A method for rapidly statistically analyzing the structure of curing pores in concrete.

By using a deep learning semantic segmentation network model, the problem of identifying internal curing holes in concrete hole structure analysis was solved, achieving efficient and accurate hole structure detection, which is suitable for rapid quality control on engineering sites.

CN123118530APending Publication Date: 2026-11-17SHANDONG UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611319114.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for analyzing concrete pore structures cannot effectively distinguish between characteristic pores left after the release of water from internal curing materials and inherent pores in concrete. Furthermore, their detection efficiency is low, failing to meet the needs of rapid quality control on engineering sites.

Method used

A method based on deep learning semantic segmentation and morphological feature optimization is adopted. Through a deep learning semantic segmentation network model with an encoding and decoding structure, high-precision automated identification of internal maintenance pores and representative sampling under sparse distribution are achieved, and pore structure feature parameters are calculated.

Benefits of technology

It achieves high-precision and rapid internal maintenance hole structure detection, reduces reliance on professional experience, improves detection efficiency and the scientific nature of results, and is suitable for rapid quality control on engineering sites.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118530A_ABST
    Figure CN123118530A_ABST
Patent Text Reader

Abstract

This invention relates to a method for rapidly statistically analyzing the structure of internal curing pores in concrete, belonging to the field of microstructure analysis and intelligent detection of building materials. The method includes: preparing a sample; cutting and photographing the sample to obtain a cross-sectional image; constructing and training a deep learning semantic segmentation network model; using the trained model for processing to classify internal curing pores, cement mortar, and coarse and fine aggregates; assigning a value of 1 to pixels identified as internal curing pores and a value of 0 to all other pixels, resulting in a binary image retaining only the internal curing pore region; calculating pore structure feature parameters based on the binary image; and automatically generating a standardized detection report based on the calculated pore structure feature parameters. This invention achieves rapid statistical analysis based on deep learning semantic segmentation and morphological feature optimization, realizing high-precision automated identification of internal curing drainage pores, representative sampling under sparse distribution, and quantitative output of key parameters of curing efficiency, greatly improving detection efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method for rapidly statistically analyzing the pore structure within concrete curing materials, specifically a rapid quantitative detection method for the characteristic pore structure formed after the release of water from curing materials (including but not limited to superabsorbent resin, lightweight aggregate, ceramic sand, zeolite, etc.) within concrete. This invention belongs to the field of microstructure analysis and intelligent detection technology for building materials. Background Technology

[0002] Internal curing technology is an effective means to solve the problems of autogenous shrinkage and early cracking in modern high-strength concrete. By incorporating internal curing materials with "water absorption-storage-release" functions (such as superabsorbent polymer (SAP) and pre-wetted lightweight aggregates) into the mix proportion, during the cement hydration process, when the relative humidity inside the matrix decreases, the internal curing materials gradually release the pre-stored moisture to compensate for the decrease in internal humidity caused by self-drying. This inhibits autogenous shrinkage, promotes later-stage cement hydration, and improves the durability of concrete.

[0003] Accurate characterization of the pore structure left after water release from internal curing materials (hereinafter referred to as "internal curing pore structure") is the core foundation for evaluating the effectiveness of internal curing, optimizing the dosage and particle size distribution of internal curing agents, and establishing water release kinetic models. However, current detection technologies for internal curing pore structure have the following serious shortcomings:

[0004] First, there is a lack of targeted identification capabilities. Existing methods for analyzing the pore structure of concrete (including mercury intrusion porosimetry, gas adsorption, and conventional image analysis) all treat all pores as a homogeneous object, failing to distinguish between "characteristic pores left after the release of water from internal curing materials" and "inherent capillary pores, gel pores, and air bubbles introduced during construction." This confusion leads to inaccurate assessments of the efficiency of internal curing.

[0005] Second, the morphology of the water release pores is complex and varied, making traditional algorithms inadequate. The pores formed after the release of water from the internal curing material are not regular spheres, but exhibit the following characteristics: SAP particles, after water release, form pores with "wrinkled inner walls, shrinkage deformation, irregular ellipsoids, or flattened sac-like shapes"; lightweight aggregates, after water release, form "annular transition zone microcrack networks" and "aggregate-matrix interface debonding pores" around them. These pores have extremely uneven grayscale distribution in RGB images, blurred edges, and are often partially filled by secondary hydration products. Traditional grayscale thresholding and watershed algorithms not only have extremely high false detection rates but also cannot extract morphological features for classification.

[0006] Third, the proportion of internal curing pores is small and their distribution is sparse, making it difficult to guarantee the representativeness of the sampling. The content of internal curing material is usually only 0.1%~0.5% (SAP) or 5%~15% (lightweight aggregate) of cementitious materials. The characteristic pore area formed after water release usually accounts for less than 3% of the total cross-sectional area, which belongs to "sparse and weak signal". Existing image analysis methods often require taking hundreds of images and manually screening them to obtain sufficient statistical data. A single analysis takes 1 to 3 days, which cannot meet the needs of rapid quality control on the engineering site.

[0007] In summary, existing methods for analyzing concrete hole structures suffer from the aforementioned problems in detecting the structure of holes during internal curing. The root cause lies in the fact that these methods remain at the level of "measuring the total number of holes" rather than "identifying the source of the holes." Summary of the Invention

[0008] To address the aforementioned issues, this invention provides a method for rapidly statistically analyzing the structure of internal curing pores in concrete. Based on deep learning semantic segmentation and morphological feature optimization using an encoding / decoding structure, this method achieves rapid statistical analysis, enabling high-precision automated identification of internal curing drainage pores, representative sampling under sparse distribution, and quantitative output of key parameters for curing efficiency.

[0009] The technical solution of the present invention is as follows:

[0010] A method for rapidly statistically analyzing the structure of curing pores within concrete includes the following steps;

[0011] S1, Prepare cement-based material samples;

[0012] S2. The sample is cut to expose a fresh and flat cross-section. The exposed cross-section is then polished. A digital microscope is used to take random multi-point photos of the cross-section at 100x magnification to obtain RGB cross-section images.

[0013] S3: Construct and train a deep learning semantic segmentation network model, and use the trained model to process the cross-sectional image to achieve pixel-level three-class classification of internal curing holes, cement mortar, and coarse and fine aggregates.

[0014] S4, assign the pixel that was determined to be an inner curing hole in step S3 to 1, and assign the value of the rest pixels to 0, to obtain a binary image that retains only the inner curing hole area;

[0015] S5 calculates pore structure feature parameters based on binary images, including pore specific surface area, pore shape factor, equivalent pore diameter, internal maintenance pore number density, water release influence zone area, and internal maintenance efficiency index.

[0016] S6 automatically generates a standardized test report based on the calculated pore structure characteristic parameters, including a heat map of the internal curing pore distribution, statistical values ​​of each parameter, and a qualification judgment conclusion.

[0017] Preferably, the process of step S1 is as follows:

[0018] S11, Selecting raw materials:

[0019] The raw materials include ordinary silicate cement, river sand, water, water-reducing agent and internal curing material, wherein the mass parts of ordinary silicate cement, river sand and water are 20-40 parts, 30-60 parts and 10-30 parts respectively.

[0020] Ordinary silicate cement, as a cementitious material, provides a strength skeleton, ensuring the basic mechanical properties of the mortar; river sand, as fine aggregate, constitutes the mortar skeleton, and its particle size distribution directly affects the fluidity, water retention, and pore structure characteristics of the fresh mortar; water, as the hydration reaction medium, directly affects the degree of cement hydration, material strength development, pore structure characteristics, and the water absorption, expansion, and release behavior of the internal curing material in the mortar; water-reducing agent imparts good fluidity to the mortar by dispersing cement particles, ensuring the workability of the mortar and the high density after casting under low water-cement ratio conditions;

[0021] The internal curing material is one or more of the following: superabsorbent polymer (SAP), pre-absorbent lightweight aggregate, ceramsite sand, and zeolite powder. It gradually releases moisture during cement hydration through a water absorption-storage-release mechanism, compensating for the decrease in internal humidity caused by self-drying, thereby inhibiting self-shrinkage and promoting subsequent hydration. When using superabsorbent polymer (SAP), its dosage is 0.1-0.5 parts by weight; when using pre-absorbent lightweight aggregate, its dosage is 5-15 parts by weight; when using ceramsite sand, its dosage is 3-10 parts by weight; when using zeolite powder, its dosage is 2-8 parts by weight. When multiple internal curing materials are used in combination, the dosage of each component shall not exceed 70% of its individual maximum dosage, and the total dosage shall not exceed 20 parts.

[0022] The water-reducing agent is 0.5%-2.0% of the mass of ordinary Portland cement;

[0023] S12, Pretreatment of internal curing materials:

[0024] When using superabsorbent polymer (SAP) as the internal curing material, soak the weighed SAP granules in deionized water at a ratio of 50 to 200 times their water absorption capacity for at least 2 hours to allow them to fully swell to saturation. Filter the granules through a 100μm stainless steel sieve to remove surface free water, obtaining pre-absorbed saturated SAP gel granules for later use. When using pre-absorbent lightweight aggregate or ceramic sand as the internal curing material, soak the aggregate in water for 24 hours until saturation. Remove the aggregate and wipe off any surface water with a damp cloth before use. When using zeolite powder, soak the material in deionized water for 24 hours. After soaking, quickly remove free water using vacuum filtration or centrifugation.

[0025] The pretreated internal curing material does not compete with the mixing water for water absorption during mixing, thus avoiding excessive loss of the instantaneous slump of the freshly mixed slurry. At the same time, it ensures that it can gradually release water according to the changes in the relative humidity inside the slurry during the cement hydration process, achieving a long-term and stable internal curing effect.

[0026] S13, Mortar preparation and curing:

[0027] After weighing the raw materials according to step S13, put the ordinary silicate cement and river sand into the mortar mixer in sequence and dry mix for 1-3 minutes until they are evenly mixed (mixing shaft speed 20±5r / min). Then, while mixing, slowly add the water-reducing agent solution dissolved in part of the mixing water and the remaining mixing water and continue mixing for 1-3 minutes until the slurry is in a uniform flow state. Then add the pretreated and drained surface water inner curing material and continue mixing for 3-5 minutes to make it evenly dispersed in the slurry without agglomeration.

[0028] The mixed mortar is poured into the mold in layers. Each layer is vibrated moderately to remove air bubbles inside the mortar while avoiding excessive vibration that could cause the inner curing material to float or settle and separate. After pouring, the surface of the mold is leveled. The mold is then cured for 24 hours under standard curing conditions (20±2℃, relative humidity not less than 95%) before demolding. The demolded specimen is placed in a curing room under the same conditions and cured until the preset age to obtain the sample, which is then ready for subsequent cross-sectional cutting and image acquisition.

[0029] The sample size can generally be 40mm×40mm×160mm or 40mm×40mm×40mm.

