Intelligent optimization method and system for concrete wear resistance based on deep learning

Abstract

The application discloses a kind of based on deep learning's concrete wear resistance intelligent optimization method and system, the method includes: by obtaining the micrograph of concrete construction material, wear resistance demand parameter and environmental parameter, utilize deep learning network to carry out cross-scale feature fusion to micrograph to extract material microstructure feature, wear resistance demand parameter and environmental parameter are combined to construct constraint condition model, then based on material microfeature, wear resistance performance prediction model is obtained by pre-training prediction value, finally according to the prediction value and constraint condition model generate optimization scheme including wear resistance material and concrete material parameter matching ratio and mixing control parameter, realize the intelligent optimization of concrete wear resistance based on deep learning.The application can improve the optimization effect of concrete wear resistance.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for intelligent optimization of concrete wear resistance performance based on deep learning. Background Technology

[0002] In the field of concrete engineering, especially in scenarios such as hydraulic structures and road surfaces that are subjected to high-speed water flow or mechanical abrasion, the abrasion resistance of concrete is a key indicator determining its durability and service life. Existing technologies for optimizing the abrasion resistance of concrete typically rely on the manual experience of technicians, adjusting material proportions (such as cement type, aggregate type and gradation, and the type and dosage of admixtures and additives) and mixing process parameters through repeated trial and error. However, this approach is inefficient.

[0003] In recent years, with the development of artificial intelligence technology, some studies have attempted to introduce machine learning methods to optimize concrete performance. These methods typically involve collecting historical mix design data and macroscopic performance indicators (such as compressive strength and abrasion resistance), using relevant modeling techniques to establish a mapping relationship between mix design parameters and abrasion resistance, and then recommending optimal material combinations or process settings. However, these methods still focus on the analysis of macroscopic parameters, failing to deeply integrate information on the microstructure of materials (such as pore distribution) or consider factors such as the construction environment and abrasion resistance requirements, resulting in less than ideal optimization effects on the abrasion resistance of concrete. Summary of the Invention

[0004] This invention provides a method and system for intelligent optimization of concrete wear resistance based on deep learning, which can improve the optimization effect on concrete wear resistance.

[0005] One embodiment of the present invention provides a method for intelligent optimization of concrete wear resistance based on deep learning, comprising: To obtain microscopic images of concrete construction material samples, abrasion resistance requirements for concrete construction, and environmental parameters for concrete construction; The microscopic image data is fused across scales using a deep learning network to extract the microscopic structural features of concrete material. Based on the wear resistance requirement parameters and the environmental parameters, a constraint condition model for optimizing the wear resistance performance of concrete is constructed. Based on the microstructure characteristics of the material, the predicted value of the wear resistance of concrete is determined by a pre-trained wear resistance prediction model. Based on the predicted wear resistance value and the constraint model, an optimization scheme for the wear resistance of concrete is generated. The optimization scheme includes the optimal ratio of wear-resistant material and concrete material and the optimal mixing control parameters of the two.

[0006] As an improvement to the above scheme, the step of performing cross-scale feature fusion on the microscopic image data through a deep learning network to extract the microscopic structural features of concrete materials includes the following sub-steps: Multi-scale structural semantic segmentation processing is performed on the microscopic image data to generate a three-level structural feature map of aggregate region, interface transition region and pore region based on the physical scale characteristics of material components. The three-level structural feature maps are processed by parallel multi-scale feature extraction branches in the deep learning network to obtain local feature sets at each scale. Specifically, the geometric morphology and spatial distribution features of the aggregate region are extracted by the first feature extraction branch of the deep learning network, the micro-topology and bond strength features of the interface transition zone are extracted by the second feature extraction branch of the deep learning network, and the size distribution and connectivity features of the pore region are extracted by the third feature extraction branch of the deep learning network. The local feature set is input into the graph neural network fusion unit in the deep learning network to construct a material structure relationship graph with aggregate, interface and pore as nodes, and cross-scale physical interaction is modeled by edge weights to obtain the fused cross-scale features. The fused cross-scale features are subjected to multi-task decoupling processing, and the aggregate-slurry interface topological features and porosity features of concrete materials are separated by an attention mechanism. The topological and porosity characteristics of the separated aggregate-slurry interface are optimized in an anti-wear manner to generate material microstructure characteristics that are geared towards wear resistance. The material microstructure characteristics include the optimized aggregate-slurry interface topological and porosity characteristics.

