment provides a granular material degradation analysis system based on deep learning and microscopic topology analysis. This system can be installed within a computer equipped with a high-performance GPU and consists of a front-end interactive interface and a back-end computing engine. (Refer to...) Figure 2 The system module structure diagram shown indicates that the system specifically includes the following five functional modules.

[0153] (a) Image acquisition and preprocessing module

[0154] This module serves as the system's data input gateway, supporting batch import of image sequences in DICOM, TIFF, or RAW formats from X-ray micro-CT equipment reconstruction. The module incorporates image metadata parsing, automatically reading voxel size information and resampling it to a uniform resolution. Simultaneously, the module sequentially performs standardized preprocessing operations such as annular artifact correction, beam hardening correction, and 3D nonlocal mean filtering for noise reduction, effectively improving the image signal-to-noise ratio and providing high-quality 3D CT image data for subsequent accurate segmentation.

[0155] (ii) Image segmentation module

[0156] This module integrates an inference engine based on the TensorFlow or PyTorch deep learning framework and pre-trained network weight files. It employs a 3D U-Net semantic segmentation network with spatial attention gating, automatically loading the model and inputting pre-processed 3D CT image data in blocks into GPU memory for accelerated inference. The network classifies each voxel, outputting voxel-level segmentation results images containing four microstructural elements: oxide cementing layer, skeletal particles, clay matrix, and pores, visualized in a color-labeled field.

[0157] (III) Morphological Parameter Extraction Module

[0158] This module consists of a series of high-performance algorithm scripts written using OpenCV, Scikit-image, and ITK computing libraries. It embeds a 3D skeletonization algorithm, a 3D Euclidean distance transformation algorithm, a 3D watershed segmentation algorithm, and a highly complex topology analysis engine. The module automatically extracts six quantitative parameters from the segmentation results: oxide cementing layer volume ratio, average thickness, pore volume ratio, pore equivalent diameter distribution, cementing layer connectivity, and inter-particle contact type ratio. The cementing layer connectivity is obtained by calculating the ratio of the total length of the particle contact surface to the cementing layer coverage length. The contact type ratio is obtained by automatically determining the proportions of point contact, surface contact, and cemented contact using a 3D morphological dilation operation.

[0159] (iv) Model Building Module

[0160] This module is the system's advanced management module, used to establish the mapping relationship between microscopic parameters and macroscopic mechanical properties during the model training phase. The module incorporates a Gaussian process regression algorithm, using six extracted quantitative parameters as independent variables and measured unconfined compressive strength and compressive modulus as dependent variables. It automatically optimizes the kernel function hyperparameters and establishes a multivariate mapping prediction model. The module supports importing new experimental data to update and refresh the model, ensuring the system's universality and prediction accuracy in engineering projects across different regions.

[0161] (v) Evaluation Output Module

[0162] This module provides an interactive interface for end users. In practical applications, the module receives six quantitative parameters of the granular material to be tested, calls the pre-trained prediction model, and calculates and displays the estimated unconfined compressive strength and compressive modulus values ​​in real time. Simultaneously, the module incorporates safety judgment logic; when the predicted value falls below the engineering safety threshold, it automatically outputs a red alarm and generates a PDF-format "Microscopic Deterioration Assessment Report of Granular Materials," including image segmentation comparison diagrams, parameter evolution curves, and mechanical evaluation conclusions, providing an intuitive basis for engineering safety decisions.

[0163] This application provides verification test data from a slope engineering project using iron oxide-rich granite residual soil in a certain region as an application illustration. In this engineering verification, the research team obtained undisturbed samples from the site and subjected them to 0, 5, 10, and 15 cycles of wet-dry degradation treatment under artificially controlled laboratory conditions. Subsequently, macroscopic mechanical specimens were simultaneously prepared for unconfined compressive strength tests, and micro-samples were analyzed and processed using this system. The evolution of the mean values ​​of the parameters and the macroscopic mechanical data automatically extracted by the system in this application are shown in the table.