[0030] Preferably, in step S2, the sample is fixed on the worktable of a water-cooled cutting machine and cut along the central axis of the sample or the direction of the specified test section. During cutting, the feed speed is controlled to be ≤5 mm / s to avoid thermal damage or microcrack propagation. After cutting, the cross section should be free of obvious saw marks and chipping defects.

[0031] During the polishing process, the cross-section is polished sequentially using 400-grit, 1000-grit, and 2000-grit silicon carbide or diamond sandpaper. Each sandpaper polishing time is 3-8 minutes, and the grinding wheel speed is controlled at 200-400 r / min. Preferably, after each polishing, the specimen is ultrasonically cleaned with anhydrous ethanol for 3-15 minutes and then dried with compressed air to remove residual abrasives and moisture until the surface of the specimen is mirror-like and the boundaries of the pore structure are clearly visible.

[0032] When taking pictures, N shooting points are determined on the polished cross section according to a uniformly distributed grid or a random number table method. It is preferred to use a digital microscope camera at 100x magnification (image resolution 640×480 pixels, corresponding to an actual area of ​​4mm×3mm, and each pixel corresponds to 10 μm).

[0033] Preferably, the method for determining N includes: fixed sampling method and adaptive sampling method; in the fixed sampling method, N≥25; in the adaptive sampling method, 20 initial images are taken first, and the porosity is statistically analyzed. When the cumulative average porosity fluctuation is ≤5%, the sampling stops, otherwise the number of images is increased to 50. During the shooting, each point maintains a fixed focal length and the same lighting conditions (illuminance 600~1500 Lux, color temperature 5000~6500K), and the camera optical axis is perpendicular to the cross section.

[0034] Preferably, in step S3, the deep learning semantic segmentation network model includes three parts: an encoder feature extraction module, a multi-scale dilated fusion pooling module, and a decoder fine restoration module.

[0035] The encoder feature extraction module includes an inlet stream, 16 sets of intermediate residual streams, and an outlet stream. The inlet stream is responsible for receiving the original cross-sectional image, performing preliminary shallow feature extraction (such as edges, colors, and simple textures), and appropriately reducing the image size. The encoder backbone adopts a multi-level depth-separable residual convolutional structure, eliminating all max pooling layers in traditional convolutional networks to avoid the loss of small hole features. The shallow edge and deep texture semantic features of the pores are extracted layer by layer through the inlet stream, 16 sets of intermediate residual streams, and the outlet stream, outputting a dense high-dimensional feature map, which solves the problem of simultaneous recognition of large, medium, and small holes in concrete with large scale spans.

[0036] The intermediate residual stream is repeated 16 times, consisting of 16 residual convolutional blocks with the same structure connected in series. Each residual convolutional block contains deep separable convolution (channel-wise convolution + pointwise convolution) and residual connections (skip connections, which directly add the input to the output). In each intermediate residual stream, the input feature map is directly added to the output feature map after the deep separable convolution is processed, and then fed into the next intermediate residual stream.

[0037] The exit stream, as the end of the encoder feature extraction module, integrates and refines the high-dimensional feature maps output from the 16 intermediate residual streams, and outputs the final high-dimensional feature map.

[0038] The multi-scale hole fusion pooling module sets up three sets of hole convolutions with dilation rates of 6, 12 and 18 in parallel to extract pore features at different scales. After feature fusion, the dimensionality is reduced by 1×1 convolution to unify the multi-scale feature channels, thus enabling pore recognition across the entire range from 10μm to 1000μm.

[0039] The decoder fine restoration module restores the low-level fine contour features of the encoder feature extraction module by skipping connections, repairs the details of the pore edges, and outputs a classification feature map with the same size as the input original image, achieving pixel-level three-class classification of internal curing pores, cement mortar, and coarse and fine aggregates.

[0040] This invention employs a multi-level depth-separable residual convolutional structure in the encoder network backbone and eliminates all max-pooling layers in traditional convolutional networks to prevent the loss of small hole features during downsampling. The image sequentially passes through the inlet stream, the intermediate residual stream repeated for sixteen sets, and the outlet stream, extracting the shallow edge features and deep texture semantic features of the inner curing holes layer by layer, and outputting a dense high-dimensional feature map, thereby solving the problem of large-scale span and difficulty in synchronous recognition of large, medium, and small holes in concrete.

[0041] The deep learning semantic segmentation network model is divided into three parts: encoder feature extraction module, multi-scale hole fusion pooling module, and decoder fine restoration module. The entire process does not require manual setting of grayscale segmentation thresholds or manual differentiation of aggregate / mortar / pores.

[0042] Preferably, the decoder fine restoration module includes an upsampling layer, skip connections, and a convolutional refinement layer. The upsampling layer progressively upsamples (enlarges) the high-dimensional feature map (small in size) finally output by the encoder feature extraction module, gradually restoring the size to the original image size. Skip connections are used in each upsampling stage to directly copy and stitch the fine contour feature map with higher spatial resolution output by the corresponding layer (lower level) of the encoder feature extraction module onto the currently upsampled feature map. After each stitching and fusion, the convolutional refinement layer connects one or more ordinary convolutional layers (such as 3×3 convolutions) to fuse and refine the stitched features and repair edge details.

[0043] Preferably, the model training process in step S3 adopts a data-driven iterative optimization process, specifically as follows:

[0044] First, the paired original RGB cross-sectional images and binary label images are normalized and divided into groups of 16 images according to batch size, and the weights of the network convolutional kernels are initialized.

[0045] Then, a single iteration is performed, that is, forward inference is performed first in each batch, and the 16 images are passed through the encoder feature extraction module, the multi-scale hole fusion pooling module and the decoder fine restoration module in sequence. The probability of each pixel belonging to three categories is output and the predicted segmentation map is obtained by taking the category with the highest probability.

[0046] To calculate the loss, the predicted segmentation map and the binary label map are substituted into the cross-entropy loss function to obtain the loss value for this batch. Considering that the proportion of internal curing holes in the cross section is usually no more than 3% and they are sparse and weak signals, the internal curing hole category is given a higher loss weight according to the ratio of background to hole of 1:5. That is, when calculating the loss, the loss value of all pixels marked as internal curing holes is multiplied by a coefficient of 5, and the loss of background (including cement mortar, coarse and fine aggregates) pixels is multiplied by a coefficient of 1, which forces the network to pay attention to sparse hole pixels and not predict all pixels as background.

[0047] Perform backpropagation, propagating the loss error back along the network layer by layer in reverse according to the chain rule, and calculating the gradient of the loss with respect to each weight;

[0048] Finally, the weights are updated using the momentum gradient descent algorithm, thus completing one iteration. Then, the next batch is loaded and the single iteration process is repeated until convergence.

[0049] Preferably, the cross-entropy loss function is:

[0050]

[0051] in, This represents the cross-entropy loss value. This represents the true pixel distribution of the binary label image. To predict the probability distribution.

[0052] The original RGB image after polishing is input into the trained model, eliminating the need for manual operations such as blackening or filling the polished surface with plaster / titanium dioxide. The encoder abstracts multi-scale semantic features through layer-by-layer convolution. A multi-scale hole fusion pooling module extracts pore features at large, medium, and small scales in parallel. The decoder then upsamples and skips through the shallow high-resolution features to restore the original image size. Finally, the model independently outputs the probability value of each pixel belonging to one of three categories: internal curing pores, cement mortar matrix, or coarse and fine aggregates. The category with the highest probability is taken as the final classification of the pixel, resulting in a three-class semantic segmentation map the same size as the original image. This segmentation criterion originates from the essential features of the components learned by the model, rather than the grayscale threshold selected by human experience in traditional methods. This is the core of this invention, avoiding manual identification and eliminating subjective errors.

[0053] During segmentation, the model relies on a triple anti-interference mechanism to ensure the reliability of results under conditions without human intervention: First, illumination robustness—when the shooting light source is at a 45° angle, causing white reflections at the edges of pores, the model relies on multi-scale void feature fusion to avoid misclassifying the reflected areas as false pores; second, spatial robustness—when the image is rotated, translated, or slightly out of focus, the model's segmentation results remain stable, without the need for manual re-cropping or correction of the shooting area; third, aggregate filtering mechanism—when the image only contains coarse aggregate and no mortar areas, the model outputs an empty mask without pores, thus preventing dark aggregate textures from being misdetected as pores. Subsequently, the model converts the three-class segmentation map into a binary mask map specifically for pore structure statistics.

[0054] Preferably, in step S5, the binary image undergoes post-processing: First, connected component analysis is performed to aggregate interconnected foreground pixels into independent hole objects, and noise points with an area less than 4 pixels (equivalent diameter less than 10μm) are removed; second, morphological closing operations are performed to repair edge breaks caused by illumination white edges and fill small-area voids inside the holes caused by secondary hydration products, making each hole object a complete and closed connected region; finally, aggregate boundary removal is performed to eliminate false holes located within the aggregate region or crossing the aggregate boundary; after post-processing, each connected region in the binary image corresponds to a real internal maintenance feature hole;

[0055] For the post-processed binary image, calculate the pore structure feature parameters:

[0056] ① Pore specific surface area:

[0057]

[0058] In the formula, This represents the specific surface area of ​​the pores. Let be the projected perimeter of a single pore. It represents the projected area of ​​a single pore; the pore surface area reflects the degree of contact between the pore wall and the slurry per unit area. The larger the value, the higher the proportion of micropores and the finer the pore structure.

[0059] This step relies on image morphology algorithms for automatic calculation, requiring no manual software operation:

[0060] Perform connected component analysis on the binary mask to identify the connected regions of each independent inner maintenance hole;

[0061] The projected area A of a single pore is calculated based on pixel coordinates: total number of pixels in the connected region × actual area of ​​a single pixel (1 pixel in this system = 10μm × 10μm).

[0062] The projected perimeter P of a single pore: Extract the edge pixels of the connected domain, use the 8-neighbor contour tracking algorithm to calculate the total edge length, and convert it into the actual physical size.

[0063] This method integrates the calculation code and performs calculations automatically after the model is segmented, eliminating the need for manual selection and statistics using third-party image software (ImageJ, Photoshop, etc.), making the entire process fully automated.

[0064] ② Stomatal shape factor:

[0065]

[0066] In the formula, For pore shape factor;

[0067] ③ Equivalent aperture:

[0068]

[0069] In the formula, For equivalent aperture, This represents the total number of internal maintenance feature holes identified within the cross-sectional field of view. The equivalent aperture directly reflects the spatial uniformity of the internal curing material particles in the concrete. The higher the number density and the more uniform the distribution, the wider the coverage of the internal curing effect.

[0070] ④ Density of internal curing holes:

[0071] Will Equivalent to a circle, the equivalent circle diameter is obtained. The proportion of the corresponding hole area to the total area of ​​the internal maintenance holes is statistically analyzed according to the diameter range, and a cumulative hole diameter distribution curve is plotted. The hole diameters corresponding to 10%, 50%, and 90% of the cumulative area proportion are then selected. , and ,in The center aperture represents the typical size of the internal curing aperture. and The distribution boundaries of the small and large pore ends are reflected respectively, and the water release expansion ratio of the internal curing agent is calculated in reverse.