[0007] As an improvement to the above solution, the step of constructing a constraint model for optimizing the wear resistance performance of concrete based on the wear resistance requirement parameters and the environmental parameters includes the following sub-steps: The wear resistance requirements parameters are quantified in multiple dimensions to identify the core requirements indicators for the wear resistance performance of concrete. Environmental sensitivity analysis was performed on the environmental parameters to identify the key environmental factors affecting the wear resistance of concrete. Based on the core requirement indicators and key environmental factors, a constraint rule base for optimizing the wear resistance of concrete is established. The constraint rule base is transformed into quantifiable constraint functions to construct the constraint model for optimizing the wear resistance of concrete.

[0008] As an improvement to the above scheme, the step of determining the predicted value of the abrasion resistance of concrete based on the microstructural characteristics of the material and through a pre-trained abrasion resistance prediction model includes the following sub-steps: Feature importance analysis was performed on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete; The key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding to obtain the encoded material microstructure features; The coded material microstructure features are transformed into wear resistance performance correlation vectors through the feature mapping mechanism in the wear resistance performance prediction model. The wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation; The weighted feature representation is calculated using the output layer of the wear resistance prediction model to obtain the predicted value of the wear resistance of concrete.

[0009] As an improvement to the above scheme, based on the predicted wear resistance value and the constraint model, an optimized scheme for the wear resistance of concrete is generated. The optimized scheme includes the optimal ratio of wear-resistant material to concrete material and the optimal mixing control parameters for the two, including the following sub-steps: Based on the predicted wear resistance value and the constraint model, a bi-objective optimization function is constructed with the objectives of maximizing wear resistance and minimizing material cost. The weighted summation method is then used to transform the bi-objective function into a single-objective optimization problem of concrete wear resistance. The single-objective optimization function is solved iteratively using the Bayesian optimization algorithm. A mapping relationship between the parameter space and the objective space is established through Gaussian process regression to generate a set of candidate solutions for the wear resistance of concrete. Based on the constraint model, the constraint satisfaction of the candidate scheme set is verified to obtain a feasible solution set that satisfies all the constraint conditions for optimizing the wear resistance of concrete. Multi-objective optimization solution set analysis is performed on the feasible solution set, and the Pareto front distribution characteristics of wear resistance performance and material cost are obtained based on the analysis results; Based on the Pareto front distribution characteristics, the optimal solution cluster in the feasible solution set is identified by density clustering algorithm, and the multiple similar candidate solutions contained in the optimal solution cluster are sorted according to the center position of the solution cluster to obtain the sorted optimal candidate solution set. Based on the Pareto optimality principle, the optimal solution is selected from the sorted set of optimal candidate solutions. By analyzing the parameter space mapping relationship corresponding to the optimal solution, the optimal ratio of the wear-resistant material and the concrete material and the optimal mixing control parameters of the two are obtained by inversion, and the optimized solution for the wear-resistant performance of concrete is output.

[0010] Another embodiment of the present invention provides a deep learning-based intelligent optimization system for the wear resistance of concrete, comprising: The acquisition module is used to acquire microscopic images of concrete construction material samples, abrasion resistance requirements parameters for concrete construction, and environmental parameters for concrete construction. The extraction module is used to perform cross-scale feature fusion on the microscopic image data through a deep learning network to extract the material microstructure features of concrete. A construction module is used to construct a constraint model for optimizing the wear resistance performance of concrete based on the wear resistance requirement parameters and the environmental parameters. The prediction module is used to determine the predicted value of the wear resistance of concrete based on the microstructure characteristics of the material and through a pre-trained wear resistance prediction model. An optimization module is used to generate an optimization scheme for the wear resistance performance of concrete based on the predicted wear resistance performance value and the constraint condition model. The optimization scheme includes the optimal ratio of wear resistance material and concrete material and the optimal mixing control parameters of the two.

