Method for resource utilization and quality grading of road steel slag aggregate

Abstract

The present application belongs to the technical field of comprehensive utilization of solid waste, and particularly relates to a method for resource utilization and quality grading of road steel slag aggregate, which comprises the following steps: step 1, obtaining multiple grades of steel slag aggregate by aging, crushing and screening the steel slag, and obtaining an expansion strain field gray image time sequence; step 2, extracting images at multiple time points from the expansion strain field gray image time sequence, stacking the images into a multi-channel input tensor, and inputting the tensor into a convolutional neural network model to output probability values of three grades of stability, i.e., excellent stability, medium stability and poor stability; and step 3, determining the mechanical durability index of each grade of steel slag aggregate with a continuous stability score reaching a set threshold value, and preparing each grade of steel slag aggregate into building materials corresponding to the road layer position. The present application breaks through the limitation of traditional autoclaved expansion rate method which only outputs a single scalar value, and realizes high-dimensional characterization of the expansion space distribution pattern and time sequence evolution pattern.

Description

Technical Field

[0001] This invention belongs to the field of solid waste comprehensive utilization technology, and involves the promotion and service of resource recycling technology. It also involves the quality classification of steel slag aggregate through image recognition technology, specifically involving a method for resource utilization and quality classification of road steel slag aggregate. Background Technology

[0002] Steel slag aggregate possesses physical and mechanical advantages such as high apparent density, abundant angularity, and strong wear resistance. In some performance indicators, it even surpasses natural basalt crushed stone, making it a potential coarse aggregate for asphalt pavement surface layers. However, the widespread application of steel slag aggregate in road engineering has consistently faced a core obstacle: volume stability. Steel slag contains a certain amount of free calcium oxide and free magnesium oxide. These two components hydrate upon contact with water, forming calcium hydroxide and magnesium hydroxide respectively, accompanied by significant volume expansion. The hydration expansion of free magnesium oxide has a latency period of several years or even more than ten years. If steel slag aggregate with unsatisfactory stability is used in road engineering, the subsequent expansion can lead to pavement heave, cracking, and even structural damage, resulting in extremely high maintenance costs and safety risks. Therefore, before using steel slag aggregate as a road construction material, its stability must be accurately evaluated and effective stabilization treatment implemented.

[0003] Currently, the mainstream method in the industry for evaluating the stability of steel slag aggregates is the autoclave expansion rate method. This method involves preparing steel slag aggregates into specimens and conducting accelerated expansion tests under high-temperature and high-pressure steam conditions. The change rate of the specimen's length or volume before and after the test is measured to determine whether the stability is up to standard. This method is standardized in operation and the results have regulatory acceptance. However, its output is only a scalar value, namely the overall expansion rate, which cannot reveal the spatial distribution characteristics of expansion on the specimen surface or the evolution process of expansion over time. In other words, the traditional autoclave expansion rate method can only provide a global statistical quantity of "how much expansion occurred," and cannot answer deeper questions such as "where the expansion is most intense" and "whether the expansion is uniform or concentrated at a certain stage." These spatial distribution patterns and temporal evolution patterns are closely related to the location, content gradient, and reactivity of free oxides inside the steel slag, and contain much richer information about the stability state than a single expansion rate value. Summary of the Invention

[0004] The main objective of this invention is to provide a method for the resource utilization and quality grading of road steel slag aggregate. This invention breaks through the limitation of the traditional autoclaving expansion rate method, which only outputs a single scalar value, and realizes a high-dimensional characterization of the expansion spatial distribution pattern and temporal evolution pattern. It establishes a closed-loop feedback control mechanism for carbonization maintenance effect and integrates stability evaluation and multi-dimensional mechanical durability indicators into a unified comprehensive grading system, so that steel slag aggregates of different qualities can be matched to road engineering layers that are compatible with their performance levels, thus realizing the full-quantity gradient resource utilization of steel slag aggregates.

[0005] To address the aforementioned technical problems, this invention provides a method for the resource utilization and quality grading of road steel slag aggregate, comprising the following steps: Step 1: Steel slag is aged, crushed and screened to obtain multiple grades of steel slag aggregate. The multiple grades of steel slag aggregate are carbonized to convert free calcium oxide and free magnesium oxide into calcium carbonate and magnesium carbonate, respectively. Samples are taken from each grade of steel slag aggregate and standard cylindrical specimens are prepared using a high-temperature resistant inorganic binder. After spraying a high-temperature resistant speckle pattern on the side surface of the standard cylindrical specimen, it is placed in an autoclave for accelerated expansion test. The speckle image pairs are acquired through a binocular stereo vision acquisition system and digital image correlation is performed to obtain the grayscale image time series of the expansion strain field. Step 2: Extract images from multiple time points in the grayscale image time series of the expansion strain field and stack them into a multi-channel input tensor. Input the tensor into the convolutional neural network model and output the probability values ​​of three levels of stability: excellent, moderate, and poor. Calculate the continuous stability score based on the probability values ​​of the three levels. When the continuous stability score is lower than the set threshold, perform supplementary carbonization curing and repeat steps 1 and 2. The cumulative number of times supplementary carbonization curing is performed does not exceed the set maximum number of times. If the continuous stability score is still lower than the set threshold after reaching the set maximum number of times, the steel slag aggregate is directly classified into the lowest quality level. Step 3: Measure the mechanical durability index of each grade of steel slag aggregate whose continuous stability score reaches the set threshold, construct a hierarchical analysis method classification framework to calculate the comprehensive quality assessment score, divide each grade of steel slag aggregate into multiple quality grades based on the comprehensive quality assessment score, and prepare the steel slag aggregate of each quality grade into building materials for the corresponding road layer.

[0006] Furthermore, in step 1, after the steel slag has been naturally piled up and aged for no less than 6 months, it is successively crushed by a jaw crusher and a secondary crusher, and then screened by a 4-layer vibrating screen to obtain four grades of steel slag aggregate with particle sizes of 19 mm to 26.5 mm, 9.5 mm to 19 mm, 4.75 mm to 9.5 mm, and 0.075 mm to 4.75 mm.

[0007] Furthermore, in step 1, the specific method of carbonization curing is as follows: multiple grades of steel slag aggregate are laid flat on the breathable mesh tray inside the sealed carbonization curing box, with a thickness of 150 mm. A mixture of carbon dioxide and nitrogen with a volume concentration of 20% is continuously introduced into the sealed carbonization curing box, and the temperature inside the sealed carbonization curing box is controlled at 60 degrees Celsius and the relative humidity at 70%, and carbonization curing is carried out continuously for 7 days.

[0008] Furthermore, in step 1, the high-temperature resistant inorganic binder is aluminum phosphate inorganic binder, and the mass of aluminum phosphate inorganic binder is 8% to 12% of the dry mass of steel slag aggregate. The diameter of the standard cylindrical specimen is 50 mm and the height is 100 mm. After preparation, it is cured at 200 degrees Celsius for 2 hours to completely ceramicize the aluminum phosphate inorganic binder. The spraying method for the high-temperature resistant speckle pattern is to first spray a layer of white high-temperature resistant inorganic primer, and after the white high-temperature resistant inorganic primer dries, spray black high-temperature resistant inorganic speckle paint. The average diameter of the speckles is 0.3 mm to 0.5 mm, and the speckle coverage is 40% to 60% of the side surface area of ​​the standard cylindrical specimen.

[0009] Furthermore, in step 1, a high-temperature resistant optical observation window with an anti-fog coating on its inner surface is installed on the side wall of the autoclave corresponding to the position of the standard cylindrical specimen. The binocular stereo vision acquisition system consists of two industrial area array cameras with a resolution of no less than 5 million pixels mounted on the outside of the high-temperature resistant optical observation window. The optical axis intersection angle of the two industrial area array cameras is 25 to 35 degrees. A narrow-band blue LED surface light source with a wavelength of 450 nanometers is configured on the outside of the high-temperature resistant optical observation window as an illumination source. The conditions for the accelerated expansion test are an autoclave temperature of 216 degrees Celsius, an autoclave pressure of 2.0 MPa, and continuous autoclaving for 3 hours. The acquisition frequency of speckle image pairs is once every 30 seconds.

[0010] Further, in step 1, the specific method of digital image correlation is as follows: the earliest acquired speckle image pair is used as the reference image pair, and the remaining speckle image pairs are used as deformed image pairs. The reference image is divided into square sub-regions with a side length of 31 pixels and a center-to-center distance of 7 pixels between adjacent square sub-regions. For each square sub-region, the correlation coefficient is calculated pixel by pixel in the deformed image using the zero-mean normalized cross-correlation criterion to determine the integer pixel matching position. In the neighborhood of the integer pixel matching position, double cubic B-spline interpolation is used to solve the displacement vector with sub-pixel accuracy. The displacement vectors of all square sub-regions are arranged according to their spatial positions to form a full-field displacement vector field. The difference between the displacement vectors of adjacent square sub-regions is taken and divided by the center-to-center distance to obtain the radial expansion strain component and the axial expansion strain component, respectively. The radial expansion strain component and the axial expansion strain component are linearly mapped to grayscale images with grayscale values ​​from 0 to 255.

[0011] Furthermore, in step 2, six time points are extracted at equal intervals. At each time point, one grayscale image of the radial expansion strain field and one grayscale image of the axial expansion strain field are taken, for a total of 12 images. These 12 images are scaled to 224 pixels by 224 pixels and then stacked to form a 224 by 224 by 12 multi-channel input tensor. The convolutional neural network model sequentially includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, a first fully connected layer, and a second fully connected layer. The first convolutional module contains 32 convolutional kernels of size 7 by 7 with a stride of 2 and padding of 3, followed by a batch normalization layer, a ReLU activation function, and a max pooling layer of size 3 by 3 with a stride of 2. The second convolutional module contains 64 kernels of size 5 by 7 with a stride of 2 and padding of 3. The first convolutional module consists of 128 kernels with a stride of 1 and padding of 2, followed by batch normalization, ReLU activation, and max pooling with a size of 3 multiplied by 3 and a stride of 2. The second convolutional module consists of 128 kernels with a size of 3 multiplied by 3 and a stride of 1 and padding of 1, followed by batch normalization and ReLU activation. The third convolutional module consists of 128 kernels with a size of 3 multiplied by 3 and a stride of 1 and padding of 1, followed by batch normalization, ReLU activation, and global average pooling. The global average pooling layer outputs a 128-dimensional feature vector. The first fully connected layer maps the 128-dimensional feature vector to 64-dimensional vectors and processes it with the ReLU activation function. The second fully connected layer maps the feature vectors to 3-dimensional vectors and outputs probability values ​​for each of the three levels through the Softmax function.