[0072]

[0073] In the formula, The water release and expansion ratio of the internal curing material. The original particle size of the internal curing material in its dry state was obtained through pre-treatment testing during the initial experimental stage: laser particle size analyzer was used to measure the particle size of the raw material SAP dry powder and dried lightweight aggregate / ceramic sand; the median particle size in the dry state was output. Store the data in the system parameter library and calculate the water release expansion ratio. You can call it directly at any time;

[0074] Under the premise that the total mass and total volume of the internal curing material are constant, the water release expansion ratio of the internal curing material and the number density of internal curing pores are inversely proportional.

[0075] The number density of internal curing pores is usually defined as the number of pores per unit area (or unit volume) of concrete cross-section. When the total volume expands, if these pores are uniformly distributed in the matrix, the spacing between the pores will increase. At the same volumetric admixture, The larger the particle (the larger the expansion), the smaller the number density of holes that can be cut on a unit cross section (because large particles are fewer in number and more sparsely distributed).

[0076] ⑤ Area of ​​the water release impact zone:

[0077] The width of the annular region extending outward from the edge of the inner curing pore to the point where the matrix porosity drops to the background value is the distance of moisture diffusion, reflecting the effective radius of action of a single inner curing agent.

[0078] The criteria for determining the location where the matrix porosity drops to the background value are as follows: Using the center of the connected domain of a single internal curing pore as the origin, divide the matrix into annular layers at equal intervals outwards; statistically analyze the capillary / gel porosity of non-internal curing pores within each annular layer; select a pure mortar area far from all internal curing pores and free from water release disturbance, and calculate its average porosity as the benchmark value; when the difference between the porosity of a certain annular layer and the benchmark background porosity is <5%, that annular layer is the outer boundary of the water release influence zone, and the distance from the annular layer to the pore edge is the water diffusion distance. .

[0079] The affected area of ​​water release is calculated using the following formula:

[0080]

[0081] In the formula, The area of ​​the water release influence zone of a single internal maintenance hole; The equivalent radius of the internal maintenance hole. ; The coverage rate is the distance affected by water diffusion; the ratio of the area affected by water release to the total cross-sectional area is the coverage rate, which reflects the degree of coverage of the internal curing effect.

[0082] ⑥ Internal maintenance efficiency index:

[0083]

[0084] In the formula, Internal maintenance efficiency index; Equivalent aperture; The roughness index of the hole wall. , For the circumference of an ideal smooth circular hole, ; The hole spacing variation coefficient is... Calculate the center-to-center distance between all adjacent internal curing holes. , These represent the mean and standard deviation of the interval, respectively. The higher the value, the more fully the internal curing material releases water, the more evenly it is distributed, and the better its effect on compensating for the moisture content of the substrate.

[0085] Preferably, in step S6, the entire report generation process does not require manual tabulation or statistics; the program automatically summarizes and outputs the data after the model segmentation is completed.

[0086] The internal curing effect is deemed satisfactory and the conclusion "Internal curing effect satisfactory" is output when both of the following conditions are met simultaneously:

[0087] A. Internal Maintenance Efficiency Index ≥Preset threshold, which is pre-calibrated according to the specific mix proportion and engineering requirements. For example, for C50 high-strength self-compacting concrete, it can be taken as ≥Preset threshold. ≥120 is used as a pass / fail criterion;

[0088] B. The coverage rate of the water release impact area is ≥80% of the cross-sectional area, that is, the internal curing effect can cover most of the cross-sectional area, avoiding the occurrence of large areas of "dry blind spots" that are not affected by internal curing.

[0089] If any one of the conditions is not met, the output will include the name of the parameter exceeding the limit, the extent of the exceeding limit, and specific adjustment suggestions. The adjustment suggestions include increasing the dosage of the internal curing agent, optimizing the particle size distribution of the internal curing agent, improving the mixing and dispersion process, and adjusting the pre-absorption ratio, etc., to provide a quantitative basis for the optimization of the mix proportion.

[0090] The logic for determining the conformity of this invention is based on The system automatically determines the maintenance effect based on two objective parameters: the coverage rate of the water release impact area, eliminating subjective human error, and the conclusion standards are reproducible. It can complete batch image analysis and judgment in a short time, adapting to rapid quality control on engineering sites. The joint judgment of the two indicators takes into account both the coverage area and the maintenance capacity, avoiding misjudgment based on a single parameter.

[0091] Recommendation generation: The system compares the measured parameters with the preset acceptable thresholds and matches the corresponding optimization strategy based on the type of parameter exceeding the limit. Scenario 1: Low concentration, insufficient coverage → Recommendation to increase the dosage of internal curing agent; Basis: Increasing the dosage can improve the pore density and expand the water release influence zone; Scenario 2: Low concentration, large agglomeration pores → Recommendation: Optimize the particle size distribution of the internal curing agent and improve the mixing and dispersion process; Basis: Overly coarse particles / insufficient mixing cause agglomeration, forming oversized pores and reducing the number of effective curing pores; Scenario 3: Low Insufficient → It is recommended to adjust the pre-absorption ratio / soaking time; Basis: The inner curing material is not saturated with water, the water storage is insufficient, the size of the water release hole is too small, and the humidity compensation ability is reduced.

[0092] The core basis is as follows: Internal curing mechanism: The number, size, and uniformity of distribution of internal curing pores directly determine the substrate's moisture compensation capability; Pore structure-macroscopic performance correlation: Coverage is linearly correlated with concrete autogenous shrinkage and early cracking risk; material action law: dosage, particle size, pre-absorption saturation, and mixing uniformity directly regulate the structural characteristic parameters of the water release pores.

[0093] In this invention, the pore surface area is used to evaluate the degree of pore refinement; The larger the diameter, the higher the proportion of micropores and the larger the cement hydration contact area; Application: In conjunction with IEI, it can be used to comprehensively determine the adequacy of water release during internal curing and explain the mechanism of early shrinkage and durability of concrete.

[0094] Pore ​​shape factor is used to distinguish pore types; ≈1 represents a near-spherical ideal water release hole; Values ​​much greater than 1 represent wrinkles, flattening, cracks, and pores (typical characteristics of SAP and lightweight aggregate water release); applications: filtering regularly spherical air bubbles introduced during construction, precisely screening for internal curing characteristic pores, and assisting in correction. ,calculate.

[0095] The water absorption and swelling ratio is used to quantitatively characterize the water absorption and swelling and water release and shrinkage behavior of internal curing materials; it reflects the water storage capacity; applications include: optimizing the pre-absorption ratio of SAP and the soaking time of lightweight aggregates, and guiding the selection of internal curing materials and pretreatment processes.

[0096] All three are auxiliary evaluation parameters: Coverage rate is a core indicator for determining the qualification of an engineering project; specific surface area of ​​pores. Stomatal shape factor Water release expansion ratio Used for mechanism analysis, explanation of mix proportion mechanism, and optimization of material pretreatment process, as a supplementary analytical indicator in test reports.

[0097] For any details not covered in this invention, please refer to the prior art.

[0098] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0099] High-precision identification: This invention applies deep learning semantic segmentation technology based on encoding and decoding structures to the analysis of concrete hole structures. Compared with traditional gray-scale thresholding and watershed algorithms, this invention achieves an overall accuracy of over 95% for mortar hole structures on the test set.

[0100] High-efficiency analysis: This invention eliminates the need for complex contrast-enhancing preprocessing (such as applying plaster or filling powder), requiring only conventional cutting and polishing for photographic analysis. It is significantly faster than traditional methods (ranging from hours to days), greatly improving detection efficiency.

[0101] Strong robustness and generalization: Data augmentation with various lighting conditions and image space transformations was introduced during model training, enabling the trained model to effectively resist interference from changes in shooting lighting, slight surface inhomogeneities, different magnifications, and image rotation and translation. Simultaneously, the model can effectively distinguish pore structures, aggregates, and the matrix, accurately segmenting even coarse aggregates with surface features similar to pores.

[0102] Simple to operate and highly applicable: This invention significantly reduces the professional experience required of experimental personnel, eliminates the need for manual selection of segmentation thresholds, and avoids subjective errors. Simultaneously, the reasonable sample size determination method ensures the scientific validity and representativeness of the statistical results, making it suitable for rapid quality control in scientific research and engineering fields. Attached Figure Description

[0103] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0104] Figure 1 This is a flowchart of the method for rapidly statistically analyzing the structure of curing pores in concrete according to the present invention;

[0105] Figure 2 This is a schematic diagram of the encoder feature extraction module of the present invention. Detailed Implementation

[0106] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. However, this is not the only description; all aspects not described in detail herein are based on conventional techniques in the art.

[0107] Example 1

[0108] A method for rapidly statistically analyzing the structure of curing pores within concrete, such as... Figure 1 As shown, it includes the following steps;

[0109] S1, Prepare cement-based material samples;

[0110] S2. The sample is cut to expose a fresh and flat cross-section. The exposed cross-section is then polished. A digital microscope is used to take random multi-point photos of the cross-section at 100x magnification to obtain RGB cross-section images.

[0111] S3: Construct and train a deep learning semantic segmentation network model, and use the trained model to process the cross-sectional image to achieve pixel-level three-class classification of internal curing holes, cement mortar, and coarse and fine aggregates.

[0112] S4, assign the pixel that was determined to be an inner curing hole in step S3 to 1, and assign the value of the rest pixels to 0, to obtain a binary image that retains only the inner curing hole area;

[0113] S5 calculates pore structure feature parameters based on binary images, including pore specific surface area, pore shape factor, equivalent pore diameter, internal maintenance pore number density, water release influence zone area, and internal maintenance efficiency index.

[0114] S6 automatically generates a standardized test report based on the calculated pore structure characteristic parameters, including a heat map of the internal curing pore distribution, statistical values ​​of each parameter, and a qualification judgment conclusion.

[0115] Example 2

[0116] A method for rapidly statistically analyzing the structure of curing pores within concrete, as described in Example 1, differs in that step S1 is as follows:

[0117] S11, Selecting raw materials:

[0118] The raw materials include ordinary silicate cement, river sand, water, water-reducing agent and internal curing material, wherein the mass parts of ordinary silicate cement, river sand and water are 20-40 parts, 30-60 parts and 10-30 parts respectively.