[0011] As an improvement to the above solution, the extraction module is specifically used for: Multi-scale structural semantic segmentation processing is performed on the microscopic image data to generate a three-level structural feature map of aggregate region, interface transition region and pore region based on the physical scale characteristics of material components. The three-level structural feature maps are processed by parallel multi-scale feature extraction branches in the deep learning network to obtain local feature sets at each scale. Specifically, the geometric morphology and spatial distribution features of the aggregate region are extracted by the first feature extraction branch of the deep learning network, the micro-topology and bond strength features of the interface transition zone are extracted by the second feature extraction branch of the deep learning network, and the size distribution and connectivity features of the pore region are extracted by the third feature extraction branch of the deep learning network. The local feature set is input into the graph neural network fusion unit in the deep learning network to construct a material structure relationship graph with aggregate, interface and pore as nodes, and cross-scale physical interaction is modeled by edge weights to obtain the fused cross-scale features. The fused cross-scale features are subjected to multi-task decoupling processing, and the aggregate-slurry interface topological features and porosity features of concrete materials are separated by an attention mechanism. The topological and porosity characteristics of the separated aggregate-slurry interface are optimized in an anti-wear manner to generate material microstructure characteristics that are geared towards wear resistance. The material microstructure characteristics include the optimized aggregate-slurry interface topological and porosity characteristics.

[0012] As an improvement to the above solution, the building module is specifically used for: The wear resistance requirements parameters are quantified in multiple dimensions to identify the core requirements indicators for the wear resistance performance of concrete. Environmental sensitivity analysis was performed on the environmental parameters to identify the key environmental factors affecting the wear resistance of concrete. Based on the core requirement indicators and key environmental factors, a constraint rule base for optimizing the wear resistance of concrete is established. The constraint rule base is transformed into quantifiable constraint functions to construct the constraint model for optimizing the wear resistance of concrete.

[0013] As an improvement to the above scheme, the prediction module is specifically used for: Feature importance analysis was performed on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete; The key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding to obtain the encoded material microstructure features; The coded material microstructure features are transformed into wear resistance performance correlation vectors through the feature mapping mechanism in the wear resistance performance prediction model. The wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation; The weighted feature representation is calculated using the output layer of the wear resistance prediction model to obtain the predicted value of the wear resistance of concrete.

[0014] As an improvement to the above solution, the optimization module is specifically used for: Based on the predicted wear resistance value and the constraint model, a bi-objective optimization function is constructed with the objectives of maximizing wear resistance and minimizing material cost. The weighted summation method is then used to transform the bi-objective function into a single-objective optimization problem of concrete wear resistance. The single-objective optimization function is solved iteratively using the Bayesian optimization algorithm. A mapping relationship between the parameter space and the objective space is established through Gaussian process regression to generate a set of candidate solutions for the wear resistance of concrete. Based on the constraint model, the constraint satisfaction of the candidate scheme set is verified to obtain a feasible solution set that satisfies all the constraint conditions for optimizing the wear resistance of concrete. Multi-objective optimization solution set analysis is performed on the feasible solution set, and the Pareto front distribution characteristics of wear resistance performance and material cost are obtained based on the analysis results; Based on the Pareto front distribution characteristics, the optimal solution cluster in the feasible solution set is identified by density clustering algorithm, and the multiple similar candidate solutions contained in the optimal solution cluster are sorted according to the center position of the solution cluster to obtain the sorted optimal candidate solution set. Based on the Pareto optimality principle, the optimal solution is selected from the sorted set of optimal candidate solutions. By analyzing the parameter space mapping relationship corresponding to the optimal solution, the optimal ratio of the wear-resistant material and the concrete material and the optimal mixing control parameters of the two are obtained by inversion, and the optimized solution for the wear-resistant performance of concrete is output.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention acquires microscopic images of concrete construction material samples, along with abrasion resistance requirements and environmental parameters. A deep learning network is used to perform cross-scale feature fusion on the microscopic image data, extracting the microstructural features of the concrete material. This feature fusion improves the comprehensiveness and accuracy of the material's microstructural characterization by integrating microstructural information at different scales. A constraint model for optimizing concrete abrasion resistance is constructed based on the abrasion resistance requirements and environmental parameters. This model correlates abrasion resistance requirements with environmental factors, adapting the constraints to actual engineering scenarios. Based on the material's microstructural features, a pre-trained abrasion resistance prediction model determines the predicted abrasion resistance value of the concrete. This process leverages the inherent correlation between microstructural features and abrasion resistance to improve prediction accuracy. Based on the predicted abrasion resistance value and the constraint model, an optimized abrasion resistance performance scheme is generated, including the optimal ratio of abrasion-resistant materials to concrete materials and optimal mixing control parameters, achieving targeted adjustments to abrasion resistance performance. This embodiment of the invention, by deeply integrating material microstructural information and combining the construction environment and abrasion resistance requirements to construct a constraint model and perform performance prediction, solves the problem of insufficient optimization effects in existing technologies, enabling effective optimization of concrete abrasion resistance performance. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an intelligent optimization method for the wear resistance of concrete based on deep learning, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a deep learning-based intelligent optimization system for the wear resistance of concrete, provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1This is a flowchart illustrating a method for intelligently optimizing the wear resistance of concrete based on deep learning, according to an embodiment of the present invention. The method includes the following steps: S10: Obtain microscopic images of concrete construction material samples, abrasion resistance requirements parameters for concrete construction, and environmental parameters for concrete construction. S11, cross-scale feature fusion is performed on the microscopic image data through a deep learning network to extract the material microstructure features of concrete. S12, Based on the wear resistance requirement parameters and the environmental parameters, construct a constraint condition model for optimizing the wear resistance performance of concrete; S13, Based on the microstructure characteristics of the material, the predicted value of the wear resistance of concrete is determined by a pre-trained wear resistance prediction model; S14. Based on the predicted wear resistance value and the constraint model, an optimization scheme for the wear resistance of concrete is generated. The optimization scheme includes the optimal ratio of wear-resistant material and concrete material and the optimal mixing control parameters of the two.