[0012] Furthermore, in step 2, the training method for the convolutional neural network model is as follows: collect no less than 500 sets of steel slag aggregate samples from different steel plants and different furnace batches; prepare standard cylindrical specimens for each set of steel slag aggregate samples according to the method in step 1 and obtain multi-channel input tensors; at the same time, take parallel samples for each set of steel slag aggregate samples and independently determine the autoclave expansion rate according to the autoclave expansion method specified by the national standard; mark autoclave expansion rate less than 0.5% as excellent stability, autoclave expansion rate between 0.5% and 2.0% as moderate stability, and autoclave expansion rate greater than 2.0% as poor stability; train the model using the cross-entropy loss function and the Adam optimizer, and save the model parameters of the round with the highest classification accuracy on the validation set.

[0013] Furthermore, in step 2, the continuous stability score is calculated by multiplying the probability value of excellent stability by 1.0, adding the probability value of moderate stability by 0.6, and adding the probability value of poor stability by 0.2; the threshold is set to 0.5, and the maximum number of times is set to 3.

[0014] Furthermore, in step 3, the mechanical durability indicators include crushing value, Los Angeles abrasion value, polishing value, needle-like particle content, and water absorption rate. The target layer of the analytic hierarchy process (AHP) grading framework is the comprehensive quality grade of road steel slag aggregate. The criterion layer sets two indicators: volume stability and mechanical durability. In the scheme layer, the volume stability indicator sets a continuous stability score, and the mechanical durability indicator sets five measured values: crushing value, Los Angeles abrasion value, polishing value, needle-like particle content, and water absorption rate. A second-order judgment matrix is ​​constructed in the criterion layer, and the Saaty 1 to 9 scaling method is used to perform pairwise comparisons and assignments on the two criteria. The normalized eigenvector corresponding to the largest eigenvalue of the second-order judgment matrix is ​​calculated to obtain the contribution values ​​of each criterion. In the scheme layer under the mechanical durability indicators, a fifth-order judgment matrix is ​​constructed, and the Saaty 1 to 9 scaling method is used to perform pairwise comparisons and assignments on the five mechanical durability indicators. The maximum eigenvalue of the fifth-order judgment matrix is ​​calculated. The normalized eigenvectors yield the contribution values ​​of the five mechanical durability indicators, and the consistency ratio is verified to be less than 0.1. The measured values ​​of crushing value, Los Angeles abrasion value, needle-like particle content, and water absorption rate are subtracted from the upper limit values ​​specified in the corresponding national standards, and the absolute value is divided by the upper limit values ​​specified in the corresponding national standards. The quotient obtained by subtracting 1.0 is used as the positive value, and 0 is taken when the positive value is less than 0. The measured value of polishing value is subtracted from the lower limit value specified in the corresponding national standards, and the quotient obtained by adding 1.0 to it and dividing by 2.0 is used as the positive value, and 1.0 is taken when the positive value is greater than 1.0. The comprehensive quality assessment score is obtained by multiplying the stability continuous score value by the criterion layer contribution value of the volume stability indicator, adding the positive values ​​of the five mechanical durability indicators by the criterion layer contribution values ​​of the mechanical durability indicators, and then multiplying them by their respective scheme layer contribution values.

[0015] The method for resource utilization and quality grading of road steel slag aggregate of the present invention has the following beneficial effects: This invention utilizes a binocular stereo vision acquisition system combined with digital image correlation to acquire the grayscale image time series of the expansion strain field throughout the entire autoclaving process. This overcomes the limitation of traditional autoclaving expansion rate methods, which can only output a single scalar value. It can simultaneously capture the spatial distribution characteristics of expansion on the specimen surface and its evolution over time, providing far richer, higher-dimensional information for characterizing the stability state than traditional methods. Furthermore, this invention uses a convolutional neural network to automatically extract and classify spatial and temporal patterns in the grayscale image time series of the expansion strain field, outputting a continuous stability score containing probability distribution information at each level. This avoids the information loss and misjudgment risks associated with manual binary judgments based on a single threshold.

[0016] This invention establishes a closed-loop control mechanism using a continuous stability score as feedback signal. For batches with insufficient carbonization, it automatically triggers supplementary carbonization curing and re-evaluation. Simultaneously, a maximum supplementary number constraint is set to balance process economy, upgrading the carbonization curing process from a traditional fixed-parameter open-loop operation to a dynamic adaptive closed-loop operation based on evaluation results. In terms of comprehensive quality grading, this invention uses an analytic hierarchy process (AHP) framework to incorporate the continuous stability score output by a convolutional neural network and multiple mechanical durability indicators into a unified multi-level weighted evaluation system. This overcomes the one-sidedness of existing technologies that rely on a single or few indicators for quality judgment, achieving an organic integration of the stability and mechanical durability dimensions.

[0017] This invention classifies steel slag aggregate into three quality grades based on comprehensive quality assessment scores and prepares them into asphalt concrete surface layer mixtures, cement-stabilized crushed stone base layer mixtures, and roadbed filling materials, respectively. This allows steel slag aggregates of different qualities to be matched to road engineering layers that are compatible with their performance levels, realizing the full-scale gradient utilization of steel slag aggregates and avoiding the resource waste caused by discarding substandard aggregates in the traditional method. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the full-field strain acquisition system for the autoclave using digital image correlation method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of sub-region matching between a reference image and a deformed image using the digital image correlation method provided in an embodiment of the present invention. Figure 3 A trend graph showing the change of the continuous stability score of steel slag aggregate with the number of carbonization curing cycles, provided in an embodiment of the present invention; Figure 4 The graph shows the comprehensive quality assessment scores and quality grade classification results of multiple steel slag aggregate samples provided in the embodiments of the present invention. Detailed Implementation

[0019] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] A method for resource utilization and quality grading of road steel slag aggregate includes the following steps: Step 1: Steel slag is aged, crushed and screened to obtain multiple grades of steel slag aggregate. The multiple grades of steel slag aggregate are carbonized to convert free calcium oxide and free magnesium oxide into calcium carbonate and magnesium carbonate, respectively. Samples are taken from each grade of steel slag aggregate and standard cylindrical specimens are prepared using a high-temperature resistant inorganic binder. After spraying a high-temperature resistant speckle pattern on the side surface of the standard cylindrical specimen, it is placed in an autoclave for accelerated expansion test. The speckle image pairs are acquired through a binocular stereo vision acquisition system and digital image correlation is performed to obtain the grayscale image time series of the expansion strain field. Step 2: Extract images from multiple time points in the grayscale image time series of the expansion strain field and stack them into a multi-channel input tensor. Input the tensor into the convolutional neural network model and output the probability values ​​of three levels of stability: excellent, moderate, and poor. Calculate the continuous stability score based on the probability values ​​of the three levels. When the continuous stability score is lower than the set threshold, perform supplementary carbonization curing and repeat steps 1 and 2. The cumulative number of times supplementary carbonization curing is performed does not exceed the set maximum number of times. If the continuous stability score is still lower than the set threshold after reaching the set maximum number of times, the steel slag aggregate is directly classified into the lowest quality level. Step 3: Measure the mechanical durability index of each grade of steel slag aggregate whose continuous stability score reaches the set threshold, construct a hierarchical analysis method classification framework to calculate the comprehensive quality assessment score, divide each grade of steel slag aggregate into multiple quality grades based on the comprehensive quality assessment score, and prepare the steel slag aggregate of each quality grade into building materials for the corresponding road layer.

[0021] In this embodiment, the first step of the method for resource utilization and quality grading of road steel slag aggregate involves carbonization curing modification treatment of steel slag aggregate and expansion strain field acquisition based on digital image correlation method. The specific implementation process of this step will be described in detail below.

[0022] Steel slag discharged during steelmaking contains a large amount of free calcium oxide and free magnesium oxide. These two components undergo a hydration reaction upon contact with water, producing calcium hydroxide and magnesium hydroxide, accompanied by significant volume expansion. This expansion manifests in road engineering as pavement bulging, cracking, and even structural damage. Therefore, before using steel slag as a road construction material, its expansion potential must be effectively eliminated or significantly reduced. This embodiment employs a two-stage stabilization strategy combining natural stocking aging and forced carbonization curing to achieve the above objective.

[0023] First, the steel slag discharged during the steelmaking process is subjected to natural stockpiling and aging for no less than 6 months. During this natural stockpiling and aging process, the larger free calcium oxide particles in the steel slag slowly hydrate and partially carbonize under the influence of atmospheric moisture and rainwater. Large pieces of steel slag naturally disintegrate due to the release of internal stress. This process stabilizes the steel slag as a whole and creates favorable conditions for subsequent mechanical crushing. In some optional embodiments, the aging time can be extended to 12 months or longer, depending on the source of the steel slag and the initial free calcium oxide content. For steel slag with a free calcium oxide content of less than 3%, the aging time can also be shortened to 3 months.

[0024] After aging, the steel slag is sequentially crushed by a jaw crusher for coarse crushing and by an impact crusher for medium crushing. The discharge particle size of the jaw crusher is controlled below 50 mm, and the discharge particle size of the impact crusher is controlled below 30 mm. An impact crusher, rather than a hammer crusher, is chosen as the medium crushing equipment because the impact crusher applies impact force to the material through high-speed hammers and achieves repeated impacts with the impact plates, resulting in cubic particles with abundant angularity and a low content of needle-like and flaky particles, making it more suitable for subsequent use as aggregate in road construction materials. The crushed steel slag is then screened through four layers of vibrating screens to obtain four grades of steel slag aggregate with particle sizes of 19 mm to 26.5 mm, 9.5 mm to 19 mm, 4.75 mm to 9.5 mm, and 0.075 mm to 4.75 mm. These four particle size ranges are consistent with the current classification method for road engineering aggregates, facilitating direct alignment with the gradation design requirements of asphalt mixtures and cement-stabilized crushed stone. In some alternative implementations, more particle size grades can be added according to specific engineering needs, or the maximum particle size limit can be adjusted to 31.5 mm.