[0119] Ordinary silicate cement, as a cementitious material, provides a strength skeleton, ensuring the basic mechanical properties of the mortar; river sand, as fine aggregate, constitutes the mortar skeleton, and its particle size distribution directly affects the fluidity, water retention, and pore structure characteristics of the fresh mortar; water, as the hydration reaction medium, directly affects the degree of cement hydration, material strength development, pore structure characteristics, and the water absorption, expansion, and release behavior of the internal curing material in the mortar; water-reducing agent imparts good fluidity to the mortar by dispersing cement particles, ensuring the workability of the mortar and the high density after casting under low water-cement ratio conditions;

[0120] The internal curing material is one or more of the following: superabsorbent polymer (SAP), pre-absorbent lightweight aggregate, ceramsite sand, and zeolite powder. It gradually releases moisture during cement hydration through a water absorption-storage-release mechanism, compensating for the decrease in internal humidity caused by self-drying, thereby inhibiting self-shrinkage and promoting subsequent hydration. When using superabsorbent polymer (SAP), its dosage is 0.1-0.5 parts by weight; when using pre-absorbent lightweight aggregate, its dosage is 5-15 parts by weight; when using ceramsite sand, its dosage is 3-10 parts by weight; when using zeolite powder, its dosage is 2-8 parts by weight. When multiple internal curing materials are used in combination, the dosage of each component shall not exceed 70% of its individual maximum dosage, and the total dosage shall not exceed 20 parts.

[0121] The water-reducing agent is 0.5%-2.0% of the mass of ordinary Portland cement;

[0122] S12, Pretreatment of internal curing materials:

[0123] When using superabsorbent polymer (SAP) as the internal curing material, soak the weighed SAP granules in deionized water at a ratio of 50 to 200 times their water absorption capacity for at least 2 hours to allow them to fully swell to saturation. Filter the granules through a 100μm stainless steel sieve to remove surface free water, obtaining pre-absorbed saturated SAP gel granules for later use. When using pre-absorbent lightweight aggregate or ceramic sand as the internal curing material, soak the aggregate in water for 24 hours until saturation. Remove the aggregate and wipe off any surface water with a damp cloth before use. When using zeolite powder, soak the material in deionized water for 24 hours. After soaking, quickly remove free water using vacuum filtration or centrifugation.

[0124] The pretreated internal curing material does not compete with the mixing water for water absorption during mixing, thus avoiding excessive loss of the instantaneous slump of the freshly mixed slurry. At the same time, it ensures that it can gradually release water according to the changes in the relative humidity inside the slurry during the cement hydration process, achieving a long-term and stable internal curing effect.

[0125] S13, Mortar preparation and curing:

[0126] After weighing the raw materials according to step S13, put the ordinary silicate cement and river sand into the mortar mixer in sequence and dry mix for 1-3 minutes until they are evenly mixed (mixing shaft speed 20±5r / min). Then, while mixing, slowly add the water-reducing agent solution dissolved in part of the mixing water and the remaining mixing water and continue mixing for 1-3 minutes until the slurry is in a uniform flow state. Then add the pretreated and drained surface water inner curing material and continue mixing for 3-5 minutes to make it evenly dispersed in the slurry without agglomeration.

[0127] The mixed mortar is poured into the mold in layers. Each layer is vibrated moderately to remove air bubbles inside the mortar while avoiding excessive vibration that could cause the inner curing material to float or settle and separate. After pouring, the surface of the mold is leveled. The mold is then cured for 24 hours under standard curing conditions (20±2℃, relative humidity not less than 95%) before demolding. The demolded specimen is placed in a curing room under the same conditions and cured until the preset age to obtain the sample, which is then ready for subsequent cross-sectional cutting and image acquisition.

[0128] The sample size can generally be 40mm×40mm×160mm or 40mm×40mm×40mm.

[0129] Example 3

[0130] A method for rapidly statistically analyzing the structure of curing pores in concrete, as described in Example 1, except that in step S2, the sample is fixed on the worktable of a water-cooled cutting machine and cut along the central axis of the sample or the direction of the specified test section. During cutting, the feed speed is controlled to be ≤5 mm / s to avoid cutting thermal damage or microcrack propagation. After cutting, the cross section should be free of obvious saw marks and chipping defects.

[0131] During the polishing process, the cross-section is polished sequentially using 400-grit, 1000-grit, and 2000-grit silicon carbide or diamond sandpaper. Each sandpaper polishing time is 3-8 minutes, and the grinding wheel speed is controlled at 200-400 r / min. Preferably, after each polishing, the specimen is ultrasonically cleaned with anhydrous ethanol for 3-15 minutes and then dried with compressed air to remove residual abrasives and moisture until the surface of the specimen is mirror-like and the boundaries of the pore structure are clearly visible.

[0132] When taking pictures, N shooting points are determined on the polished cross section according to a uniformly distributed grid or a random number table method. It is preferred to use a digital microscope camera at 100x magnification (image resolution 640×480 pixels, corresponding to an actual area of ​​4mm×3mm, and each pixel corresponds to 10 μm).

[0133] Preferably, the method for determining N includes: fixed sampling method and adaptive sampling method; in the fixed sampling method, N≥25; in the adaptive sampling method, 20 initial images are taken first, and the porosity is statistically analyzed. When the cumulative average porosity fluctuation is ≤5%, the sampling stops, otherwise the number of images is increased to 50. During the shooting, each point maintains a fixed focal length and the same lighting conditions (illuminance 600~1500 Lux, color temperature 5000~6500K), and the camera optical axis is perpendicular to the cross section.

[0134] Example 4

[0135] A method for rapidly statistically analyzing the curing pore structure within concrete, as described in Example 3, differs in that, in step S3, the deep learning semantic segmentation network model comprises three parts: an encoder feature extraction module, a multi-scale void fusion pooling module, and a decoder fine restoration module.

[0136] The encoder feature extraction module includes an inlet stream, 16 sets of intermediate residual streams, and an outlet stream. The inlet stream is responsible for receiving the original cross-sectional image, performing preliminary shallow feature extraction (such as edges, colors, and simple textures), and appropriately reducing the image size. The encoder backbone adopts a multi-level depth-separable residual convolutional structure, eliminating all max pooling layers in traditional convolutional networks to avoid the loss of small hole features. The shallow edge and deep texture semantic features of the pores are extracted layer by layer through the inlet stream, 16 sets of intermediate residual streams, and the outlet stream, outputting a dense high-dimensional feature map, which solves the problem of simultaneous recognition of large, medium, and small holes in concrete with large scale spans.

[0137] The intermediate residual stream is repeated 16 times, consisting of 16 residual convolutional blocks with the same structure connected in series. Each residual convolutional block contains deep separable convolution (channel-wise convolution + pointwise convolution) and residual connections (skip connections, which directly add the input to the output). In each intermediate residual stream, the input feature map is directly added to the output feature map after the deep separable convolution is processed, and then fed into the next intermediate residual stream.

[0138] The intermediate residual flow extracts increasingly abstract and semantic features layer by layer. Shallow blocks focus on details such as edges and corners; deep blocks focus on high-level semantics such as the overall shape and texture of pores and their relative relationship with aggregate / matrix. Through 16 stacked groups, it achieves multi-scale feature representation capabilities from 10μm micropores to 1000μm macropores.

[0139] As the end of the encoder feature extraction module, the exit stream integrates and refines the high-dimensional feature maps output from 16 sets of intermediate residual streams, and outputs the final high-dimensional feature map, which contains both shallow fine contour information and deep semantic category information of the pores.

[0140] The connection methods for each layer are as follows, such as Figure 2 As shown:

[0141] Serial connection: Image → Inlet flow → Group 1 intermediate residual flow → Group 2 intermediate residual flow → ... → Group 16 intermediate residual flow → Outlet flow. Data flows unidirectionally in sequence.

[0142] Residual connections (internal): In each set of intermediate residual streams, the input feature map is directly added to the output feature map after depthwise separable convolution via a shortcut before being fed into the next set. This solves the vanishing gradient problem in deep networks while preserving low-level details.

[0143] Cancel max pooling layers: The downsampling (size reduction) task is entirely controlled by the stride in the depthwise separable convolution (if stride > 1), or only downsampling is performed a limited number of times in the ingress stream, and the size of subsequent feature maps remains relatively constant to prevent the loss of small hole features during pooling.

[0144] The multi-scale hole fusion pooling module sets up three sets of hole convolutions with dilation rates of 6, 12 and 18 in parallel to extract pore features at different scales. After feature fusion, the dimensionality is reduced by 1×1 convolution to unify the multi-scale feature channels, thus enabling pore recognition across the entire range from 10μm to 1000μm.

[0145] The decoder fine restoration module restores the low-level fine contour features of the encoder feature extraction module by skipping connections, repairs the details of the pore edges, and outputs a classification feature map with the same size as the input original image, achieving pixel-level three-class classification of internal curing pores, cement mortar, and coarse and fine aggregates.

[0146] This invention employs a multi-level depth-separable residual convolutional structure in the encoder network backbone and eliminates all max-pooling layers in traditional convolutional networks to prevent the loss of small hole features during downsampling. The image sequentially passes through the inlet stream, the intermediate residual stream repeated for sixteen sets, and the outlet stream, extracting the shallow edge features and deep texture semantic features of the inner curing holes layer by layer, and outputting a dense high-dimensional feature map, thereby solving the problem of large-scale span and difficulty in synchronous recognition of large, medium, and small holes in concrete.

[0147] The deep learning semantic segmentation network model is divided into three parts: encoder feature extraction module, multi-scale hole fusion pooling module, and decoder fine restoration module. The entire process does not require manual setting of grayscale segmentation thresholds or manual differentiation of aggregate / mortar / pores.

[0148] The specific structure of the multi-scale hole fusion pooling module is as follows:

[0149] (1) Structural composition:

[0150] The upsampling layer uses parallel dilated convolutional layers: the module's input (i.e., the high-dimensional feature map output by the encoder) is simultaneously copied into three copies and fed into three parallel dilated convolutional layers with different dilation rates.

[0151] Skip connections, or feature fusion layers, are used to concatenate the feature maps from the three sets of convolutions above, which have the same size (H / 2 × W / 2) but different receptive fields, together along the channel direction to form a multi-scale feature fusion map.

[0152] The convolutional refinement layer is a 1×1 convolutional dimensionality reduction layer: the fused high-channel-count feature map is input into a 1×1 convolutional layer, compressing the number of channels to a uniform dimension that is the same as or lower than the number of channels output by the encoder.

[0153] Where H and W are the height and width of the original image, and C is the number of channels.

[0154] (2) Connections and functions of each structure:

[0155] Connection: Encoder output → Parallel copy → (Dilated convolution 6, dilated convolution 12, dilated convolution 18) → Three-way output → Channel splicing → 1×1 convolution → Output unified feature map.

[0156] Functions: 1) Three sets of dilated convolutions: Each independently extracts features of pores at different scales. A dilation rate of 6 focuses on small scales (small receptive field), 12 on medium scales, and 18 on medium-to-large scales (large receptive field), covering the entire range of 10-1000μm in parallel. 2) Feature fusion (stitching): Merges information from different scales along the channel dimension, enabling subsequent networks to make comprehensive judgments. 3) 1×1 convolutional dimensionality reduction: This is a standard convolutional layer (1×1 kernel size). It serves two purposes: first, it fuses information between channels, allowing for interaction and filtering of the stitched multi-scale features; second, it reduces dimensionality, decreasing the number of channels, reducing subsequent computation, and unifying the feature representation dimension.

[0157] Three sets of dilated convolutions refer to three parallel, depth-separable convolutional layers with the same kernel size (all 3×3) but dilation rates of 6, 12, and 18, respectively.