[0019] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention acquires microscopic images of concrete construction material samples, along with abrasion resistance requirements and environmental parameters. A deep learning network is used to perform cross-scale feature fusion on the microscopic image data, extracting the microstructural features of the concrete material. This feature fusion improves the comprehensiveness and accuracy of the material's microstructural characterization by integrating microstructural information at different scales. A constraint model for optimizing concrete abrasion resistance is constructed based on the abrasion resistance requirements and environmental parameters. This model correlates abrasion resistance requirements with environmental factors, adapting the constraints to actual engineering scenarios. Based on the material's microstructural features, a pre-trained abrasion resistance prediction model determines the predicted abrasion resistance value of the concrete. This process leverages the inherent correlation between microstructural features and abrasion resistance to improve prediction accuracy. Based on the predicted abrasion resistance value and the constraint model, an optimized abrasion resistance performance scheme is generated, including the optimal ratio of abrasion-resistant materials to concrete materials and optimal mixing control parameters, achieving targeted adjustments to abrasion resistance performance. This embodiment of the invention, by deeply integrating material microstructural information and combining the construction environment and abrasion resistance requirements to construct a constraint model and perform performance prediction, solves the problem of insufficient optimization effects in existing technologies, enabling effective optimization of concrete abrasion resistance performance.

[0020] In this embodiment, firstly, microscopic images of concrete construction material samples are obtained. Standard specimens are prepared by mixing aggregates, cement paste, and admixtures used in construction according to the design ratio. After vacuum drying, images are captured using an electron microscope at a resolution of 10nm-1μm, with the observation scale covering 0.5μm-15mm to ensure the capture of aggregates, interface transition zones, and pore structures. Secondly, wear resistance requirements are obtained by extracting parameters from engineering design documents and owner requirements, including the cumulative wear limit within the design service life (e.g., ≤0.3kg / m²), the maximum allowable wear rate per unit time (e.g., ≤0.01g / m²・h), and the wear resistance stability requirements corresponding to the load level. Finally, environmental parameters are obtained, such as data collected continuously at different times by temperature and humidity sensors and freeze-thaw monitoring instruments deployed at the construction site, combined with data from the local meteorological department in recent years, to determine the annual average temperature fluctuation (e.g., -10℃~35℃), the number of annual freeze-thaw cycles (e.g., ≥50 times), the annual average humidity (e.g., 40%~85%RH), and the type of media erosion (e.g., sandy river water).

[0021] As one example, the step of performing cross-scale feature fusion on the microscopic image data using a deep learning network to extract the microscopic structural features of concrete material includes the following sub-steps: Multi-scale structural semantic segmentation processing is performed on the microscopic image data to generate a three-level structural feature map of aggregate region, interface transition region and pore region based on the physical scale characteristics of material components. The three-level structural feature maps are processed by parallel multi-scale feature extraction branches in the deep learning network to obtain local feature sets at each scale. Specifically, the geometric morphology and spatial distribution features of the aggregate region are extracted by the first feature extraction branch of the deep learning network, the micro-topology and bond strength features of the interface transition zone are extracted by the second feature extraction branch of the deep learning network, and the size distribution and connectivity features of the pore region are extracted by the third feature extraction branch of the deep learning network. The local feature set is input into the graph neural network fusion unit in the deep learning network to construct a material structure relationship graph with aggregate, interface and pore as nodes, and cross-scale physical interaction is modeled by edge weights to obtain the fused cross-scale features. The fused cross-scale features are subjected to multi-task decoupling processing, and the aggregate-slurry interface topological features and porosity features of concrete materials are separated by an attention mechanism. The topological and porosity characteristics of the separated aggregate-slurry interface are optimized in an anti-wear manner to generate material microstructure characteristics that are geared towards wear resistance. The material microstructure characteristics include the optimized aggregate-slurry interface topological and porosity characteristics.