[0025] Subsequently, the four grades of steel slag aggregate were subjected to carbonization curing treatment. Each grade of steel slag aggregate was spread evenly on a breathable mesh tray inside a sealed carbonization curing chamber, with a thickness controlled at 150 mm. The thickness of the spread directly affects the carbonization curing effect; too thick a layer will prevent the bottom aggregate from fully contacting the carbon dioxide, while too thin a layer will reduce the throughput per batch. 150 mm represents a compromise between carbonization uniformity and processing efficiency. A mixture of carbon dioxide and nitrogen with a volume concentration of 20% was continuously introduced into the sealed carbonization curing chamber, maintaining a temperature of 60 degrees Celsius and a relative humidity of 70%, for a continuous carbonization curing period of 7 days. Carbonization curing at 60 degrees Celsius, rather than room temperature, was chosen because the diffusion rate and reaction kinetic constant of carbon dioxide molecules are significantly higher at this temperature than at room temperature, accelerating the penetration and reaction process of carbon dioxide into the steel slag particles. The consideration for controlling the relative humidity at 70% is that the carbonation reaction requires a certain amount of water as a reaction medium to dissolve carbon dioxide into carbonic acid, which then reacts with free calcium oxide and free magnesium oxide. However, excessive humidity can cause a continuous water film to form on the particle surface, blocking the gas diffusion channels. A relative humidity of 70% achieves a balance between water film thickness and gas permeability. During the carbonation curing process, free calcium oxide in the steel slag aggregate reacts with carbon dioxide to form calcium carbonate, and free magnesium oxide reacts with carbon dioxide to form magnesium carbonate. These two products are volume stable and no longer exhibit the characteristic of swelling upon contact with water, thus fundamentally eliminating the expansion potential energy of the steel slag. In some optional embodiments, the volume concentration of carbon dioxide can be adjusted to other values ​​between 10% and 30%, the carbonation curing temperature can be set between 40°C and 80°C, the relative humidity can be set between 50% and 80%, and the carbonation curing time can be extended to 14 days.

[0026] After carbonization curing, the expansion strain field acquisition stage begins. Quarter samples are taken from each grade of steel slag aggregate using the quartering method. Each sample is then dried at 105°C to constant weight to eliminate residual moisture interference with subsequent cementing and autoclaving processes. After drying, each grade of steel slag aggregate is prepared into standard cylindrical specimens with a diameter of 50 mm and a height of 100 mm using aluminum phosphate inorganic binder. The aluminum phosphate inorganic binder is incorporated at a rate of 8% to 12% of the dry weight of the steel slag aggregate. This dosage range provides sufficient cementing strength to maintain the specimen's integrity during handling and installation without significantly altering the expansion behavior of the aggregate system due to excessive binder content. The core reason for choosing aluminum phosphate inorganic binder instead of conventional epoxy resin organic binder is that the subsequent autoclaving test reaches a temperature of 216 degrees Celsius. Conventional epoxy resin undergoes thermal decomposition and softening at this temperature, resulting in a non-negligible volume change. This leads to interference from the binder's thermal deformation in the strain field data acquired by digital image correlation, failing to accurately reflect the expansion characteristics of the steel slag aggregate. After high-temperature curing, the aluminum phosphate inorganic binder forms a ceramicized phosphate glass phase network structure. Its coefficient of thermal expansion is on the same order of magnitude as that of steel slag, and it undergoes almost no additional volume change at 216 degrees Celsius. Therefore, it ensures that the acquired strain field data faithfully reflects the expansion behavior of the steel slag aggregate. After preparation, the standard cylindrical specimens are cured at 200 degrees Celsius for 2 hours to fully ceramicize the aluminum phosphate inorganic binder. In some alternative embodiments, magnesium phosphate inorganic binder or water glass inorganic binder can be used instead of aluminum phosphate inorganic binder, and the curing temperature can be adjusted between 150 and 250 degrees Celsius depending on the type of binder.

[0027] Next, a high-temperature resistant speckle pattern was sprayed onto the side surface of a standard cylindrical specimen for tracking surface displacement using digital image correlation (DIR). The spraying method involved first applying a layer of white high-temperature resistant inorganic primer as a uniform background, followed by spraying black high-temperature resistant inorganic speckle paint after the primer had completely dried, forming a randomly distributed speckle array. The average diameter of the speckles was controlled between 0.3 mm and 0.5 mm, and the speckle coverage was controlled between 40% and 60% of the side surface area of ​​the standard cylindrical specimen. The selection of speckle size and coverage needed to meet the speckle quality requirements of DIR. Excessively large speckles would reduce the spatial resolution of the displacement field, while excessively small speckles would be difficult to distinguish clearly in camera imaging. A diameter range of 0.3 mm to 0.5 mm corresponds to an imaging size of approximately 3 to 5 pixels at the typical working distance of a 5-megapixel industrial area array camera, ensuring reliable identification without excessively blurring the displacement gradient information. When the coverage is below 30%, the grayscale distribution of the speckle pattern is too sparse, resulting in a decrease in the signal-to-noise ratio of cross-correlation calculations. When it is above 70%, the large-area overlap between speckles leads to reduced independent recognizability. 40% to 60% is a widely validated optimal range. High-temperature resistant inorganic paint was chosen instead of ordinary acrylic paint because the 216°C autoclaving environment would cause organic paint films to bubble, peel, or discolor, causing the speckle pattern to fail during testing.

[0028] refer to Figure 1 ,like Figure 1As shown, the autoclave is the core test chamber of the entire acquisition system. It creates a high-temperature, high-pressure steam environment at 216 degrees Celsius and 2.0 MPa, used for accelerated expansion tests on standard cylindrical specimens of steel slag aggregate. The standard cylindrical specimen has a diameter of 50 mm and a height of 100 mm, with random speckle patterns covering its side surface. It is placed vertically on a specimen holder inside the autoclave. The specimen holder is located at the bottom of the autoclave, serving to fix and support the standard cylindrical specimen, preventing displacement or tipping due to steam impact during autoclaving. A high-temperature resistant optical observation window is installed on the side wall of the autoclave at the location of the standard cylindrical specimen. The inner surface of the high-temperature resistant optical observation window is coated with an anti-fog coating to suppress the condensation of high-temperature steam on the inner surface of the window, which would affect image clarity. Two industrial area array cameras, each with a resolution of at least 5 megapixels, are mounted on the outside of the high-temperature resistant optical observation window. The angle between their optical axes is set to 25 to 35 degrees, forming a binocular stereo vision acquisition system. The two cameras simultaneously acquire speckle images of the side surface of a standard cylindrical specimen from different perspectives. After stereo matching, the three-dimensional displacement field of the specimen surface can be reconstructed, distinguishing between radial and axial displacements—something a single camera cannot achieve by acquiring only two-dimensional projected displacement. A narrow-band blue LED surface light source with an emission wavelength of 450 nanometers, located in the short-wavelength visible light band, is also configured on the outside of the high-temperature resistant optical observation window. This light is spectrally significantly separated from the infrared thermal radiation generated by the high-temperature environment inside the autoclave. Combined with a narrow-band filter at the camera end, it effectively filters out background light interference from thermal radiation, ensuring stable image quality of the speckle image throughout the autoclaving process. Illumination light emitted from a narrow-band blue LED surface light source passes through a high-temperature resistant optical observation window and illuminates the speckle pattern on the side surface of a standard cylindrical specimen. The reflected light from the speckle pattern then returns through the high-temperature resistant optical observation window to two industrial area array cameras for imaging. Throughout the autoclaving process, a binocular stereo vision acquisition system operates continuously at a frequency of once every 30 seconds, acquiring approximately 360 speckle image pairs over the 3-hour autoclaving duration. This constitutes a complete time-series acquisition covering the entire accelerated expansion test process, providing raw image data for subsequent digital image correlation calculations.

[0029] The standard cylindrical specimen was then placed vertically on the specimen holder inside the autoclave. A high-temperature resistant optical observation window was installed on the side wall of the autoclave at the position corresponding to the standard cylindrical specimen. The inner surface of this window was coated with an anti-fog coating. The anti-fog coating was used to prevent the condensation of high-temperature steam on the inner surface of the observation window during the autoclaving process, as water mist would scatter the incident light and severely reduce the contrast and clarity of the speckle image, directly leading to a significant decrease or even complete failure of the matching accuracy of subsequent digital image correlation calculations. Two industrial area array cameras with a resolution of no less than 5 megapixels were mounted outside the high-temperature resistant optical observation window. The optical axis intersection angle of the two industrial area array cameras was set to 25 to 35 degrees, forming a binocular stereo vision acquisition system. The reason for using a binocular stereo vision acquisition system instead of a single camera is that the standard cylindrical specimen is a three-dimensional curved surface. A single camera can only acquire two-dimensional projected displacement and cannot distinguish between in-plane and out-of-plane displacement. The binocular stereo vision acquisition system uses two cameras to simultaneously image from different angles. After stereo matching, it can reconstruct the three-dimensional displacement field of the surface, thus completely capturing the radial and axial displacement components of the standard cylindrical specimen during the autoclaving expansion process. The selection of the optical axis intersection angle follows a trade-off between measurement accuracy and matching stability. If the angle is too small, the parallax change between the two cameras will be insensitive, resulting in a decrease in out-of-plane displacement resolution. If the angle is too large, the perspective distortion difference of the speckle pattern in the left and right images will increase, resulting in a decrease in the matching success rate. 25 degrees to 35 degrees is a commonly used angle range in binocular stereo digital image correlation. A narrowband blue LED surface light source with a wavelength of 450 nm is also configured on the outside of the high-temperature optical observation window as the illumination source. Narrowband blue light illumination is chosen to suppress the interference of infrared thermal radiation background light generated by the high-temperature environment inside the autoclave on the speckle image. This is because the 450 nm wavelength is in the short-wavelength visible light band, which is significantly separated spectrally from the infrared band where thermal radiation from the high-temperature body is concentrated. By adding a narrowband filter matching the illumination wavelength in front of the camera lens, the thermal radiation background light can be effectively filtered out, retaining only the speckle reflection light, thus ensuring that the speckle image maintains stable contrast throughout the autoclaving process. In some optional embodiments, the illumination wavelength can also be selected as 532 nm green light or other narrowband wavelengths separated from the infrared band.