[0158] "1×1 convolution" refers to a standard convolutional layer with a kernel space size of 1×1, which operates on each pixel position of the feature map and is used for cross-channel information integration and channel number transformation.

[0159] Furthermore, the decoder's fine restoration module uses shallow skip connections to access the encoder's low-level fine contour features, repairing pore edge details and outputting a classification feature map with the same size as the original input image, achieving pixel-level three-class classification of internal curing pores, cement mortar, and coarse and fine aggregates. The decoder's fine restoration module includes an upsampling layer, skip connections, and a convolutional refinement layer. The upsampling layer (deconvolution or interpolation) progressively upsamples (enlarges) the high-dimensional feature map (smaller in size) output by the encoder's feature extraction module, gradually restoring its size to the original image size. Skip connections (fusion layers) are used at each upsampling stage to directly copy and stitch the higher spatial resolution fine contour feature map output by the corresponding layer (lower level) of the encoder's feature extraction module onto the currently upsampled feature map. After each stitching and fusion, the convolutional refinement layer connects one or more ordinary convolutional layers (such as 3×3 convolutions) to fuse and refine the stitched features, repairing edge details.

[0160] Connections and functions of each structure:

[0161] 1) The encoder's 16th group output (low resolution, high semantics → upsampling (×2) → obtaining feature map A).

[0162] 2) The encoder's 8th group output (semantic in high resolution, via skip connections) → concatenates with feature map A → fuses feature map B.

[0163] 3) Fuse feature map B → 3×3 convolution refinement → obtain the output of this stage. 4) Repeat the above process until the size is restored to the original image H×W.

[0164] Functions: 1) Upsampling layer: Restores the spatial resolution of abstract semantic features, providing a locational basis for pixel-level classification. 2) Skip connection: A key function. It directly introduces the fine spatial information (such as sharp edges of pores and minute contours) preserved in the shallow layers (lower layers) of the encoder into the decoder, compensating for details lost during upsampling and thus repairing pore edges. 3) Convolutional refining layer: Fuses features from high and low layers, eliminates discontinuities caused by splicing, and generates smoother and more accurate feature maps.

[0165] Mathematical principles of core operators:

[0166] 1) Dilated separable convolution

[0167] Standard convolution computational cost:

[0168]

[0169] In the formula, This represents the total computational cost of a standard convolutional layer. The side length of the convolution kernel is 3, which is 3x3 for this network. The number of channels in the input feature map; This represents the number of channels in the output feature map. This is the side length of the output feature map.

[0170] The depth-separable convolution is split into channel-wise dilated convolution + 1×1 pointwise convolution, with a computational cost of:

[0171] In the network architecture of this invention, all depthwise separable convolutional units are uniformly split into a two-stage cascaded structure of channel-wise dilated convolution and 1×1 pointwise convolution. The encoder inlet stream, the 16 intermediate residual stream main branches, and the outlet stream are sequentially processed by 3×3 channel-wise dilated convolution to extract single-channel spatial features, followed by ReLU nonlinear activation and then connected to 1×1 pointwise convolution to achieve cross-channel feature fusion and channel dimension adjustment. Each residual block is further coupled with a shortcut jump connection to add the input features and the main branch output pixel by pixel, preserving details and mitigating gradient vanishing. The three parallel branches of the core multi-scale dilated fusion pooling module also follow this splitting paradigm, using dilation rates of 6, 4, 5, 6, 7, 8, 9, 10 ... The 3×3 per-channel dilated convolutions in modules 12 and 18 capture the spatial features of maintenance pores at small, medium, and large scales, respectively. These features are then uniformly output as channels via 1×1 pointwise convolutions within their respective branches. The three features are then concatenated along the channels and further processed by independent 1×1 pointwise convolutions at the end of the module to achieve dimensionality reduction and multi-scale feature interaction. The convolutional refinement layer after each round of upsampling and skip-concatenation in the decoder also uses a combination of per-channel dilated convolution and 1×1 pointwise convolution to repair broken edges of pores, refine contours, and fuse high and low layer features to eliminate splicing breaks. This two-level split structure decouples spatial feature extraction and channel information fusion, with per-channel dilated convolution contributing significantly to computational efficiency. · · · Spatial convolution computational power, 1×1 pointwise convolution contribution · · · The channel fusion computing power, when combined with the two, equals the total computational cost of a fully dilated separable convolution, significantly reducing computational overhead compared to standard convolution.

[0172]

[0173] Computational cost compression ratio (the ratio of computational cost of dilated separable convolution to that of standard convolution; a smaller value indicates a lighter computational load):

[0174]

[0175] This is used to quantify the degree of lightweighting. Among them, the computing power of the 3×3 convolution kernel is only 1 / 9 of that of the conventional convolution, ensuring ultra-fast inference of a single image in 0.2s; the hole spacing expands the receptive field, capturing a large area of ​​pores without increasing the size of the convolution kernel.

[0176] 2) ReLU nonlinear activation function:

[0177]

[0178] In the formula, x is the output value of the activation function; x is the input value obtained by the linear operation of the convolution kernel. The larger of 0 and x is usually chosen: Output 0 (hard saturation in the negative half-region) at this time. The output x is constant.

[0179] All convolutional units in the network are followed by ReLU activation functions, covering the following structures: Encoder: inlet stream, all 16 sets of intermediate residual streams (channel-wise dilated convolution, 1×1 pointwise convolution followed by a channel-wise dilated convolution), and outlet stream convolutional layer; Multi-scale dilated pooling module: three-way parallel dilated convolution, and a terminal 1×1 dimensionality reduction convolutional layer; Decoder: 3×3 convolutional refinement layer for each cycle.

[0180] Example 5

[0181] A method for rapidly statistically analyzing the structure of curing pores in concrete, as described in Example 4, differs in that the model training process in step S3 employs a data-driven iterative optimization process, specifically as follows:

[0182] First, the paired original RGB cross-sectional images and binary label images are normalized and divided into groups of 16 images according to batch size, and the weights of the network convolutional kernels are initialized.

[0183] Binary label images can be obtained manually by using image annotation software (LabelMe, VGG ImageAnnotator) to manually outline all SAP pores and lightweight aggregate water release feature pores on the original micrograph.

[0184] After the annotation is completed, the software exports a pixel label image, with the inner maintenance hole area marked as foreground 1 and the rest as background 0, which is a binary label image;

[0185] The labeled images are synchronously rotated, translated, and their brightness is changed to expand the training dataset.

[0186] Normalization method: Linear normalization of image pixels is adopted to adapt to neural network input: the original RGB image pixel range is 0-255; the normalization calculation method is as follows: Map all pixels to the [0, 1] interval; Optional mean and variance standardization: Standardize using the mean and variance of the concrete microscopic image dataset to accelerate model convergence and improve illumination robustness.

[0187] Network convolution kernels refer to the weights of all trainable convolutional layers in the network: encoder inlet / residual / outlet convolutions, multi-scale dilated fusion module three-way dilated convolutions + 1×1 dimensionality reduction convolutions, decoder all 3×3 refined convolutions, and output classification convolutional layers; network bias terms are uniformly initialized to 0.

[0188] The method for initializing the network convolutional kernel weights is as follows:

[0189] He Kaiming normal initialization is adopted: adapting to the ReLU activation function, generating normally distributed random weights based on the convolution kernel size and the number of input channels; solving the problems of activation value saturation and gradient vanishing in deep networks, and ensuring effective feature extraction in the early stage of training.

[0190] Then, a single iteration is performed, that is, forward inference is performed first in each batch, and the 16 images are passed through the encoder feature extraction module, the multi-scale hole fusion pooling module and the decoder fine restoration module in sequence. The probability of each pixel belonging to three categories is output and the predicted segmentation map is obtained by taking the category with the highest probability.

[0191] To calculate the loss, the predicted segmentation map and the binary label map are substituted into the cross-entropy loss function to obtain the loss value for this batch. Considering that the proportion of internal curing holes in the cross section is usually no more than 3% and is a sparse and weak signal, the internal curing hole category is given a higher loss weight according to the ratio of background to hole of 1:5. That is, when calculating the loss, the loss value of all pixels marked as internal curing holes is multiplied by a coefficient of 5, and the loss of background (including cement mortar, coarse and fine aggregates) pixels is multiplied by a coefficient of 1. This can amplify the error contribution of sparse hole pixels, force the model to learn hole features first, and avoid ignoring low proportion internal curing holes.

[0192] Perform backpropagation, propagating the loss error back along the network layer by layer in reverse according to the chain rule, and calculating the gradient of the loss with respect to each weight;

[0193] The process for calculating the gradient of the loss with respect to each weight is as follows:

[0194] 1) Output layer gradient: The predicted probability is output by the classification layer at the end of the decoder. The first-order partial derivative of the weighted cross-entropy loss is used to obtain the prediction error gradient of the output pixel.

[0195] 2) Decoder reverse layer-by-layer differentiation: Convolutional refinement layer: According to the differentiation rules of convolution operation, calculate the gradient of loss with respect to the weights and biases of the 3×3 convolution kernel; Upsampling layer: The gradient is back-interpolated to restore the size and passed to the skip concatenation layer; Skip connection layer: The gradient is split, one path is passed back to the current decoder input and the other path is passed back to the corresponding shallow feature layer of the encoder.

[0196] 3) Gradient backpropagation of multi-scale dilated fusion module: The gradient of 1×1 dimensionality reduction convolution is backpropagated to the three-way dilated convolution, and the weight gradients of the three sets of dilated depth separable convolution (channel-wise and point-wise convolution) are calculated respectively.

[0197] 4) Encoder gradient backpropagation: The derivatives are calculated layer by layer from the exit stream to the 16 sets of residual streams and the inlet stream; the gradient inside the residual block is divided into two paths: the gradient of the main convolution branch + the gradient of the shortcut residual pass-through are superimposed; the gradients of all dilated separable convolution weights are calculated in sequence.

[0198] 5) Summarize all parameter gradients: Iterate through all convolutional layer weights and biases to obtain the gradient corresponding to each set of parameters. The weights are then fed into the momentum gradient descent optimizer to complete the weight update.

[0199] Finally, the weights are updated using the momentum gradient descent algorithm, thus completing one iteration. Then, the next batch is loaded and the single iteration process is repeated until convergence.

[0200] Preferably, the cross-entropy loss function is:

[0201]

[0202] in, The cross-entropy loss value represents the distance between two probability distributions; the smaller the value, the closer the prediction is to the true value. This represents the true pixel distribution of the binary label image. To predict the probability distribution, the stomatal prediction error is quantified; a momentum gradient descent optimizer is then used.