[0022] In this embodiment, multi-scale structural semantic segmentation can accurately divide the core structural regions of concrete materials, avoiding confusion of features in different regions and improving the targeting of subsequent feature extraction. Parallel multi-scale feature extraction branches can focus on the key characteristics of aggregates, interface transition zones, and pores respectively, ensuring the accuracy of feature representation of each structural region and reducing feature omissions caused by a single extraction method. The graph neural network fusion unit achieves deep fusion of features at different scales by constructing a material structure relationship graph and modeling cross-scale physical interactions, eliminating the problem of information fragmentation between scales. The attention mechanism can accurately separate the aggregate-slurry interface topological features and porosity features that are key to the wear resistance of concrete. Adverse optimization further strengthens the correlation between features and wear resistance, and the final generated material microstructure features can more accurately reflect the intrinsic influencing factors of concrete wear resistance.

[0023] Specifically, the working process of this embodiment is as follows: First, multi-scale structural semantic segmentation processing is performed on the microscopic image data. First, microscopic images of concrete construction material samples are acquired using a scanning electron microscope. For example, the image resolution is set to 1024×1024 pixels, and the observation scale covers 0.5μm-15mm to ensure that the structure of each component of the material is covered. The images are processed using the U-Net semantic segmentation network. This network adds three multi-scale receptive field modules at the encoding end, corresponding to the physical scale ranges of 0.5μm-10μm (pore scale), 10μm-100μm (interface transition zone scale), and 100μm-15mm (aggregate scale), respectively. Based on the physical scale characteristics of the material components (aggregate particle size is usually 0.2mm-15mm, interface transition zone thickness is usually 8μm-60μm, and pore diameter is usually 0.5μm-40μm), the image is divided into aggregate region, interface transition zone, and pore region through pixel-level classification. Finally, a pixel-level mask map of each region is generated, that is, a three-level structural feature map. Each feature map completely preserves the spatial location and morphological information of the corresponding region, providing a precise region segmentation basis for subsequent feature extraction.

[0024] Next, the three-level structure feature map is processed through parallel multi-scale feature extraction branches in the deep learning network. The first feature extraction branch uses an improved ResNet-50 network, taking the aggregate region feature map as input. It first extracts basic features such as aggregate edges and contours through 3×3 convolutional layers, and then calculates geometric morphological parameters such as equivalent aggregate size and shape factor through global average pooling layers. At the same time, a spatial attention module is introduced to capture the spatial distribution information of aggregates (such as the spacing between adjacent aggregates and the distribution density) by analyzing the differences in pixel gray-level distribution. Finally, a 128-dimensional local feature vector of aggregates is output. The second feature extraction branch uses a DenseNet-121 network, taking the interface transition area feature map as input, and strengthens the extraction of interface detail features through dense connection layers. The first branch captures microscopic topological information such as interface roughness and texture complexity. Then, it combines the correlation model fitted based on 100 sets of experimental data on interface density and bonding strength to transform the topological features into a 64-dimensional bonding strength feature vector, forming a local feature set of the interface transition zone. The second branch uses the MobileNet-V3 network to perform depthwise separable convolution on the feature map of the pore region to extract dimensional information such as the equivalent diameter and area of ​​the pores. Then, it uses a connected component analysis algorithm to count connectivity parameters such as the number of connected pores and the length of connected paths, and integrates them into a 64-dimensional pore local feature vector. The feature vectors of the three branches together constitute the local feature set.