[0030] The autoclave was started and autoclaved for 3 hours at a temperature of 216 degrees Celsius and a pressure of 2.0 MPa. Throughout the autoclaving process, a binocular stereoscopic vision acquisition system synchronously acquired left and right speckle image pairs on the side surface of the standard cylindrical specimen at a frequency of once every 30 seconds. The 3-hour autoclaving period generated approximately 360 sets of left and right speckle image pairs, forming a complete time acquisition sequence.

[0031] After autoclaving, digital image correlation was performed on all left and right speckle image pairs to extract expansion strain field information. All left and right speckle image pairs were numbered sequentially according to the acquisition time. The first group was used as the reference image pair, which corresponds to the initial state of the standard cylindrical specimen before significant expansion during the initial stage of autoclaving. The second group and subsequent groups were used as deformation image pairs. Each deformation image pair recorded the change of the speckle pattern on the surface of the standard cylindrical specimen relative to the initial state at the corresponding time.

[0032] The specific method for performing digital image correlation (DIR) operations on each pair of deformed images relative to the reference image pair is as follows: The reference image is divided into square sub-regions with a side length of 31 pixels, and the center-to-center distance between adjacent square sub-regions is set to 7 pixels. The 31-pixel side length ensures that each sub-region contains a sufficient number of speckles to provide unique gray-level distribution characteristics, and the 7-pixel center-to-center distance means that there is a large area of ​​overlap between adjacent sub-regions. This overlap strategy can improve the spatial sampling density of the displacement field without increasing the size of the sub-regions. For each square sub-region, the search starts from the corresponding position of the square sub-region in the reference image in the deformed image, and the search is performed pixel by pixel within the search area. The zero-mean normalized cross-correlation criterion is used to calculate the correlation coefficient between the reference sub-region and the candidate sub-region. The calculation process of the zero-mean normalized cross-correlation criterion is as follows: the gray values ​​of the reference sub-region and the candidate sub-region are respectively subtracted from their respective gray-level mean values ​​to eliminate the influence of average brightness differences, and then the sum of the pixel-by-pixel products of the two is calculated and divided by the square root of the product of the sum of the squares of their respective gray-level deviations to normalize to the nearest square. The interval, in which This indicates a perfect match. After the pixel-by-pixel sliding search is completed, the position where the correlation coefficient reaches its maximum value is determined as the integer-pixel matching position of the square sub-region. To further improve the displacement measurement accuracy, double cubic B-spline interpolation is used in the neighborhood of the integer-pixel matching position to perform sub-pixel-level reconstruction of the grayscale field of the deformed image. Then, the search is refined with a sub-pixel step size on the continuous grayscale field to achieve sub-pixel accuracy displacement vector solution, with typical sub-pixel accuracy reaching 0.01 to 0.05 pixels. The coordinate difference between the center coordinates of each square sub-region in the reference image and the corresponding sub-pixel matching coordinates in the deformed image is obtained, which is the 2D displacement vector of the square sub-region. The horizontal component corresponds to the axial displacement of the standard cylindrical specimen, and the vertical component corresponds to the radial displacement. The 2D displacement vectors of all square sub-regions on the side surface of the standard cylindrical specimen are arranged according to their spatial positions to form a full-field displacement vector field.

[0033] refer to Figure 2 ,like Figure 2 As shown, the left side is the reference image, corresponding to the start time of the autoclave test. The first image shows a speckle pattern on the side surface of a standard cylindrical specimen. At this point, the specimen has not yet undergone significant expansion, and the speckle pattern is in its initial, undeformed state. The second image shows a deformed pattern, corresponding to a specific moment during the autoclaving process. The collected speckle image shows the specimen having undergone expansion and deformation under high-temperature and high-pressure steam. Comparing the reference image and the deformed image reveals that the speckle pattern in the deformed image has undergone overall stretching and displacement relative to the reference image. This overall stretching originates from the uniform thermal expansion of the steel slag aggregate and the hydration expansion of free oxides under autoclaving conditions. Simultaneously, two localized areas with significantly concentrated speckle pattern distortion exist in the deformed image, marked with dashed circles. These areas of high local deformation correspond to locations within the standard cylindrical specimen with higher levels of free calcium oxide or free magnesium oxide. The expansion reaction is more intense at these locations, resulting in significantly greater local surface displacement than the surrounding areas. A square sub-region is marked in the reference image, with a side length corresponding to 31 pixels. The center of the sub-region is marked with a cross and the pixel coordinates are indicated. The core computational process of the digital image correlation method involves searching the deformed image for the region with the most similar grayscale distribution to each square sub-region in the reference image. The search process uses a zero-mean normalized cross-correlation criterion to calculate the correlation coefficient pixel-by-pixel, and then refines the matching accuracy to the sub-pixel level through bicubic B-spline interpolation. In the deformed image, the target position of the square sub-region determined after matching is marked with a dashed box. The center of the dashed box is also marked with a cross and the matching coordinates with sub-pixel precision are noted. The coordinate difference between the center of the sub-region in the reference image and the matching position in the deformed image constitutes the displacement vector of the sub-region. The direction and magnitude of this displacement vector are indicated by arrows in the figure. After performing the above matching operation on all square sub-regions on the side surface of the standard cylindrical specimen, a full-field displacement vector field covering the entire side surface can be obtained. Then, by taking the difference between the displacement vectors of adjacent sub-regions, the radial expansion strain field and the axial expansion strain field are obtained, providing basic data for the subsequent construction of the multi-channel input tensor of the convolutional neural network.

[0034] Based on the full-field displacement vector field, the strain field is further solved. The radial expansion strain components are obtained by taking the difference between the displacement vectors of adjacent square sub-regions and dividing by the center-to-center distance. and axial expansion strain components ,in It represents the elongation per unit length in the radial direction of a standard cylindrical specimen. This represents the elongation per unit length in the axial direction of a standard cylindrical specimen. and The numerical values ​​at various spatial locations on the side surface of the standard cylindrical specimen are linearly mapped to grayscale images with grayscale values ​​ranging from 0 to 255, where grayscale value 0 corresponds to the minimum value of the strain component and grayscale value 255 corresponds to the maximum value of the strain component. One grayscale image of the radial expansion strain field and one grayscale image of the axial expansion strain field are generated for each acquisition moment. The grayscale images of the radial and axial expansion strain fields generated at all acquisition moments are arranged in chronological order to obtain a time series of grayscale images of the expansion strain field covering the entire autoclaving process.

[0035] This grayscale image time series of the expansion strain field not only records the overall expansion of the standard cylindrical specimen during the autoclaving process, but more importantly, it preserves the spatial distribution characteristics and temporal evolution of the expansion in the form of a full-field two-dimensional image. For example, local areas with high free calcium oxide content appear as concentrated bright spots with high grayscale values ​​in the grayscale image of the expansion strain field, while areas with sufficient carbonization curing appear as uniform dark areas with low grayscale values. This spatial distribution pattern and temporal evolution pattern contains richer information on the stability state than the traditional single autoclaving expansion rate value, providing a high-dimensional input data foundation for subsequent intelligent determination of stability level using convolutional neural networks.

[0036] In this embodiment, the second step of the method for resource utilization and quality grading of road steel slag aggregate involves the evaluation of carbonization curing effect and determination of stability level based on convolutional neural network, as well as the supplementary carbonization curing closed-loop control driven by the determination result. The specific implementation process of this step is described in detail below.

[0037] The grayscale image time series of the expansion strain field obtained in Step 1 contains approximately 360 radial expansion strain field grayscale images and approximately 360 axial expansion strain field grayscale images. Directly feeding all images into the classification model would incur significant computational overhead and introduce a large amount of redundant information, as the variation in the strain field within adjacent 30-second intervals is extremely small. Therefore, images at equal intervals of 6 time points are extracted from the grayscale image time series of the expansion strain field. Specifically, the total number of frames acquired throughout the entire autoclaving process is evenly divided into 6 equally spaced time nodes, and one radial expansion strain field grayscale image and one axial expansion strain field grayscale image are taken at each time node, resulting in a total of 12 images. The reason for extracting 6 time points, rather than more or fewer, is that 6 equally spaced time points are sufficient to cover the typical evolutionary stages of autoclaving expansion, including the initial heating and thermal deformation stage, the intermediate rapid hydration and expansion stage of free oxides, and the later stage where expansion tends to stabilize. Furthermore, taking two images at each time point, one radial and one axial, can fully characterize the differences in expansion behavior in two orthogonal directions. In some alternative implementations, the number of extraction times can also be set to 4, 8 or 10, and the number of multichannel input tensor channels will vary accordingly.

[0038] The 12 extracted images were uniformly scaled to a spatial size of 224 pixels by 224 pixels. This size was chosen because it aligns with the typical input size of mainstream convolutional neural network architectures. Bilinear interpolation was used during scaling to avoid introducing jagged artifacts. After scaling, the 12 images were stacked sequentially along the channel dimension in chronological order, forming a multi-channel input tensor with a spatial size of 224 pixels by 224 and 12 channels. In this multi-channel input tensor, the first two channels correspond to the radial and axial expansion strain field grayscale images at the first extraction time, the third and fourth channels correspond to the radial and axial expansion strain field grayscale images at the second extraction time, and so on, with the 11th and 12th channels corresponding to the radial and axial expansion strain field grayscale images at the sixth extraction time. This stacking method enables convolutional neural networks to simultaneously perceive strain changes at the same spatial location at different times and in different directions through cross-channel convolutional kernels during convolution operations, thereby jointly extracting spatial distribution patterns and temporal evolution patterns in a single forward propagation.