[0203] The forward inference and loss calculation stages during model training are the core loss functions for optimizing network weights; the specific usage process is as follows:

[0204] 1) Batch RGB images are input into the network, encoder-hole fusion module-decoder forward propagation, outputting the predicted probability distribution of each pixel belonging to the three categories of [internal curing hole / mortar / aggregate]. ;

[0205] 2) Load the manually labeled ground truth map to obtain the true category distribution of pixels. ;

[0206] 3) , Substitute the values ​​into the weighted cross-entropy loss formula to calculate the total batch loss:

[0207] To address the issue of sparse internal curing pores (accounting for <3%), category weights are introduced: the weight of background (mortar + aggregate) is 1, and the weight of internal curing pores is 5, which weights and amplifies the contribution of pore pixel loss, thus preventing the network from predicting all of them as background.

[0208] 4) The loss value is used as the input for backpropagation. The gradient is calculated in reverse along the network and updated with momentum gradient descent to update all convolution and 1×1 convolution weights.

[0209]

[0210]

[0211] In the formula, Let be the momentum (cumulative gradient) of the t-th iteration; The momentum of the previous (t-1) iteration; The momentum parameter has a range of values. Control the retention ratio of historical gradients; This represents the gradient of the loss function with respect to the parameters in the current iteration. These are the weight parameters of the network to be optimized; The learning rate controls the step size for each parameter update.

[0212] The calculation is performed step-by-step using the chain-derivative backpropagation algorithm:

[0213] 1) Input a batch of images to complete forward inference, output the predicted segmentation map, and calculate the weighted cross-entropy total loss Loss;

[0214] 2) Starting from the output layer at the end of the decoder, calculate the derivative layer by layer in reverse: first calculate the partial derivative of the loss with respect to the convolution weights of the last layer of the decoder;

[0215] 3) The gradient is passed back layer by layer along the decoder → multi-scale dilated fusion module → encoder. At each layer, the partial derivatives of the loss with respect to the weights and bias parameters of all convolution kernels in this layer are calculated in turn using the derivative rules of convolution, dilated convolution and skip connections.

[0216] 4) After all layers have been differentiated, the set of partial derivatives corresponding to all parameters is the gradient of the current iteration. .

[0217] Regarding training hyperparameters, the initial learning rate is set to 1×10. -4 A segmented decay strategy is employed, reducing the value to half and one-tenth of its original value at 10,000 and 15,000 iterations, respectively, to achieve rapid approximation in the early stages and fine-tuning in the later stages. During training, the mean intersection-union ratio (mIoU) is calculated on the test set at regular intervals, and the trends of training loss and test loss are monitored to determine overfitting. Convergence is determined when the loss value tends to stabilize and the validation set no longer improves.

[0218] The original RGB image after polishing is input into the trained model, eliminating the need for manual operations such as blackening or filling the polished surface with plaster / titanium dioxide. The encoder abstracts multi-scale semantic features through layer-by-layer convolution. A multi-scale hole fusion pooling module extracts pore features at large, medium, and small scales in parallel. The decoder then upsamples and skips through the shallow high-resolution features to restore the original image size. Finally, the model independently outputs the probability value of each pixel belonging to one of three categories: internal curing pores, cement mortar matrix, or coarse and fine aggregates. The category with the highest probability is taken as the final classification of the pixel, resulting in a three-class semantic segmentation map the same size as the original image. This segmentation criterion originates from the essential features of the components learned by the model, rather than the grayscale threshold selected by human experience in traditional methods. This is the core of this invention, avoiding manual identification and eliminating subjective errors.

[0219] During segmentation, the model relies on a triple anti-interference mechanism to ensure the reliability of results under conditions without human intervention: First, illumination robustness—when the shooting light source is at a 45° angle, causing white reflections at the edges of pores, the model relies on multi-scale void feature fusion to avoid misclassifying the reflected areas as false pores; second, spatial robustness—when the image is rotated, translated, or slightly out of focus, the model's segmentation results remain stable, without the need for manual re-cropping or correction of the shooting area; third, aggregate filtering mechanism—when the image only contains coarse aggregate and no mortar areas, the model outputs an empty mask without pores, thus preventing dark aggregate textures from being misdetected as pores. Subsequently, the model converts the three-class segmentation map into a binary mask map specifically for pore structure statistics.

[0220] Example 6

[0221] A method for rapidly statistically analyzing the internal curing pore structure of concrete, as described in Example 5, differs in that, in step S5, the binary image undergoes post-processing: First, connected component analysis is performed to aggregate interconnected foreground pixels into independent pore objects, and noise points with an area less than 4 pixels (equivalent diameter less than 10μm) are removed; second, morphological closing operations are performed to repair edge breaks caused by illumination white edges and fill small-area voids inside the pores caused by secondary hydration products, making each pore object a complete and closed connected region; finally, aggregate boundary removal is performed to eliminate false pores located within the aggregate region or crossing the aggregate boundary; after post-processing, each connected region in the binary image corresponds to a real internal curing feature pore;

[0222] For the post-processed binary image, calculate the pore structure feature parameters:

[0223] ① Pore specific surface area:

[0224]

[0225] In the formula, This represents the specific surface area of ​​the pores. Let be the projected perimeter of a single pore. It represents the projected area of ​​a single pore; the pore surface area reflects the degree of contact between the pore wall and the slurry per unit area. The larger the value, the higher the proportion of micropores and the finer the pore structure.

[0226] This step relies on image morphology algorithms for automatic calculation, requiring no manual software operation:

[0227] Perform connected component analysis on the binary mask to identify the connected regions of each independent inner maintenance hole;

[0228] The projected area A of a single pore is calculated based on pixel coordinates: total number of pixels in the connected region × actual area of ​​a single pixel (1 pixel in this system = 10μm × 10μm).

[0229] The projected perimeter P of a single pore: Extract the edge pixels of the connected domain, use the 8-neighbor contour tracking algorithm to calculate the total edge length, and convert it into the actual physical size.

[0230] This method integrates the calculation code and performs calculations automatically after the model is segmented, eliminating the need for manual selection and statistics using third-party image software (ImageJ, Photoshop, etc.), making the entire process fully automated.

[0231] ② Stomatal shape factor:

[0232]

[0233] In the formula, For pore shape factor;

[0234] ③ Equivalent aperture:

[0235]

[0236] In the formula, For equivalent aperture, This represents the total number of internal maintenance feature holes identified within the cross-sectional field of view. The equivalent aperture directly reflects the spatial uniformity of the internal curing material particles in the concrete. The higher the number density and the more uniform the distribution, the wider the coverage of the internal curing effect.

[0237] ④ Density of internal curing holes:

[0238] Will Equivalent to a circle, the equivalent circle diameter is obtained. The proportion of the corresponding hole area to the total area of ​​the internal maintenance holes is statistically analyzed according to the diameter range, and a cumulative hole diameter distribution curve is plotted. The hole diameters corresponding to 10%, 50%, and 90% of the cumulative area proportion are then selected. , and ,in The center aperture represents the typical size of the internal curing aperture. and The distribution boundaries of the small and large pore ends are reflected respectively, and the water release expansion ratio of the internal curing agent is calculated in reverse.

[0239]

[0240] In the formula, The water release and expansion ratio of the internal curing material. The original particle size of the internal curing material in its dry state was obtained through pre-treatment testing during the initial experimental stage: laser particle size analyzer was used to measure the particle size of the raw material SAP dry powder and dried lightweight aggregate / ceramic sand; the median particle size in the dry state was output. Store the data in the system parameter library and calculate the water release expansion ratio. You can call it directly at any time;

[0241] Under the premise that the total mass and total volume of the internal curing material are constant, the water release expansion ratio of the internal curing material and the number density of internal curing pores are inversely proportional.

[0242] The number density of internal curing pores is usually defined as the number of pores per unit area (or unit volume) of concrete cross-section. When the total volume expands, if these pores are uniformly distributed in the matrix, the spacing between the pores will increase. At the same volumetric admixture, The larger the particle (the larger the expansion), the smaller the number density of holes that can be cut on a unit cross section (because large particles are fewer in number and more sparsely distributed).

[0243] ⑤ Area of ​​the water release impact zone:

[0244] The width of the annular region extending outward from the edge of the inner curing pore to the point where the matrix porosity drops to the background value is the distance of moisture diffusion, reflecting the effective radius of action of a single inner curing agent.

[0245] The criteria for determining the location where the matrix porosity drops to the background value are as follows: Using the center of the connected domain of a single internal curing pore as the origin, divide the matrix into annular layers at equal intervals outwards; statistically analyze the capillary / gel porosity of non-internal curing pores within each annular layer; select a pure mortar area far from all internal curing pores and free from water release disturbance, and calculate its average porosity as the benchmark value; when the difference between the porosity of a certain annular layer and the benchmark background porosity is <5%, that annular layer is the outer boundary of the water release influence zone, and the distance from the annular layer to the pore edge is the water diffusion distance. .

[0246] The affected area of ​​water release is calculated using the following formula:

[0247]

[0248] In the formula, The area of ​​the water release influence zone of a single internal maintenance hole; The equivalent radius of the internal maintenance hole. ; The coverage rate is the distance affected by water diffusion; the ratio of the area affected by water release to the total cross-sectional area is the coverage rate, which reflects the degree of coverage of the internal curing effect.

[0249] ⑥ Internal maintenance efficiency index:

[0250]

[0251] In the formula, Internal maintenance efficiency index; Equivalent aperture; The roughness index of the hole wall. , For the circumference of an ideal smooth circular hole, ; The hole spacing variation coefficient is... Calculate the center-to-center distance between all adjacent internal curing holes. , These represent the mean and standard deviation of the interval, respectively. The higher the value, the more fully the internal curing material releases water, the more evenly it is distributed, and the better its effect on compensating for the moisture content of the substrate.

[0252] Example 7

[0253] A method for rapidly statistically analyzing the structure of curing pores in concrete, as described in Example 6, differs in that, in step S6, the entire report generation process does not require manual tabulation or statistics; the program automatically summarizes and outputs the data after the model is segmented.

[0254] The internal curing effect is deemed satisfactory and the conclusion "Internal curing effect satisfactory" is output when both of the following conditions are met simultaneously:

[0255] A. Internal Maintenance Efficiency Index ≥Preset threshold, which is pre-calibrated according to the specific mix proportion and engineering requirements. For example, for C50 high-strength self-compacting concrete, it can be taken as ≥Preset threshold. ≥120 is used as a pass / fail criterion;

[0256] B. The coverage rate of the water release impact area is ≥80% of the cross-sectional area, that is, the internal curing effect can cover most of the cross-sectional area, avoiding the occurrence of large areas of "dry blind spots" that are not affected by internal curing.

[0257] If any one of the conditions is not met, the output will include the name of the parameter exceeding the limit, the extent of the exceeding limit, and specific adjustment suggestions. The adjustment suggestions include increasing the dosage of the internal curing agent, optimizing the particle size distribution of the internal curing agent, improving the mixing and dispersion process, and adjusting the pre-absorption ratio, etc., to provide a quantitative basis for the optimization of the mix proportion.

[0258] The logic for determining the conformity of this invention is based on The system automatically determines the maintenance effect based on two objective parameters: the coverage rate of the water release impact area, eliminating subjective human error, and the conclusion standards are reproducible. It can complete batch image analysis and judgment in a short time, adapting to rapid quality control on engineering sites. The joint judgment of the two indicators takes into account both the coverage area and the maintenance capacity, avoiding misjudgment based on a single parameter.