[0025] Then, the local feature set is input into the graph neural network (GNN) fusion unit. First, the local feature vectors of aggregate, interface transition zone, and pores are used as the three types of nodes in the graph (denoted as...). , , ,in This is a 128-dimensional aggregate node feature vector. It is a 64-dimensional interface node feature vector. (64-dimensional pore node feature vectors), and then construct node connection edges based on the spatial positional relationships of each region in the microscopic image (e.g., aggregate nodes connected to adjacent interface nodes, and interface nodes connected to adjacent pore nodes); to model cross-scale physical interactions, an improved edge weight calculation formula is designed: , For nodes With nodes The edge weight is used to characterize the strength of the association between two nodes; , The weighting coefficients are calibrated through 50 sets of cross-scale interactive experiments to balance the influence of feature correlation and spatial distance. The cosine similarity between the features of two nodes reflects the degree of correlation at the feature level. The spatial distance decay function, This represents the shortest spatial distance between two regions in the microscopic image. Based on the diffusion experiments of concrete material components, the attenuation effect of simulated distance on physical interaction is determined to be a constant; the features of each node are updated through the GNN message passing mechanism, and finally the 256-dimensional fused cross-scale features are output.

[0026] Subsequently, multi-task decoupling processing is performed on the cross-scale features. A multi-task learning framework is adopted, with two parallel decoupled branches: topological feature extraction from the input interface and porosity feature extraction. An improved attention module is set before each branch to enhance the feature channels relevant to the target task. The attention weight calculation formula is as follows: , For the first The attention weights for each feature channel, with values ​​ranging from [0,1]; For the first The correlation score between each channel feature and the target task (calculated by the mutual information between the feature and the task label, where the interface topology task label is the measured value of concrete interface roughness and the porosity task label is the measured value of concrete pore volume fraction). The sum of the exponential scores of all feature channels is used to normalize the weights. Through this mechanism, the interface topology branch strengthens the feature channels related to texture and thickness, while the porosity branch strengthens the feature channels related to pore quantity and size, outputting a 64-dimensional aggregate-slurry interface topology feature vector (including parameters such as interface roughness and interface thickness) and a 64-dimensional porosity feature vector (including parameters such as pore volume fraction and pore size distribution), respectively.

[0027] Finally, adversarial optimization is performed on the separated features. An adversarial optimization network is constructed, including a generator. With discriminator Generator Using the separated interface topology and porosity features as input, the optimized features are output through three fully connected layers. Discriminator enter Microstructural characteristics of truly wear-resistant concrete (Obtained through testing 100 sets of concrete samples that met the wear resistance standard), output the similarity score between the two. Design an improved loss function: , For expectation operators; This represents the probability that the discriminator recognizes truly superior features. The discriminator's recognition probability of the optimized features output by the generator; This is the regularization coefficient (balancing the magnitude of feature optimization with the authenticity of the original features). To optimize the L2 norm of the features before and after training, the optimization magnitude is constrained. The generator is trained alternately. With discriminator The process continues until the preset loss function converges, and finally outputs the material microstructure features, including the optimized aggregate-slurry interface topology and porosity features.

[0028] As one example, the process of constructing a constraint model for optimizing the wear resistance performance of concrete based on the wear resistance requirement parameters and the environmental parameters includes the following sub-steps: The wear resistance requirements parameters are quantified in multiple dimensions to identify the core requirements indicators for the wear resistance performance of concrete. Environmental sensitivity analysis was performed on the environmental parameters to identify the key environmental factors affecting the wear resistance of concrete. Based on the core requirement indicators and key environmental factors, a constraint rule base for optimizing the wear resistance of concrete is established. The constraint rule base is transformed into quantifiable constraint functions to construct the constraint model for optimizing the wear resistance of concrete.

[0029] In this embodiment, the wear resistance requirements of concrete construction are first quantified from multiple dimensions. By decomposing the requirement dimensions and converting them into comparable quantitative indicators, the core requirements for concrete wear resistance are accurately identified. This process eliminates interference from non-critical requirements, allowing subsequent constraints to focus on indicators that play a decisive role in wear resistance performance in actual engineering. Next, environmental sensitivity analysis is performed on the environmental parameters of concrete construction. By calculating the correlation between different environmental factors and the attenuation of concrete wear resistance, key environmental factors affecting concrete wear resistance are identified, thereby filtering out environmental parameters with weak influence and reducing redundant data on the constraint model. Interference is eliminated to ensure that constraints accurately relate to factors that truly affect wear resistance performance in the actual environment. Subsequently, based on the identified core requirement indicators and determined key environmental factors, a constraint rule base for optimizing concrete wear resistance performance is established. The core requirements and key environmental factors are mapped to form executable constraint logic, making the constraint logic directly fit the actual engineering scenario. Finally, the constraint rule base is transformed into quantifiable constraint functions to construct a constraint model. Ultimately, the constraint model can effectively limit the optimization direction, ensuring that the generated optimization scheme not only meets the core wear resistance requirements but also adapts to key environmental influences, thereby improving the practical applicability and reliability of the optimization scheme.