[0039] Multi-channel input tensors are fed into a pre-trained convolutional neural network (CNN) model for stability level determination. This CNN model consists of four convolutional modules and two fully connected layers connected in series. The first convolutional module contains 32 convolutional layers with kernels of size 7x7, stride 2, and padding 3, followed by batch normalization layers, ReLU activation functions, and max-pooling layers of size 3x3 with stride 2. The use of large 7x7 kernels in the first convolutional module is to capture large-scale spatial distribution features in the dilatational strain field, such as the overall dilatational gradient direction and contour information of large high-strain regions, with a large receptive field at the front of the network. The stride is set to 2, and the max-pooling layer rapidly reduces the spatial size of the feature map, reducing the computational cost of subsequent layers. The second convolutional module contains 64 convolutional layers with kernels of size 5x5, stride 1, and padding 2, followed by batch normalization layers, ReLU activation functions, and max-pooling layers of size 3x3 with stride 2. The second convolutional module doubles the number of channels to 64, using medium-sized 5x5 convolutional kernels to extract more refined features from locally dilated concentrated regions. The third convolutional module contains 128 convolutional layers consisting of 3x3 kernels with a stride of 1 and padding of 1, followed by batch normalization layers and ReLU activation functions. The fourth convolutional module also contains 128 convolutional layers consisting of 3x3 kernels with a stride of 1 and padding of 1, followed by batch normalization layers, ReLU activation functions, and global average pooling layers. The third and fourth convolutional modules use small 3x3 kernels for deep feature extraction. The continuously stacked small kernels expand the effective receptive field by layer stacking while maintaining a small number of parameters, enabling the capture of complex dilated textures and local anomaly patterns. The global average pooling layer calculates the arithmetic mean of the gray values ​​at all spatial locations for each feature map output by the fourth convolutional module, compressing each channel into a scalar, thus outputting a 128-dimensional feature vector. The advantage of global average pooling over fully connected flattening is that it is not affected by small spatial offsets in the input image, and the number of parameters is zero, so it does not introduce additional overfitting risk.

[0040] The 128-dimensional feature vector is mapped to a 64-dimensional vector through the first fully connected layer, then nonlinearly transformed using the ReLU activation function, and finally mapped to a 3-dimensional vector through the second fully connected layer. The three components of the 3-dimensional vector are processed by the Softmax function, outputting the probability values ​​for each of the three stability levels: excellent, moderate, and poor. The Softmax function converts the three components into a probability distribution with values ​​ranging from 0 to 1, and the sum of these values ​​is always equal to 1. A higher probability value indicates a stronger confidence level in the model's determination of that level.

[0041] This convolutional neural network model requires supervised training to acquire the aforementioned decision-making ability. The training method is as follows: Collect no fewer than 500 sets of steel slag aggregate samples from different heats of different steel plants to ensure that the training data covers the widest possible range of variations in the chemical composition and phase structure of steel slag. For each set of steel slag aggregate samples, prepare standard cylindrical specimens according to the method in step 1, perform autoclaving tests, complete digital image correlation (DIR) calculations, and generate multi-channel input tensors. At the same time, take parallel samples for each set of steel slag aggregate samples and independently determine the autoclaving expansion rate according to the autoclaving method specified in the national standard. It is emphasized here that parallel samples are taken for independent measurement rather than directly converting the expansion rate from the DIR specimens. The purpose is to ensure that the source of the training labels is completely independent of the model input features and avoid homologous structural bias between the labels and the input. Autoclaving expansion rates less than 0.5% are labeled as excellent stability, autoclaving expansion rates between 0.5% and 2.0% are labeled as moderate stability, and autoclaving expansion rates greater than 2.0% are labeled as poor stability. The above threshold settings refer to the volume stability requirements in the current standards for steel slag used in roads. A content below 0.5% can be directly used in scenarios with the highest stability requirements, such as asphalt pavement surface layers. Content between 0.5% and 2.0% can be used in base courses after appropriate measures are taken. Content above 2.0% poses a significant risk of expansion. In some optional implementations, the threshold can also be adjusted according to specific engineering standards; for example, the boundary between excellent and medium grades can be set at 1.0%.

[0042] The training set consists of all multi-channel input tensors and their corresponding stability level labels, with 20% randomly selected as the validation set. During training, the cross-entropy loss function is used to measure the difference between the model's predicted probability distribution and the actual labels. Cross-entropy loss function. The calculation method is as follows ,in This is a stability level number, with values ​​from 1 to 3 corresponding to excellent stability, moderate stability, and poor stability, respectively. The one-hot encoded value of the actual annotation, that is, if the actual level is the first Class otherwise , The first output of the Softmax function The network parameters are updated using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs. After each training epoch, the classification accuracy is calculated on the validation set. After all training epochs are completed, the model parameters corresponding to the epoch with the highest classification accuracy on the validation set are saved as the final model. In some optional implementations, the learning rate can be set to other values ​​between 0.0005 and 0.005, and the number of training epochs can be adjusted from 50 to 200 epochs depending on the convergence of the validation set accuracy. A learning rate decay strategy or early stopping mechanism can also be introduced to prevent overfitting.

[0043] When using the trained convolutional neural network model to determine the stability level of steel slag aggregate in different batches, the probability values ​​of the three levels output by the Softmax function are further converted into a single continuous stability score to facilitate the quantitative calculation of subsequent comprehensive quality assessment. (Continuous stability score) The calculation method is as follows ,in This represents the probability value for superior stability. This is a probability value with moderate stability. This represents the probability value for a grade of poor stability. This calculation method assigns the highest coefficient (1.0) to excellent stability, the next highest (0.6) to moderate stability, and the lowest (0.2) to poor stability. This ensures that the continuous stability score not only reflects the final grade category but also includes the model's confidence distribution for each grade. For example, a sample with a probability of 0.7 for excellent stability, 0.2 for moderate stability, and 0.1 for poor stability has a continuous stability score of 0.84. Conversely, a sample with a probability of 0.5 for excellent stability, 0.3 for moderate stability, and 0.2 for poor stability has a continuous stability score of 0.72. Although the latter also has the highest probability of excellent stability, its lower model confidence results in a lower continuous stability score. This continuous scoring method conveys more information about the stability state than simply taking the discrete grade label corresponding to the highest probability value. In some alternative implementations, the coefficients corresponding to the three levels can also be adjusted to other combinations of values ​​based on engineering experience, as long as the monotonically decreasing relationship of the coefficient for superior grade being greater than that for medium grade and the coefficient for medium grade being greater than that for poor grade is maintained.

[0044] A closed-loop feedback control system for carbonation curing effect is implemented based on the continuous stability score. The threshold for the continuous stability score is set at 0.5. When the continuous stability score of a certain grade of steel slag aggregate is below 0.5, it indicates that the aggregate still retains a significant amount of unconverted free calcium oxide or free magnesium oxide after carbonation curing, and its expansion potential energy has not been fully eliminated. Direct application to the pavement surface or base course poses a stability risk. In this case, the entire grade of steel slag aggregate is reintroduced into a sealed carbonation curing chamber and subjected to supplementary carbonation curing for 7 days under the same conditions as in step 1. After the supplementary carbonation curing is completed, standard cylindrical specimens are prepared from samples of the steel slag aggregate, and all operations of step 1 (autoclave expansion strain field acquisition) and step 2 (convolutional neural network stability evaluation) are repeated, and the continuous stability score is recalculated. The design intent of this closed-loop feedback mechanism is that the initial free oxide content of steel slag aggregates from different sources and furnace batches varies greatly. A single carbonization curing may not achieve sufficient stabilization for all batches. By repeating carbonization curing and evaluation operations, additional treatment can be carried out on batches that are not sufficiently carbonized until their stability reaches an acceptable level.

[0045] The cumulative number of supplementary carbonization curing treatments is capped at three. This cap is set for two reasons. First, the carbonization reaction progresses layer by layer from the particle surface to the interior. As the carbonization depth increases, the diffusion path of carbon dioxide into the particle core lengthens, and the reaction rate decreases marginally with each cycle. After three supplementary carbonization curing treatments, the improvement in stability from further carbonization is very limited. Second, excessive carbonization curing cycles significantly increase production time and costs, contradicting the economic goals of steel slag resource utilization. If a grade of steel slag aggregate still has a stability score below 0.5 after three supplementary carbonization curing treatments, it is considered to have inherently high-expansion-risk components that cannot be fully stabilized through carbonization curing. This grade of steel slag aggregate is directly classified into the lowest quality grade, Grade 3 aggregate, and subsequently used in roadbed filling layers with the lowest stability requirements. In some alternative implementations, the maximum number of supplementary carbonization curing cycles can be set to 2 or 4 times depending on the actual production line conditions, and the threshold for the continuous stability score can be adjusted to other values ​​between 0.4 and 0.6 according to engineering standards.

[0046] After the above operations, each grade of steel slag aggregate either enters the subsequent comprehensive quality assessment stage with a continuous stability score of 0.5 or higher, or is directly classified as Grade 3 aggregate because its continuous stability score is consistently below 0.5. The continuous stability score, a continuous value between 0.2 and 1.0, is retained and subsequently used as a volumetric stability index in the analytic hierarchy process (AHP) grading framework to directly participate in the weighted calculation of the comprehensive quality assessment score, thereby organically integrating the expansion pattern recognition results extracted by the convolutional neural network model with traditional mechanical durability indicators.