[0164] Table 1. Microscopic parameters extracted from CT scan images and measured macroscopic peak compressive strength

[0165]

[0166] Table 2 Model Data Error Analysis Table

[0167]

[0168] The detailed parameters in the two tables above provide a profound insight into the evolutionary logic of the degradation mechanism:

[0169] Initially (0 wet-dry cycles), goethite and amorphous iron oxides formed a good cement between soil particles, at which point the cement thickness T was significant. C Large, its connectivity C rUp to 82.3%, and the particle skeleton is mainly connected by strong adhesive bonding (R cem The macroscopic unconfined compressive strength is at its highest peak (68.5%).

[0170] As environmental water erosion intensifies to 15 cycles, the iron oxide cementing material undergoes phase transformation or dissolution, resulting in a cemented layer area A. C and thickness T C The data shows a precipitous drop. The dissolution caused by the wet-dry cycle leads to the collapse of the contact bridges between particles, resulting in a decrease in the cementation connectivity C. r The proportion of cemented contacts dropped to 35.8%, indicating a significant degradation of the previously strong bonded contact. Under geometric morphological recognition, this degenerated into highly unstable point and surface contacts, with the proportion of cemented contacts falling to a mere 18.2%. This resulted in the expansion and interconnection of micropores, leading to an increase in the equivalent diameter D. eq The load-bearing capacity of the skeleton increases exponentially. The loss of load-bearing capacity is macroscopically manifested as compressive strength σ. c The drastic decay, with a decrease of nearly 65%, fully verifies that the six-dimensional microscopic index selected in this method is highly consistent with the actual mechanical decay logic in terms of physical mechanism.

[0171] After importing the first three sets of data into the multiple regression module to establish the equation, we used the six microscopic parameters obtained from 15 iterations of pure image analysis to calculate the predicted compressive strength, which was 163.5 kPa. Comparing this to the measured mechanical strength of 168.2 kPa in the same batch in Table 2, the system's prediction error was only about 2.8%. This fully demonstrates the reliability of the method in this application in terms of prediction accuracy, enabling high-precision and rapid prediction of the mechanical properties of severely degraded strata based solely on scanning electron microscope images without the need for destructive macroscopic experiments.

[0172] Figure 3 The study demonstrates the relationship between six microscopic characteristic parameters and macroscopic unconfined compressive strength at different stages of degradation. The correlation analysis results of the compressive modulus Es were also presented. Combined with the data from the examples, the absolute values ​​of the correlation coefficients between each microscopic parameter and the mechanical index were all above 0.5, with the cemented contact ratio R... cem and The correlation with [a specific particle type] reached a high of 0.91, and with Es reached 0.84, indicating that the evolution of interparticle contact type is the core controlling factor for mechanical property degradation. The proportion of cemented layer area A C and the equivalent diameter of the pores D eq The correlation is also strong, further confirming the direct impact of cementing material loss and pore expansion on structural weakening. This correlation analysis verifies that the selected six microscopic parameters are highly consistent with the macroscopic mechanical attenuation law in terms of physical mechanism, laying a solid statistical foundation for the establishment of the Gaussian process regression model.

[0173] Figure 4High-resolution micron-scale CT scan sequences of granular material samples under different wet-dry cycles (0, 5, 10, and 15 cycles) are presented. With increasing cycle number, significant changes in image grayscale levels are observed, with bright cemented areas gradually decreasing and dark pore areas expanding and becoming interconnected. This visually reflects the dissolution, loss, and pore expansion of the cemented layer caused by environmental factors. Combined with the data in Table 1 of the examples, the cemented layer volume percentage decreased from 12.45% to 5.34%, the connectivity decreased from 82.3% to 35.8%, and the equivalent pore diameter increased several times, indicating that the strong cemented contact between particles gradually degraded into unstable point and surface contact. This image sequence provides an intuitive visualization of structural degradation for subsequent deep learning multi-component segmentation and microscopic topological parameter extraction.