[0030] The specific process of this embodiment is illustrated below: First, the wear resistance requirements of concrete construction are quantified from multiple dimensions, and core requirement indicators are identified. The wear resistance requirements proposed by the project are collected, including three types of parameters: cumulative wear limit within the design service life, maximum allowable wear rate per unit time, and wear resistance stability requirements under different load levels. An improved entropy weighted analytic hierarchy process is used for multi-dimensional quantification: the cumulative wear limit is converted into "annual average wear," the maximum allowable wear rate per unit time is unified into "hourly wear rate," and the wear resistance stability requirements corresponding to load levels are converted into "load fluctuation coefficients." Then, weight calculation is performed using an improved weighting formula. , For the first The comprehensive weight of the class requirement parameters, The information entropy of this parameter (reflecting the parameter's dispersion; the lower the dispersion) The smaller the value, the higher the weighting percentage. Subjective weights determined for the analytic hierarchy process (scored by multiple engineering experts based on project importance). This represents the number of demand parameter categories. As an example, through calculation, if the overall weight of the annual average wear is the highest, then "annual average wear ≤ 0.3 kg / m²·a" is identified as the core demand indicator for the wear resistance performance of concrete.

[0031] Next, an environmental factor sensitivity analysis was conducted on the environmental parameters of concrete construction to identify key environmental factors. First, environmental parameters of the construction area were collected, including four categories of factors: annual average temperature fluctuation range, annual average humidity, annual freeze-thaw cycle count, and type of media erosion (e.g., fresh water, sandy water, chemical solutions). A modified grey relational analysis was used for sensitivity analysis: the abrasion resistance degradation rate of standard concrete specimens under different environments was used as the reference sequence (denoted as X_0), and the measured values ​​of the four environmental factors were used as the comparison sequence (denoted as...). , ), Introducing the environmental impact coefficient to improve the correlation calculation formula: In the formula For the first The correlation between environmental factors and the rate of wear resistance degradation (the higher the value, the higher the sensitivity). The resolution coefficient, The influence weight of environmental factors (calibrated by multiple sets of environmental-wear resistance experiments). For the first The actual fluctuation range of the class factor (e.g., the fluctuation range of the number of freeze-thaw cycles is taken as "annual maximum value - annual minimum value"). For example, if the correlation of the annual number of freeze-thaw cycles is calculated... Correlation between annual average temperature fluctuation range and Therefore, "annual freeze-thaw cycles ≥ 50 times" and "annual average temperature fluctuation range ≥ 15℃" are identified as key environmental factors affecting the wear resistance of concrete.

[0032] Then, a rule base for constraint conditions to optimize the wear resistance of concrete is established based on core requirement indicators and key environmental factors. The rule base is constructed with "core requirement indicators as the benchmark and key environmental factors as the basis for correction". Each rule includes two parts: "preconditions" and "constraint requirements". For example, the first rule, if the construction environment meets the requirement of "annual freeze-thaw cycle count ≥ 50 times" and the core requirement is "annual average wear amount ≤ 0.3 kg / m²·a", then the constraint requirements are "concrete interface transition zone bond strength ≥ 3.5 MPa" and "porosity ≤ 8%"; the second rule, if the construction environment meets the requirement of "annual average temperature fluctuation range ≥ 15℃" and the core requirement is "annual average wear amount ≤ 0.3 kg / m²·a", then the constraint requirements are "concrete aggregate compressive strength ≥ 120 MPa" and "wear-resistant material content ≥ 5%"; the third rule, if two key environmental factors are met simultaneously, then the constraint requirements are increased by 10% on the basis of single-factor constraints (e.g., interface transition zone bond strength ≥ 3.85 MPa). The rule base is pre-reviewed and verified by engineering experts to ensure that each rule can cover the coupled impact of core requirements and key environmental factors, and is directly related to the microstructural characteristics of concrete (such as interfacial bond strength and porosity), providing a clear logical basis for subsequent quantitative constraints.