[0047] refer to Figure 3 ,like Figure 3 As shown, the horizontal axis represents the number of carbonization treatments, from left to right: first carbonization treatment, one supplementary carbonization treatment, two supplementary carbonization treatments, and three supplementary carbonization treatments. The vertical axis represents the continuous stability score. The graph shows a horizontal straight line, corresponding to a threshold of 0.5. Only steel slag aggregates with a continuous stability score of 0.5 or higher can proceed to the subsequent comprehensive quality assessment stage. The graph also shows four broken lines, corresponding to four different particle size ranges of steel slag aggregates. Grade A steel slag aggregates have a particle size range of 19 mm to 26.5 mm. Their continuous stability score reaches 0.85 after the first carbonization curing, far exceeding the threshold, indicating that the free oxides in this grade of coarse-sized aggregates have been fully converted after carbonization curing, resulting in excellent stability, and no further carbonization curing is required. Grade B steel slag aggregates have a particle size range of 9.5 mm to 19 mm. Their continuous stability score after the first carbonization curing is 0.42, below the threshold. After one additional carbonization curing, the continuous stability score rises to 0.67, exceeding the threshold and entering the qualified range. Grade C steel slag aggregate has a particle size range of 4.75 mm to 9.5 mm. Its continuous stability score after the first carbonization curing is only 0.35. After one additional carbonization curing, it rises to 0.44, but still fails to meet the standard. After two additional carbonization curings, it rises to 0.58, exceeding the set threshold. Grade D steel slag aggregate has a particle size range of 0.075 mm to 4.75 mm. Its continuous stability score gradually increases from 0.28 initially to 0.46 after three additional carbonization curings. Although each additional carbonization curing brings a certain degree of improvement, even after reaching the maximum set number of three carbonization curings, it still fails to reach the set threshold of 0.5. Therefore, this grade of steel slag aggregate is directly classified as Grade 3 aggregate for use in roadbed filling layers. Overall, the smaller the particle size of the steel slag aggregate, the slower the rate of improvement in the stability continuous score during the carbonization curing process. This is because although fine particles have a large specific surface area and a sufficient carbonization reaction interface, the accumulation between fine particles in the carbonization curing box is more compact, making it difficult for carbon dioxide gas to penetrate evenly into the depth of the material layer, thus limiting the carbonization efficiency.

[0048] In this embodiment, the third step of the method for resource utilization and quality grading of road steel slag aggregates involves comprehensive quality assessment and directional preparation of road building materials based on the hierarchical analysis framework. The specific implementation process of this step will be described in detail below.

[0049] After the stability evaluation and closed-loop carbonization curing control in step 2, steel slag aggregates with a continuous stability score of 0.5 or higher have basic volume stability assurance. However, volume stability is only one of the necessary conditions for steel slag aggregates to be used in road engineering. Actual road use scenarios also require the aggregates to meet corresponding indicators in terms of mechanical strength, abrasion resistance, particle morphology, and water stability. Therefore, it is necessary to further introduce mechanical durability indicators based on the stability evaluation, and incorporate both types of indicators into a unified evaluation framework for comprehensive scoring and grading.

[0050] For steel slag aggregates with a stability continuous score of 0.5 or higher, the following five mechanical durability indicators were determined according to national standard test methods: crushing value, Los Angeles abrasion value, polishing value, needle-like and flaky particle content, and water absorption rate. The crushing value reflects the aggregate's resistance to crushing under static pressure; a lower crushing value indicates higher aggregate strength. The Los Angeles abrasion value reflects the aggregate's resistance to wear under combined rolling friction and impact; a lower Los Angeles abrasion value indicates better wear resistance. The polishing value reflects the aggregate's ability to maintain micro-roughness after repeated tire rolling, and is directly related to the skid resistance of asphalt pavement layers; a higher polishing value indicates that the aggregate can maintain a good skid resistance contribution under long-term traffic conditions. The polishing value is introduced because the Los Angeles abrasion value alone is insufficient to comprehensively evaluate the surface service performance of aggregates. Existing engineering practice shows that while some aggregates may have high Los Angeles abrasion values, their skid resistance in the field remains acceptable. Conversely, some aggregates may have acceptable Los Angeles abrasion values ​​but low polishing values, leading to a rapid decline in skid resistance after the road is opened to traffic. The two indicators are complementary. The content of needle-shaped and flaky particles reflects the proportion of flat and elongated particles in the aggregate. These particles are prone to breakage along weak points under compaction and load, which is detrimental to the interlocking stability and uniformity of the void structure of the mixture. Water absorption reflects the aggregate's ability to absorb water. Excessive water absorption can lead to ice expansion stress within the aggregate during freeze-thaw cycles, causing deterioration. Furthermore, excessive water absorption consumes the effective asphalt film thickness, reducing the adhesion strength between asphalt and aggregate. These five indicators comprehensively cover the road suitability of steel slag aggregates from four dimensions: mechanical strength, surface durability, particle morphology, and water stability. In some alternative implementations, indicators such as apparent density, strength, or adhesion to asphalt can be added according to specific engineering requirements.

[0051] After obtaining the continuous stability score and the measured values ​​of five mechanical durability indicators, a hierarchical framework based on the Analytic Hierarchy Process (AHP) was constructed for comprehensive quality assessment. The core idea of ​​AHP is to decompose the complex multi-indicator comprehensive evaluation problem into several levels, determine the relative importance of each indicator by comparing each indicator pairwise at each level, and finally obtain a set of contribution values ​​reflecting decision preferences for weighted calculation.

[0052] The analytic hierarchy process (AHP) grading framework consists of three layers. The target layer represents the comprehensive quality grade of road-grade steel slag aggregate, i.e., the final grading result to be determined. The criteria layer sets two criteria: volumetric stability and mechanical durability, representing two major dimensions affecting the road-grade quality of steel slag aggregate. In the scheme layer, the volumetric stability index uses the continuous stability score output from step 2 as the sole indicator, while the mechanical durability index includes five measured values: crushing value, Los Angeles abrasion value, polishing value, needle-like and flaky particle content, and water absorption rate.

[0053] First, a second-order judgment matrix is ​​constructed at the criterion level. The Saaty 1-9 scale is used to pairwise compare and assign values ​​to the relative importance of volumetric stability and mechanical durability indices. The Saaty scale specifies that a value of 1 is assigned when both criteria are equally important, 3 when the former is slightly more important than the latter, 5 when significantly important, 7 when strongly important, and 9 when extremely important. 2, 4, 6, and 8 are compromise values ​​between adjacent odd-numbered scales. In this embodiment, considering that volumetric stability is the most critical differentiating factor between steel slag and natural aggregates, steel slag with unqualified stability cannot be safely used in road engineering even with excellent mechanical properties. Therefore, a group of experts in road engineering and metallurgical solid waste assigned a value of 3 to the volumetric stability index relative to the mechanical durability index, meaning that the volumetric stability index is slightly more important than the mechanical durability index. In the resulting second-order judgment matrix, the element in the first row and first column is 1, the element in the first row and second column is 3, the element in the second row and first column is its reciprocal (1 divided by 3), and the element in the second row and second column is 1. Calculating the normalized eigenvector corresponding to the largest eigenvalue of this second-order judgment matrix yields a criterion layer contribution value of 0.75 for volume stability and 0.25 for mechanical durability. Since the second-order judgment matrix inherently satisfies the consistency requirement, no consistency check is required. In some alternative implementations, the expert group can also adjust the assignment to 2 or 5, and the corresponding criterion layer contribution values ​​will change accordingly.

[0054] Then, a 5th-order judgment matrix was constructed at the scheme layer under the mechanical durability index. The Saaty 1-9 scaling method was used to pairwise compare and assign values ​​to the relative importance of the five indicators: crushing value, Los Angeles abrasion value, polishing value, flaky particle content, and water absorption rate. In this embodiment, the expert group assigned values ​​based on the degree of influence of each performance indicator on pavement performance under actual service conditions of road engineering. For example, a value of 3 was assigned to the crushing value relative to the flaky particle content, indicating slightly important, and a value of 2 was assigned to the Los Angeles abrasion value relative to the water absorption rate, indicating slightly important. The largest eigenvalue of this 5th-order judgment matrix was calculated. The corresponding normalized feature vector, where The largest eigenvalue of the 5th-order judgment matrix is ​​represented by the eigenvector. The five components of the normalized eigenvector are the contributions of crushing value, Los Angeles abrasion value, polishing value, needle-like particle content, and water absorption rate to the scheme layer, respectively. A consistency check is performed on this 5th-order judgment matrix, and a consistency index is calculated. , equal The quotient obtained by subtracting the matrix order 5 and then dividing by the matrix order 5 minus 1 is: The largest eigenvalue of the 5th-order judgment matrix. Consistency ratio. Equal to consistency index Divide by the average random consistency index corresponding to the 5th order judgment matrix The resulting quotient, of which the 5th order matrix The value is 1.12. When the consistency ratio... If the value is less than 0.1, the 5th-order judgment matrix is ​​considered to have passed the consistency test, indicating that the pairwise comparison assignments by the expert group have acceptable logical consistency. If the consistency ratio... If the value is not less than 0.1, the expert panel needs to re-examine and adjust the pairwise comparison assignments until the test is passed.

[0055] After obtaining all contribution values ​​for the criterion layer and the scheme layer, it is necessary to uniformly convert the continuous stability score and the measured values ​​of the five mechanical durability indicators to a comparable dimension range of 0 to 1.0. The continuous stability score, which is already a continuous value between 0.2 and 1.0 in step 2, naturally possesses the positive characteristic that "the larger the value, the better the stability," and therefore requires no additional conversion. It can be directly used as the evaluation value of the volumetric stability indicator in subsequent weighted calculations.

[0056] Of the five mechanical durability indicators, crushing value, Los Angeles abrasion value, needle-like particle content, and water absorption rate are inverse indicators, meaning that the lower the measured value, the better the quality. Therefore, a forward conversion is required. The forward conversion method is to subtract the upper limit specified in the corresponding national standard from the measured value, take the absolute value, divide it again by the upper limit specified in the corresponding national standard, and then subtract the quotient from 1.0 to obtain the forward value. Taking the crushing value as an example, if the measured crushing value of a certain grade of steel slag aggregate is 18%, and the upper limit specified in the national standard is 26%, then first calculate the absolute value of 18% minus 26% (8%), then divide it by 26% to obtain a quotient of approximately 0.31, and finally subtract 0.31 from 1.0 to obtain a forward value of approximately 0.69. The essence of this forward conversion method is to use the upper limit of the national standard as a benchmark to measure the superiority of the measured value relative to the standard limit. The farther the measured value is from the upper limit, the closer the forward value is to 1.0, representing better quality. When the measured value is exactly equal to the upper limit, the positive value is 0, indicating that it is just within the standard acceptable boundary. When the measured value exceeds the upper limit, i.e., it is unacceptable, the positive value will be negative. In this case, the positive value is set to 0 to avoid the unacceptable index making a negative contribution to the comprehensive evaluation. In some optional implementations, other positive values ​​such as extreme value normalization or piecewise linear mapping can also be used.