[0174] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0175] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0176] It should be particularly noted that, through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, or of course, by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for analyzing the degradation of granular materials based on deep learning and microscopic topological analysis, characterized in that, Includes the following steps: Step S1: Obtain granular material samples that have undergone different environmental action stages, and perform computed tomography (CT) tests on them to obtain high-resolution micron-scale CT scan sequence images; Step S2: Use a deep learning algorithm to perform multi-component pixel-level image segmentation on the high-resolution micron CT scan sequence image, and output a pixel-level segmentation result image containing four types of microstructure elements: oxide cement layer, skeleton particles, clay matrix and pores. Step S3: Use image analysis algorithms to extract quantitative parameters characterizing the micro-deterioration degree of granular materials from the pixel-level segmentation result image. The quantitative parameters include at least the area ratio of oxide cementing layer, the average thickness of oxide cementing layer, the area ratio of pores, the equivalent diameter distribution of pores, the connectivity of cementing layer, and the proportion of inter-particle contact types. Step S4: Obtain macroscopic mechanical test data of the same batch of granular material samples corresponding to the different environmental action stages; based on the extracted quantitative parameters and the macroscopic mechanical test data, establish a multivariate mapping model from microstructural parameters to macroscopic mechanical property degradation. Step S5: Using the established multivariate mapping model, analyze the degree of degradation and predict the strength of the granular material under test.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: The granular material samples that have undergone wet-dry or freeze-thaw cycles are subjected to liquid nitrogen quick-freezing and vacuum freeze-drying to preserve their original microstructure. The dried sample was vacuum impregnated with low-viscosity epoxy resin. After the resin cured, the embedded sample was scanned with high-resolution micron CT using X-ray computed tomography equipment to obtain a sequence of images containing information on the internal three-dimensional microstructure. The reconstructed CT sequence images were preprocessed sequentially with ring artifact correction, beam hardening correction, and three-dimensional nonlocal mean filtering for noise reduction.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: The acquired high-resolution micron-sized CT scan sequence images were subjected to adaptive contrast enhancement and resampling normalization preprocessing. The preprocessed image is input into a pre-trained deep learning semantic segmentation network. The network adopts the U-Net architecture with spatial attention mechanism and is optimized by a combination of multi-class cross-entropy loss and Dice loss to output a pixel-level segmentation result image with accurate boundaries for the four types of micro-structural elements.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: Quantitative parameters of four basic morphology parameters were extracted from the pixel-level segmentation result image: the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the area ratio of the pores, and the distribution of the equivalent diameter of the pores. Extract the total length of the contact surface of the skeleton particles in the pixel-level segmentation result image, and the length of the contact surface continuously covered by the oxide cementing layer. Calculate the ratio between the two to obtain the cementing layer connectivity, which characterizes the three-dimensional topological features of the cemented network. Extract and calculate the number of point contacts, surface contacts, and cemented contacts, and calculate the percentage of each contact type to the total number of contacts to obtain the proportion of interparticle contact types that characterize the stability of the skeleton.

5. The method according to claim 4, characterized in that, The process of extracting the quantitative parameters is as follows: include: The connectivity of the adhesive layer is obtained in the following way: Extract the total length of all skeleton particle contact surfaces in the pixel-level segmentation result image, and the length of the contact surface continuously covered by the oxide cementing layer. Calculate the ratio of the covered contact surface length to the total length, and use this as the cementing layer connectivity rate to characterize the three-dimensional topological features of the cemented network. The ratio of interparticle contact types is obtained in the following way: A morphological dilation operation with a specified radius is performed on the discrete skeleton particle region in the pixel-level segmentation result image to obtain the dilated overlapping feature region; based on the geometric features of the dilated overlapping feature region and its coverage by the oxide cementing layer, the contact type between the skeleton particles is determined to be point contact, surface contact or cementing contact; the number of point contact, surface contact and cementing contact are counted respectively, and the percentage of each contact type to the total number of contacts is calculated. The average thickness of the oxide cement layer is obtained by the following method: Extract the central skeleton line of the oxide cement layer region and perform a three-dimensional Euclidean distance transformation on the region to obtain the distance transformation matrix; take the corresponding value of each voxel point on the central skeleton line in the distance transformation matrix as the half thickness at that point, and calculate twice the mean of all half thicknesses as the average thickness of the oxide cement layer. The equivalent pore diameter distribution is obtained through the following method: For the pore region in the pixel-level segmentation result image, local extrema are found based on three-dimensional distance transformation as seed points and a three-dimensional watershed segmentation algorithm is run to segment the interconnected pore network into independent pore unit connected domains; the volume of each connected domain is counted and the equivalent diameter of each connected domain is calculated using the sphere volume inverse formula, thereby obtaining the pore equivalent diameter distribution.