[0259] Recommendation generation: The system compares the measured parameters with the preset acceptable thresholds and matches the corresponding optimization strategy based on the type of parameter exceeding the limit. Scenario 1: Low concentration, insufficient coverage → Recommendation to increase the dosage of internal curing agent; Basis: Increasing the dosage can improve the pore density and expand the water release influence zone; Scenario 2: Low concentration, large agglomeration pores → Recommendation: Optimize the particle size distribution of the internal curing agent and improve the mixing and dispersion process; Basis: Overly coarse particles / insufficient mixing cause agglomeration, forming oversized pores and reducing the number of effective curing pores; Scenario 3: Low Insufficient → It is recommended to adjust the pre-absorption ratio / soaking time; Basis: The inner curing material is not saturated with water, the water storage is insufficient, the size of the water release hole is too small, and the humidity compensation ability is reduced.

[0260] The core basis is as follows: Internal curing mechanism: The number, size, and uniformity of distribution of internal curing pores directly determine the substrate's moisture compensation capability; Pore structure-macroscopic performance correlation: Coverage is linearly correlated with concrete autogenous shrinkage and early cracking risk; material action law: dosage, particle size, pre-absorption saturation, and mixing uniformity directly regulate the structural characteristic parameters of the water release pores.

[0261] In this invention, the pore surface area is used to evaluate the degree of pore refinement; The larger the diameter, the higher the proportion of micropores and the larger the cement hydration contact area; Application: In conjunction with IEI, it can be used to comprehensively determine the adequacy of water release during internal curing and explain the mechanism of early shrinkage and durability of concrete.

[0262] Pore ​​shape factor is used to distinguish pore types; ≈1 represents a near-spherical ideal water release hole; Values ​​much greater than 1 represent wrinkles, flattening, cracks, and pores (typical characteristics of SAP and lightweight aggregate water release); applications: filtering regularly spherical air bubbles introduced during construction, precisely screening for internal curing characteristic pores, and assisting in correction. ,calculate.

[0263] The water absorption and swelling ratio is used to quantitatively characterize the water absorption and swelling and water release and shrinkage behavior of internal curing materials; it reflects the water storage capacity; applications include: optimizing the pre-absorption ratio of SAP and the soaking time of lightweight aggregates, and guiding the selection of internal curing materials and pretreatment processes.

[0264] All three are auxiliary evaluation parameters: Coverage rate is a core indicator for determining the qualification of an engineering project; specific surface area of ​​pores. Stomatal shape factor Water release expansion ratio Used for mechanism analysis, explanation of mix proportion mechanism, and optimization of material pretreatment process, as a supplementary analytical indicator in test reports.

[0265] Example 8

[0266] A method for rapidly statistically analyzing the pore structure within concrete curing mortar, as described in Example 7, differs in that this example focuses on detecting the pore structure of SAP curing mortar, including:

[0267] Sample preparation: 450 kg / m³ of P.O42.5 silicate cement, 1350 kg / m³ of medium sand, and a water-cement ratio of 0.45 were used. The internal curing material was superabsorbent polymer (SAP), added at 0.3% of the cement mass. SAP particles were immersed in deionized water at a water absorption ratio of 100 times until saturated. After filtering through a 100 μm stainless steel sieve to remove surface free water, the samples were ready for use. Mortar specimens (40 mm × 40 mm × 160 mm) were cast according to the preparation method and process in step 1 and cured for 28 days before testing.

[0268] Cross-section preparation and image acquisition: The specimen was transversely cut along its center and successively polished with 400-grit, 1000-grit, and 2000-grit sandpaper. After each polishing, it was ultrasonically cleaned with anhydrous ethanol for 5 minutes and dried with compressed air until the cross-section was mirror-like. Images were acquired using a 100x digital microscope camera, with each image corresponding to an actual area of ​​4mm × 3mm and each pixel being 10μm. The sampling size was determined according to the minimum statistical sample size, with 80 images acquired from a 40mm × 40mm mortar cross-section, avoiding the 5mm edge area.

[0269] Model recognition and parameter calculation: Eighty images were input into the model in batches. Inference time per image was approximately 0.18 seconds, and the segmentation of all images was completed in approximately 14.4 seconds. The model automatically identified the wrinkled pores formed by SAP water release and calculated various parameters.

[0270] Test results: The total porosity of the internal curing mortar cross-section was 11.3%, the average pore size was 243.09 μm, and the specific surface area was 0.0117 μm. -1 The SAP water release characteristic pore number density is 126 pores / cm², the median pore size D50 is 185μm, and the water release expansion ratio λ is [missing value]. swell The value is 18.5, and the Internal Maintenance Efficiency Index (IEI) is 142.

[0271] Comparative verification: Compared with manually labeled true values, the false detection rate of this method is 0.8%, and the recognition accuracy is higher than 99%. Compared with the traditional plaster application + threshold segmentation method, the porosity test results of the traditional method are 18.3% higher, mainly because plaster residue is misjudged as a pore structure, and the pretreatment time exceeds 1 hour. The efficiency of this method is improved by more than 4 times.

[0272] Conclusion: The method of this invention can achieve high-precision and high-efficiency pore structure identification and parameter statistics for SAP internal curing mortar, accurately distinguishing between SAP folded water release pores and inherent capillary pores, and improving the internal curing efficiency index. It can effectively evaluate the water release effect.

[0273] Example 9

[0274] A method for rapidly statistically analyzing the pore structure within concrete curing, as described in Example 7, differs in that this example focuses on the detection of pore structure in lightweight aggregate-cured concrete, including:

[0275] Sample preparation: P.O42.5 silicate cement, 5-16mm continuously graded crushed stone coarse aggregate, and medium sand were used, with a water-cement ratio of 0.5 and coarse aggregate accounting for 40% of the volume. Pre-absorbent shale ceramsite lightweight aggregate was used as the internal curing material, replacing 10% of the crushed stone coarse aggregate by volume. The lightweight aggregate was soaked in water for 24 hours until saturated, and then wiped dry with a damp cloth. The concrete was mixed using a forced mixer and poured into 100mm×100mm×100mm cubic specimens, which were then cured for 28 days under standard conditions.

[0276] Cross-section preparation and image acquisition: The concrete section was cut transversely along the center using a water-cooled cutting machine, and then polished sequentially with 400-grit, 1000-grit, and 2000-grit sandpaper. It was then ultrasonically cleaned and dried with anhydrous ethanol. Images were acquired using a 100x digital microscope camera. A total of 150 images were collected from each 100mm × 100mm concrete section, using the minimum statistical sample size, avoiding the 5mm edge area.

[0277] Model recognition and parameter calculation: 150 images were batch-input into the model, and the total inference time was approximately 27 seconds. The model automatically distinguishes the annular water-releasing transition zone pores around the coarse aggregate, mortar matrix, and lightweight aggregate.

[0278] Test results: The total porosity of the concrete was 5.93%, the average pore size was 50.61 μm, the specific surface area was 0.0134 μm⁻¹, and the average shape factor was 1.14; the characteristic pore number density in the water release transition zone of the lightweight aggregate was 89 pores / cm², and the median pore size was... The thickness is 42 μm, the coverage rate of the water release impact zone is 83.5%, and the internal maintenance efficiency index is [missing information]. It is 96.

[0279] Comparative verification: Compared with manually labeled true values, the false detection rate is 1.5%, and the recognition accuracy is higher than 98%; compared with the watershed algorithm, the watershed method has a false detection rate of 27.8%, and a large number of microcracks at the edges of aggregates are misjudged as pore structures. This method has particularly significant advantages in working conditions containing coarse aggregates.

[0280] Conclusion: The method of the present invention has good identification accuracy and aggregate anti-interference ability for internally cured concrete containing coarse aggregate. It can accurately identify the pores in the annular water release transition zone around the lightweight aggregate, and the coverage rate of the water release influence zone can quantitatively evaluate the spatial coverage of the internal curing effect.

[0281] Example 10

[0282] A method for rapidly statistically analyzing the internal curing pore structure of concrete, as described in Example 7, differs in that this example optimizes the mix proportions of internal curing mortar with different SAP admixtures, including:

[0283] Sample preparation: Three SAP dosage gradients were set up, which were 0.1%, 0.3% and 0.5% of the cement mass, respectively, and were numbered M1, M2 and M3; the other mix proportions were the same as in Example 1, the water-cement ratio was 0.45, and the mortar test blocks (40mm×40mm×160mm) were cured for 28 days under standard conditions.

[0284] Testing method: Three test blocks were taken in each group, and two cross sections were taken from each test block. Eighty images were collected from each cross section. All images were automatically segmented by the model and the pore structure parameters and internal curing efficiency index were statistically analyzed.

[0285] Test results: Group M1 (0.1% SAP): Porosity 8.7%, internal curing pore density 52 pores / cm². =58, water release impact zone coverage 56%; M2 group (0.3% SAP): porosity 11.3%, internal curing pore density 126 / cm², =142, water release impact zone coverage rate 82%; M3 group (0.5% SAP): porosity 14.6%, internal curing pore density 198 / cm², =176, the coverage rate of the water release influence area is 91%, but a small number of SAP agglomerates with large pores (pore size > 500 μm) are present.

[0286] Mix ratio optimization judgment: based on The criteria for qualification are ≥120, ≥80% coverage of the water release influence zone and no obvious agglomeration macropores. Group M1 failed to meet all three criteria, indicating insufficient internal curing effect. Group M2 met all criteria and had a uniform pore structure, making it the optimal dosage. Group M3 had higher criteria but agglomeration defects, and its overall cost-effectiveness and pore structure uniformity were inferior to Group M2.

[0287] Conclusion: The method of this invention can complete the comparative analysis of the internal curing pore structure of multiple mix proportions in a short time, and quickly screen the optimal dosage through comprehensive parameters such as the internal curing efficiency index (IEI). Compared with traditional methods such as shrinkage test and strength comparison, it significantly shortens the mix proportion design cycle.

[0288] The following are the methods for determining the absence of obvious aggregated macropores:

[0289] Based on the statistical analysis of equivalent pore size distribution, a dual-criteria joint judgment is used: size threshold: internal curing pores with an equivalent diameter > 500 μm are defined as agglomerated macropores; quantity threshold: if the number of pores with an equivalent diameter > 500 μm in a single cross section is > 3, it is judged that there is obvious agglomeration defect; if the number of super-large pores in a single cross section is ≤ 2, it is considered that there are no significant agglomerated macropores and the pore structure uniformity meets the standard.