[0057] The polishing value is a positive indicator, meaning a higher measured value indicates better quality. Its positive conversion method varies. The measured polishing value is subtracted from the corresponding lower limit specified in the national standard, then divided by the same lower limit. The quotient is added to 1.0 and divided by 2.0 to obtain the positive conversion value. For example, if the measured polishing value of a certain grade of steel slag aggregate is 46, and the corresponding lower limit specified in the national standard is 42, then 46 minus 42 divided by 42 gives a quotient of approximately 0.095. Adding 1.0 gives 1.095, and dividing by 2.0 gives a positive conversion value of approximately 0.55. When the measured polishing value is exactly equal to the lower limit, the positive conversion value is 0.5, which is at an intermediate level. The higher the measured value is above the lower limit, the closer the positive conversion value is to 1.0. When the positive conversion value is greater than 1.0, 1.0 is taken as the cutoff value.

[0058] After all indicators have been converted, calculate the comprehensive quality assessment score for each grade of steel slag aggregate. . Equal to the stability continuous score The contribution value of the criterion layer multiplied by the volume stability index In addition, the positive values ​​of the five mechanical durability indicators are multiplied by the contribution values ​​of the criterion layer of the mechanical durability indicators. Then multiply each by the contribution value of its respective scheme layer and sum the results. The stability continuous score value output from step 2. This represents the contribution value of the volumetric stability index to the criterion layer. The contribution values ​​of the mechanical durability indices to the criterion layer are denoted as follows: The contribution values ​​of each of the five mechanical durability indices to the scheme layer are denoted as follows: to The positive values ​​of the five mechanical durability indices are denoted as follows: to Taking the contribution value of the criterion layer in this embodiment as an example, It is 0.75. With a value of 0.25, the volume stability index contributes 75% of the overall quality assessment score, while the five mechanical durability indexes contribute a combined 25% of the score. This ratio reflects the priority given to the stability of steel slag aggregates.

[0059] Based on the comprehensive quality assessment score, steel slag aggregates of each grade are divided into three quality levels and prepared in a targeted manner as building materials for the corresponding road layers.

[0060] Steel slag aggregate with a comprehensive quality assessment score greater than 0.8 is classified as Grade 1 aggregate. Grade 1 aggregate exhibits excellent performance in both stability and mechanical durability, making it suitable for asphalt pavement surface layers with the highest aggregate quality requirements. Grade 1 steel slag aggregate is used as coarse aggregate, replacing natural basalt crushed stone at a volumetric ratio of 60% to 80% of the total coarse aggregate volume. Volumetric replacement, rather than mass replacement, is used because the apparent density of steel slag aggregate is higher than that of natural basalt crushed stone. If replaced by mass, the actual volume occupied by steel slag aggregate would be too small, leading to a deviation of the mixture's skeleton structure from the original gradation design. The replacement ratio is set between 60% and 80%, rather than 100% replacement, to retain a portion of natural crushed stone to complement the steel slag aggregate's gradation. This is because the particle morphology and size distribution characteristics of crushed steel slag aggregate differ from those of natural crushed stone; appropriate blending of natural crushed stone helps optimize the mixture's skeleton interlocking state and porosity distribution. Grade 1 steel slag aggregate, limestone powder with a particle size less than 0.075 mm, and SBS modified asphalt are mixed at a mixing temperature of 170°C to 180°C. The optimal asphalt content is determined according to the Marshall mix design method to prepare AC-13 type steel slag asphalt concrete surface course mixture. AC-13 type is a dense-graded asphalt concrete with a nominal maximum particle size of 13.2 mm, widely used in the surface and intermediate layers of high-grade highway asphalt pavements. In some optional embodiments, SMA-13 ​​type steel slag asphalt mastic aggregate mixture can also be prepared, or the Superpave volumetric design method can be used instead of the Marshall mix design method according to engineering needs.

[0061] Steel slag aggregate with a comprehensive quality assessment score between 0.5 and 0.8 is classified as Grade 2 aggregate. Grade 2 aggregate exhibits acceptable stability and moderate mechanical properties, making it suitable for road base courses. Grade 2 steel slag aggregate is mixed with natural limestone crushed stone at a mass ratio of 7:3 to form coarse and fine aggregates. 5% (by mass) of 42.5 grade ordinary Portland cement is added, and the mixture is stirred with water under optimal moisture content conditions to prepare a cement-stabilized steel slag crushed stone base course mixture. The 7:3 mass ratio of steel slag aggregate to natural limestone crushed stone ensures high utilization of the steel slag in the mixture, while the 30% natural limestone crushed stone fills the gap in the fine-particle size range of the steel slag aggregate and improves the gradation continuity of the mixture. The 5% cement content is a commonly used proportion for cement-stabilized crushed stone base courses, which can meet strength requirements while controlling the risk of drying shrinkage cracks caused by excessive cement content in the base course. In some alternative implementations, the mass ratio of steel slag aggregate to natural crushed stone can be adjusted to 6:4 or 8:2, and the cement content can be adjusted between 3% and 6%.

[0062] like Figure 4 As shown, the horizontal axis represents the sample number of the steel slag aggregate, and the vertical axis represents the comprehensive quality assessment score. Each sample corresponds to one vertical bar, and the height of the bar represents the sample's overall quality assessment score. Two horizontal dashed lines are drawn in the figure, corresponding to the two thresholds of 0.8 and 0.5 for the overall quality assessment score, dividing the vertical axis into three intervals. Samples with an overall quality assessment score greater than 0.8 are located in the uppermost interval and are classified as Grade 1 aggregate, suitable for asphalt pavement surface layers, used to prepare AC-13 type steel slag asphalt concrete surface layer mixtures. Samples with an overall quality assessment score between 0.5 and 0.8 are located in the middle interval and are classified as Grade 2 aggregate, suitable for base course aggregates, used to prepare cement-stabilized steel slag crushed stone base course mixtures. Samples with an overall quality assessment score less than 0.5 are located in the lowermost interval and are classified as Grade 3 aggregate, suitable for subgrade fill layers, mixed with native soil for subgrade layered backfilling and compaction. Figure 4It can be seen that 6 out of the 20 samples had a comprehensive quality assessment score exceeding 0.8, falling into the Grade 1 aggregate category. These samples performed excellently in both the continuous stability score and the five mechanical durability indicators. 8 samples had a comprehensive quality assessment score between 0.5 and 0.8, classifying them as Grade 2 aggregate. Their stability was acceptable, but their mechanical durability indicators had some shortcomings. Another 6 samples had a comprehensive quality assessment score below 0.5, classifying them as Grade 3 aggregate. This includes those samples that, after three rounds of supplementary carbonization curing in step 2, still had a continuous stability score below the set threshold and were directly classified into the lowest quality grade. This three-tiered classification ensures that steel slag aggregates of different qualities can be matched to road engineering layers appropriate to their performance levels. High-quality aggregates are used in high-value applications, medium-quality aggregates in medium-requirement applications, and low-quality aggregates in low-requirement applications, thus achieving full resource utilization of steel slag aggregates and avoiding the resource waste caused by discarding substandard aggregates entirely in traditional methods.

[0063] Steel slag aggregate with a comprehensive quality assessment score less than 0.5, and steel slag aggregate that, after three supplementary carbonization curing treatments in step 2, still has a stability score below 0.5 and is therefore directly classified into the lowest quality grade, are uniformly designated as Grade 3 aggregate. Grade 3 aggregate has relatively weak stability or mechanical properties, but it can still play a role in resource utilization in the subgrade fill layer, which has the lowest requirements for aggregate quality. The subgrade fill layer is located below the pavement structure layer, and the stress level it bears is much lower than that of the surface layer and base layer. Furthermore, the direct contact between the fill layer and atmospheric moisture is significantly reduced after the entire fill layer is covered by the upper structure layer, making the impact of residual expansion risk on the overall pavement structure controllable. Grade 3 steel slag aggregate is mixed evenly with native soil at a mass ratio of 6:4 and then used for layered backfilling of the subgrade fill layer. Each layer has a loose thickness of 300 mm and is compacted to a compaction degree of over 95%. Steel slag aggregate and subsoil are mixed at a mass ratio of 6:4. The subsoil fills the voids between the steel slag aggregate particles and provides cohesion, resulting in a dense and stable subgrade structure after compaction. The required loose layer thickness of 300 mm and a compaction degree of over 95% are consistent with current subgrade construction specifications. In some alternative implementations, the mass ratio of steel slag aggregate to subsoil can be adjusted to 5:5 or 7:3, and the loose layer thickness can be adjusted between 200 mm and 400 mm depending on the compaction equipment capacity.

[0064] The aforementioned three-tiered resource utilization scheme ensures that steel slag aggregates of different qualities can find road engineering application scenarios that match their performance levels, achieving 100% resource utilization of steel slag aggregates rather than simply preferential use or one-size-fits-all downgrading. The continuous numerical characteristics of the comprehensive quality assessment score make the grading boundary more scientific, avoiding misjudgments caused by index fluctuations when relying solely on a single index for grading. The analytic hierarchy process (AHP) grading framework organically integrates the stability information output by the convolutional neural network in step 2 with traditional mechanical durability testing information, so that the final grading result simultaneously takes into account the special stability issues of steel slag aggregates compared to natural aggregates and the general mechanical performance requirements of road construction materials.