6. The method according to claim 5, characterized in that, The extraction process of the ratio of interparticle contact types also includes: A morphological dilation operation with a specified radius is performed on the discrete skeleton particle region in the pixel-level segmentation result image to obtain the dilated overlapping feature region. If the overlapping feature area is extremely small and not covered by any adhesive, it is determined to be point contact; if the overlapping feature area is distributed as a line segment and not covered by any adhesive, it is determined to be surface contact; if the overlapping feature area is mainly covered by oxide adhesive layer pixels, it is determined to be adhesive contact.

7. The method according to claim 1, characterized in that, Step S4 specifically includes: Macroscopic mechanical tests were conducted on the same batch of granular material samples corresponding to different environmental action stages to obtain macroscopic mechanical test data including unconfined compressive strength and compressive modulus. The extracted quantitative parameters are used as the independent variable matrix, and the macroscopic mechanical test data are used as the dependent variable. The Gaussian process regression algorithm is used to eliminate multicollinearity and establish the multivariate mapping model.

8. The method according to claim 1, characterized in that, Step S5 specifically includes: Acquire CT scan images of the granular material to be tested, and extract quantitative parameters of the soil to be tested according to steps S2 to S3, including the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the pore area ratio, the pore equivalent diameter distribution, the cement layer connectivity rate, and the proportion of interparticle contact types. The six quantitative parameters of the soil under test, namely the area ratio of the oxide cement layer, the average thickness of the oxide cement layer, the pore area ratio, the pore equivalent diameter distribution, the cement layer connectivity rate, and the proportion of interparticle contact types, are input into the multivariate mapping model to quickly invert and calculate and output the quantitative prediction result of the degree of mechanical property degradation of the soil under test.

9. A degradation analysis system for granular materials based on deep learning and microscopic topological analysis, characterized in that, include: The image acquisition and preprocessing module is used to acquire high-resolution micron-scale CT scan sequence images of in-situ processed granular material samples and perform standardized preprocessing. The image segmentation module is used to run the built-in deep learning semantic segmentation network to perform pixel-level classification of the image and output a pixel-level segmentation result image containing four types of micro-structure elements: oxide cement layer, skeleton particles, clay matrix and pores. The morphological parameter extraction module is used to automatically extract multiple quantitative parameters from the pixel-level segmentation result image using morphological and topological algorithms. The quantitative parameters include at least the area ratio of the oxide cementing layer, the average thickness of the oxide cementing layer, the pore area ratio, the pore equivalent diameter distribution, the cementing layer connectivity, and the proportion of interparticle contact types. The model building module is used to establish a multivariate mapping model between the quantitative parameters and the measured macroscopic mechanical test data during the model training phase. It also establishes a predictive model between the oxide cement layer area ratio, average thickness of oxide cement layer, pore area ratio, pore equivalent diameter distribution, cement layer connectivity, and particle contact type ratio of granular material samples and the unconfined compressive strength and compressive modulus of granular materials. The evaluation output module is used to input the quantitative parameters of the granular material to be tested into the multivariate mapping model during the practical application stage, and output a degradation degree analysis and strength prediction report.

Images