[0290] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for rapidly statistically analyzing the structure of curing pores within concrete, characterized in that, Includes the following steps; S1, Prepare cement-based material samples; S2. The sample is cut to expose a fresh and flat cross-section. The exposed cross-section is then polished. A digital microscope is used to take random multi-point photos of the cross-section at 100x magnification to obtain RGB cross-section images. S3: Construct and train a deep learning semantic segmentation network model, and use the trained model to process the cross-sectional image to achieve pixel-level three-class classification of internal curing holes, cement mortar, and coarse and fine aggregates. S4, assign the pixel that was determined to be an inner curing hole in step S3 to 1, and assign the value of the rest pixels to 0, to obtain a binary image that retains only the inner curing hole area; S5 calculates pore structure feature parameters based on binary images, including pore specific surface area, pore shape factor, equivalent pore diameter, internal maintenance pore number density, water release influence zone area, and internal maintenance efficiency index. S6 automatically generates a standardized test report based on the calculated pore structure characteristic parameters, including a heat map of the internal curing pore distribution, statistical values ​​of each parameter, and a qualification judgment conclusion.

2. The method for rapidly statistically analyzing the internal curing pore structure of concrete according to claim 1, characterized in that, The process of step S1 is as follows: S11, Selecting raw materials: The raw materials include ordinary silicate cement, river sand, water, water-reducing agent and internal curing material, wherein the mass parts of ordinary silicate cement, river sand and water are 20-40 parts, 30-60 parts and 10-30 parts respectively. The internal curing material is one or more of the following: super absorbent polymer (SAP), pre-absorbent lightweight aggregate, ceramsite sand, and zeolite powder. When super absorbent polymer (SAP) is used, its mass fraction is 0.1-0.5 parts; when pre-absorbent lightweight aggregate is used, its mass fraction is 5-15 parts; when ceramsite sand is used, its mass fraction is 3-10 parts; when zeolite powder is used, its mass fraction is 2-8 parts. When multiple internal curing materials are used in combination, the mass fraction of each component shall not exceed 70% of its upper limit for use alone, and the total mass fraction shall not exceed 20 parts. The water-reducing agent is 0.5%-2.0% of the mass of ordinary Portland cement; S12, Pretreatment of internal curing materials: When using superabsorbent polymer (SAP) as the internal curing material, soak the weighed SAP granules in deionized water at a ratio of 50 to 200 times their water absorption capacity for at least 2 hours to allow them to fully swell to saturation. Filter the granules through a 100μm stainless steel sieve to remove surface free water, obtaining pre-absorbed saturated SAP gel granules for later use. When using pre-absorbent lightweight aggregate or ceramic sand as the internal curing material, soak the aggregate in water for 24 hours until saturation. After soaking, wipe off any excess water with a damp cloth before use. When using zeolite powder, soak the material in deionized water for 24 hours. After soaking, quickly remove free water using vacuum filtration or centrifugation. S13, Mortar preparation and curing: After weighing the raw materials according to step S13, put the ordinary silicate cement and river sand into the mortar mixer in sequence and dry mix them evenly. Then, while stirring, add the water-reducing agent solution dissolved in part of the mixing water and the remaining mixing water and continue stirring for 1-3 minutes until the slurry is in a uniform flow state. Then add the pre-treated and drained surface water inner curing material and continue stirring for 3-5 minutes to make it evenly dispersed in the slurry without agglomeration. The mixed mortar was poured into the mold in layers, each layer was vibrated, and the surface of the mold was leveled after pouring. The mold was then demolded after curing under standard curing conditions for 24 hours. The demolded specimen was placed in a curing room under the same conditions and continued to be cured until the preset age to obtain the specimen for subsequent cross-sectional cutting and image acquisition.

3. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 1, characterized in that, In step S2, the sample is fixed on the worktable of the water-cooled cutting machine and cut along the central axis of the sample or the direction of the specified test section. During cutting, the feed speed is controlled to be ≤5 mm / s. During the polishing process, the cross-section is polished sequentially using 400-grit, 1000-grit, and 2000-grit silicon carbide or diamond sandpaper. Each sandpaper polishing time is 3-8 minutes, and the grinding wheel speed is controlled at 200-400 r / min. After each polishing, the surface is ultrasonically cleaned with anhydrous ethanol for 3-15 minutes and then dried with compressed air to remove residual abrasives and moisture until the surface of the specimen is mirror-like. When taking photos, N shooting points are determined on the polished cross section using a uniformly distributed grid or a random number table method.

4. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 3, characterized in that, The methods for determining N include: fixed sampling method and adaptive sampling method; in the fixed sampling method, N≥25; in the adaptive sampling method, first take 20 initial images, count the porosity, and stop when the cumulative average porosity fluctuation is ≤5%, otherwise increase to 50 images; during shooting, each point maintains a fixed focal length and the same lighting conditions, and the camera optical axis is perpendicular to the cross section.

5. The method for rapidly statistically analyzing the internal curing pore structure of concrete according to claim 1, characterized in that, In step S3, the deep learning semantic segmentation network model includes three parts: encoder feature extraction module, multi-scale dilated fusion pooling module, and decoder fine restoration module. The encoder feature extraction module includes an inlet stream, 16 sets of intermediate residual streams, and an outlet stream. The inlet stream is responsible for receiving the original cross-sectional image, performing preliminary shallow feature extraction, and reducing the image size. The intermediate residual stream is repeated 16 times, consisting of 16 residual convolutional blocks with the same structure connected in series. Each residual convolutional block contains deep separable convolution and residual connections. In each intermediate residual stream, the input feature map is directly added to the output feature map after deep separable convolution, and then fed into the next intermediate residual stream. The exit stream, as the end of the encoder feature extraction module, integrates and refines the high-dimensional feature maps output from the 16 intermediate residual streams, and outputs the final high-dimensional feature map. The multi-scale hole fusion pooling module sets up three sets of hole convolutions with dilation rates of 6, 12 and 18 in parallel to extract pore features at different scales. After feature fusion, the dimensionality is reduced by 1×1 convolution to unify the multi-scale feature channels, thus enabling pore recognition across the entire range from 10μm to 1000μm. The decoder fine restoration module restores the low-level fine contour features of the encoder feature extraction module by skipping connections, repairs the details of the pore edges, and outputs a classification feature map with the same size as the input original image, achieving pixel-level three-class classification of internal curing pores, cement mortar, and coarse and fine aggregates.

6. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 5, characterized in that, The decoder fine restoration module includes an upsampling layer, a skip connection, and a convolution refinement layer; the upsampling layer performs stepwise upsampling on the high-dimensional feature map finally output by the encoder feature extraction module, so that the size gradually returns to the size of the original image; the skip connection is used to directly copy and splice the fine contour feature map with higher spatial resolution output by the encoder feature extraction module to the current upsampled feature map through the skip connection at each upsampling stage; after each splicing and fusion, the convolution refinement layer is connected to one or more ordinary convolution layers, which are used to fuse and refine the spliced features and repair edge details.

7. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 6, characterized in that, The model training process in step S3 adopts a data-driven iterative optimization process, which is specifically as follows: First, the paired original RGB cross-section images and binary label maps are normalized and divided into groups with a batch size of 16, and the weights of the network convolution kernels are initialized; Then a single iteration is entered, that is, forward inference is performed first in each batch, 16 images are sequentially passed through the encoder feature extraction module, the multi-scale dilated fusion pooling module and the decoder fine restoration module, the probability that each pixel belongs to three categories is output, and the category with the maximum probability is taken to obtain a predicted segmentation map; Loss calculation is performed, the predicted segmentation map and the binary label map are substituted into the cross-entropy loss function to obtain the loss value of this batch, and a higher loss weight is assigned to the internal curing pore category according to a ratio of 1:5 between the background and the pores, that is, when calculating the loss, the loss value of all pixels marked as internal curing pores is multiplied by a coefficient of 5, and the loss value of background pixels is multiplied by a coefficient of 1; Backpropagation is performed, the loss error is backpropagated layer by layer along the network according to the chain rule of derivation, and the gradient of the loss with respect to each weight is calculated; Finally, weight update is performed, the momentum gradient descent algorithm is used to update all weights, so far one iteration is completed, then the next batch is loaded and the single iteration process is repeated until convergence.

8. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 7, characterized in that, The cross-entropy loss function is; in, This represents the cross-entropy loss value. This represents the true pixel distribution of the binary label image. To predict the probability distribution.

9. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 8, characterized in that, In step S5, the binary image is post-processed: first, connected domain analysis is performed to aggregate mutually connected foreground pixels into independent pore objects, and noise points with an area smaller than 4 pixels are eliminated; second, morphological closing operation is performed to repair edge fractures caused by illumination white edges and fill small-area voids inside the pores caused by secondary hydration products, so that each pore object becomes a complete and closed connected region; finally, aggregate boundary elimination is performed to exclude pseudo-pores located in the aggregate region or crossing the aggregate boundary; after post-processing, each connected domain in the binary image corresponds to a real internal curing characteristic pore; For the post-processed binary image, the pore structure characteristic parameters are calculated: ① Specific surface area of pores: In the formula, This represents the specific surface area of ​​the pores. Let be the projected perimeter of a single pore. The projected area of ​​a single pore; ② Pore shape factor: In the formula, For pore shape factor; ③ Equivalent pore diameter: In the formula, For equivalent aperture, This represents the total number of internal maintenance feature holes identified within the cross-sectional field of view. This represents the actual area of ​​the cross-sectional view. ④ Number density of internal curing pores: Will Equivalent to a circle, the equivalent circle diameter is obtained. The proportion of the corresponding hole area to the total area of ​​the internal maintenance holes is statistically analyzed according to the diameter range, and a cumulative hole diameter distribution curve is plotted. The hole diameters corresponding to 10%, 50%, and 90% of the cumulative area proportion are then selected. , and ,in The center aperture represents the typical size of the internal curing aperture. and These respectively reflect the distribution boundaries of the small-hole end and the large-hole end; In the formula, The water release and expansion ratio of the internal curing material. The original particle size of the internal curing material in its dry state; On the premise that the total mass and total volume of the internal curing material are fixed, the water release expansion ratio of the internal curing material is inversely proportional to the number density of internal curing pores; ⑤ Area of water release influence zone: In the formula, The area of ​​the water release influence zone of a single internal maintenance hole; The equivalent radius of the internal maintenance hole. ; The distance affected by moisture diffusion; ⑥ Internal curing efficiency index: In the formula, Internal maintenance efficiency index; Equivalent aperture; The roughness index of the hole wall. , For the circumference of an ideal smooth circular hole, ; The hole spacing variation coefficient is... Calculate the center-to-center distance between all adjacent internal curing holes. , These represent the mean and standard deviation of the interval, respectively.

10. The method for rapidly statistically analyzing the structure of curing pores in concrete according to claim 9, characterized in that, In step S6, when the following two conditions are satisfied at the same time, the internal curing effect is determined to be qualified, and the conclusion of "internal curing effect is qualified" is output: A. Internal Maintenance Efficiency Index ≥Preset threshold; B. The area coverage rate of the water release influence zone ≥ 80% of the cross-sectional area; If any condition is not met, the output will include the name of the parameter exceeding the limit, the extent of the exceeding limit, and specific adjustment suggestions, providing a quantitative basis for mix design optimization.