[0065] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A method for resource utilization and quality grading of road steel slag aggregate, characterized in that, Includes the following steps: Step 1: Steel slag is aged, crushed and screened to obtain multiple grades of steel slag aggregate. The multiple grades of steel slag aggregate are carbonized to convert free calcium oxide and free magnesium oxide into calcium carbonate and magnesium carbonate, respectively. Samples are taken from each grade of steel slag aggregate and standard cylindrical specimens are prepared using a high-temperature resistant inorganic binder. After spraying a high-temperature resistant speckle pattern on the side surface of the standard cylindrical specimen, it is placed in an autoclave for accelerated expansion test. The speckle image pairs are acquired through a binocular stereo vision acquisition system and digital image correlation is performed to obtain the grayscale image time series of the expansion strain field. Step 2: Extract images from multiple time points in the grayscale image time series of the expansion strain field and stack them into a multi-channel input tensor. Input the tensor into the convolutional neural network model and output the probability values ​​of three levels of stability: excellent, moderate, and poor. Calculate the continuous stability score based on the probability values ​​of the three levels. When the continuous stability score is lower than the set threshold, perform supplementary carbonization curing and repeat steps 1 and 2. The cumulative number of times supplementary carbonization curing is performed does not exceed the set maximum number of times. If the continuous stability score is still lower than the set threshold after reaching the set maximum number of times, the steel slag aggregate is directly classified into the lowest quality level. Step 3: Measure the mechanical durability index of each grade of steel slag aggregate whose continuous stability score reaches the set threshold, construct a hierarchical analysis method classification framework to calculate the comprehensive quality assessment score, divide each grade of steel slag aggregate into multiple quality grades based on the comprehensive quality assessment score, and prepare the steel slag aggregate of each quality grade into building materials for the corresponding road layer.

2. The method according to claim 1, characterized in that, In step 1, after the steel slag is naturally piled and aged for no less than 6 months, it is crushed in sequence by a jaw crusher and a secondary crusher by an impact crusher, and then screened by a 4-layer vibrating screen to obtain four grades of steel slag aggregate with particle sizes of 19 mm to 26.5 mm, 9.5 mm to 19 mm, 4.75 mm to 9.5 mm, and 0.075 mm to 4.75 mm.

3. The method according to claim 1, characterized in that, In step 1, the specific method of carbonization curing is as follows: multiple grades of steel slag aggregate are laid flat on the breathable mesh tray inside the sealed carbonization curing box, with a thickness of 150 mm. A mixture of carbon dioxide and nitrogen with a volume concentration of 20% is continuously introduced into the sealed carbonization curing box. The temperature inside the sealed carbonization curing box is controlled at 60 degrees Celsius and the relative humidity is 70%, and carbonization curing is carried out continuously for 7 days.

4. The method according to claim 1, characterized in that, In step 1, the high-temperature resistant inorganic binder is aluminum phosphate inorganic binder, and the mass of aluminum phosphate inorganic binder is 8% to 12% of the dry mass of steel slag aggregate. The diameter of the standard cylindrical specimen is 50 mm and the height is 100 mm. After preparation, it is cured at 200 degrees Celsius for 2 hours to completely ceramicize the aluminum phosphate inorganic binder. The spraying method for the high-temperature resistant speckle pattern is to first spray a layer of white high-temperature resistant inorganic primer, and after the white high-temperature resistant inorganic primer dries, spray black high-temperature resistant inorganic speckle paint. The average diameter of the speckles is 0.3 mm to 0.5 mm, and the speckle coverage is 40% to 60% of the side surface area of ​​the standard cylindrical specimen.

5. The method according to claim 1, characterized in that, In step 1, a high-temperature resistant optical observation window with an anti-fog coating on its inner surface is installed on the side wall of the autoclave corresponding to the position of the standard cylindrical specimen. The binocular stereo vision acquisition system consists of two industrial area array cameras with a resolution of no less than 5 million pixels mounted on the outside of the high-temperature resistant optical observation window. The optical axis intersection angle of the two industrial area array cameras is 25 to 35 degrees. A narrow-band blue LED surface light source with a wavelength of 450 nanometers is configured on the outside of the high-temperature resistant optical observation window as an illumination source. The conditions for the accelerated expansion test are an autoclave temperature of 216 degrees Celsius, an autoclave pressure of 2.0 MPa, and continuous autoclaving for 3 hours. The acquisition frequency of speckle image pairs is once every 30 seconds.

6. The method according to claim 1, characterized in that, In step 1, the specific method of digital image correlation is as follows: the earliest acquired speckle image pair is used as the reference image pair, and the remaining speckle image pairs are used as deformable image pairs. The reference image is divided into square sub-regions with a side length of 31 pixels and a center-to-center distance of 7 pixels between adjacent square sub-regions. For each square sub-region, the correlation coefficient is calculated pixel by pixel in the deformable image using the zero-mean normalized cross-correlation criterion to determine the integer pixel matching position. In the neighborhood of the integer pixel matching position, double cubic B-spline interpolation is used to solve the displacement vector with sub-pixel accuracy. The displacement vectors of all square sub-regions are arranged according to their spatial positions to form a full-field displacement vector field. The difference between the displacement vectors of adjacent square sub-regions is taken and divided by the center-to-center distance to obtain the radial expansion strain component and the axial expansion strain component, respectively. The radial expansion strain component and the axial expansion strain component are linearly mapped to grayscale images with grayscale values ​​from 0 to 255.

7. The method according to claim 1, characterized in that, In step 2, six time points are extracted at equal intervals. At each time point, one grayscale image of the radial expansion strain field and one grayscale image of the axial expansion strain field are taken, for a total of 12 images. These 12 images are scaled to 224 pixels by 224 pixels and then stacked to form a 224 by 224 by 12 multi-channel input tensor. The convolutional neural network model sequentially includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, a first fully connected layer, and a second fully connected layer. The first convolutional module contains 32 convolutional kernels of size 7 by 7 with a stride of 2 and padding of 3, followed by a batch normalization layer, a ReLU activation function, and a max pooling layer of size 3 by 3 with a stride of 2. The second convolutional module contains 64 convolutional kernels of size 5 by 7 with a stride of 2 and padding of 3. The first convolutional module consists of 5 convolutional kernels with a stride of 1 and padding of 2, followed by batch normalization layers, ReLU activation functions, and max pooling layers with a size of 3 x 3 and a stride of 2. The second convolutional module contains 128 convolutional kernels with a size of 3 x 3 and a stride of 1 and padding of 1, followed by batch normalization layers and ReLU activation functions. The third convolutional module contains 128 convolutional kernels with a size of 3 x 3 and a stride of 1 and padding of 1, followed by batch normalization layers, ReLU activation functions, and global average pooling layers. The global average pooling layer outputs a 128-dimensional feature vector. The first fully connected layer maps the 128-dimensional feature vector to 64 dimensions and processes it with the ReLU activation function. The second fully connected layer maps it to a 3-dimensional vector and outputs the probability values ​​of the three levels through the Softmax function.

8. The method according to claim 1, characterized in that, In step 2, the training method for the convolutional neural network model is as follows: collect no less than 500 sets of steel slag aggregate samples from different steel plants and different furnace batches; prepare standard cylindrical specimens for each set of steel slag aggregate samples according to the method in step 1 and obtain multi-channel input tensors; at the same time, take parallel samples for each set of steel slag aggregate samples and independently determine the autoclave expansion rate according to the autoclave expansion method specified by the national standard; mark autoclave expansion rate less than 0.5% as excellent stability, autoclave expansion rate between 0.5% and 2.0% as moderate stability, and autoclave expansion rate greater than 2.0% as poor stability; train the model using the cross-entropy loss function and the Adam optimizer, and save the model parameters of the round with the highest classification accuracy on the validation set.

9. The method according to claim 1, characterized in that, In step 2, the continuous stability score is calculated by multiplying the probability value of excellent stability by 1.0, adding the probability value of moderate stability by 0.6, and adding the probability value of poor stability by 0.2; the threshold is set to 0.5, and the maximum number of times is set to 3.

10. The method according to claim 1, characterized in that, In step 3, the mechanical durability indicators include crushing value, Los Angeles abrasion value, polishing value, needle-like and flaky particle content, and water absorption rate. The target layer of the analytic hierarchy process (AHP) grading framework is the comprehensive quality grade of road steel slag aggregate. The criterion layer sets two indicators: volumetric stability and mechanical durability. In the scheme layer, the volumetric stability indicator is set with a continuous stability score, and the mechanical durability indicator is set with five measured values: crushing value, Los Angeles abrasion value, polishing value, needle-like and flaky particle content, and water absorption rate. A second-order judgment matrix is ​​constructed in the criterion layer, and the Saaty 1 to 9 scaling method is used to perform pairwise comparisons and assign values ​​to the two criteria. The normalized eigenvector corresponding to the largest eigenvalue of the second-order judgment matrix is ​​calculated to obtain the contribution value of each criterion. In the scheme layer under the mechanical durability indicators, a fifth-order judgment matrix is ​​constructed, and the Saaty 1 to 9 scaling method is used to perform pairwise comparisons and assign values ​​to the five mechanical durability indicators. The maximum eigenvalue of the fifth-order judgment matrix is ​​calculated. The normalized eigenvectors yielded the contribution values ​​of the five mechanical durability indices, and the consistency ratio was verified to be less than 0.

1. The measured values ​​of crushing value, Los Angeles abrasion value, needle-like particle content, and water absorption rate were subtracted from the corresponding national standard upper limit, and the absolute value was divided by the corresponding national standard upper limit. The quotient obtained by subtracting 1.0 was used as the positive quantification value; when the positive quantification value was less than 0, it was set to 0. The measured value of polishing value was subtracted from the corresponding national standard lower limit, and the quotient was divided by the corresponding national standard lower limit. The quotient obtained by adding 1.0 to the quotient and dividing by 2.0 was used as the positive quantification value; when the positive quantification value was greater than 1.0, it was set to 1.

0. The comprehensive quality assessment score was obtained by multiplying the stability continuous score by the criterion layer contribution value of the volume stability index, adding the positive quantification values ​​of the five mechanical durability indices by the criterion layer contribution values ​​of the mechanical durability index, and then multiplying them by their respective scheme layer contribution values.

Images