[0033] Finally, the constraint rule base is transformed into quantifiable constraint functions to construct a constraint model. For example, based on the core demand indicator "annual average wear ≤ 0.3 kg / m²·a", and combined with the environmental correction logic in the rule base, a linear constraint function is established: In the formula This represents the average annual wear of concrete under actual working conditions (parameter to be optimized). As the benchmark value for core demand indicators, The number of key environmental factors, For the first The influence coefficients of key environmental factors (determined by fitting experimental data) For the first The actual compliance coefficient of each key environmental factor (if the annual freeze-thaw cycle count is ≥50 times) ,otherwise (The same applies to temperature fluctuations). Simultaneously, the microstructure constraints in the rule base, such as "interface transition zone bond strength ≥ 3.5 MPa," are transformed into: ( For interfacial bond strength, , (Multi-factor superposition coefficient) ( Porosity (This is a predefined base value). Integrating the above functions forms a multi-dimensional constraint model that includes wear amount and microstructure parameters. This model can be directly connected to the subsequent anti-wear performance prediction and optimization process to achieve quantitative calculation of the constraint conditions.

[0034] As one example, determining the predicted abrasion resistance of concrete based on the microstructural characteristics of the material using a pre-trained abrasion resistance prediction model includes the following sub-steps: Feature importance analysis was performed on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete; The key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding to obtain the encoded material microstructure features; The coded material microstructure features are transformed into wear resistance performance correlation vectors through the feature mapping mechanism in the wear resistance performance prediction model. The wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation; The weighted feature representation is calculated using the output layer of the wear resistance prediction model to obtain the predicted value of the wear resistance of concrete.

[0035] In this embodiment, firstly, feature importance analysis is performed on the microstructure characteristics of the material to screen the key feature subset that dominates wear resistance performance, reducing interference from non-key features and making the features input into the model more targeted. Then, the key feature subset is input into the model's feature encoding layer for encoding. By optimizing the feature expression, the features are made more compatible with the model's processing mechanism, enhancing the resolvability of the features within the model. Subsequently, a feature mapping mechanism transforms the encoded features into wear resistance performance correlation vectors, directly establishing the correlation between microstructure characteristics and wear resistance performance, making the feature expression more closely aligned with the intrinsic laws of the prediction target. Next, the correlation vectors are dynamically weighted to generate weighted feature representations. By strengthening the weights of features that significantly affect wear resistance performance, the feature representations more accurately reflect the determinants of wear resistance performance. Finally, the model output layer calculates the weighted feature representations to obtain the predicted wear resistance performance value. Therefore, this embodiment enables a closer correlation between features and wear resistance performance, resulting in higher accuracy of the final predicted value.

[0036] As a specific example, this embodiment first performs feature importance analysis on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete. The collected microstructure characteristics include 64-dimensional aggregate-slurry interface topological features (such as interface roughness, interface thickness, interface density, etc.) and 64-dimensional porosity features (such as pore volume fraction, pore size distribution, pore connectivity, etc.), totaling 128 features. An improved random forest feature importance evaluation method is used for analysis: by constructing a random forest model composed of multiple decision trees, using the measured value of concrete wear resistance as the label, the average reduction in impurity for each feature across all decision trees is calculated. Then, a feature mutual information correction factor is introduced to obtain the improved feature importance scoring formula: ,in For the first The importance score of each feature This represents the average reduction in impurity for this feature. This is the mutual information value between the feature and the wear resistance performance (range 0-1). An importance score threshold of 0.015 is set, and features with scores higher than this threshold (such as interface roughness, interface density, pore volume fraction, maximum pore diameter, etc.) are selected to form a 64-dimensional subset of key features. This ensures that features significantly affecting wear resistance performance are retained, while redundant information is removed.

[0037] Next, the key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding. The feature encoding layer adopts a three-layer fully connected neural network structure, with 128, 96, and 64 neurons in each layer, and LeakyReLU activation function. The key feature subset is first standardized (scaled to the [0,1] interval) and then input into the encoding layer: the first layer maps the 64-dimensional features to a 128-dimensional space to enhance the feature expressive power; the second layer reduces the risk of overfitting through sparse connections (connection probability 0.7) and compresses the 128-dimensional features to 96 dimensions; the third layer uses batch normalization to output the 64-dimensional encoded material microstructure features. This feature retains the core information of the original key features and enhances the feature discrimination through nonlinear transformation, providing a better input form for subsequent mapping processes.

[0038] Then, through the feature mapping mechanism in the wear resistance performance prediction model, the encoded material microstructure features are transformed into wear resistance performance correlation vectors. The feature mapping mechanism is implemented using an improved radial basis function network, consisting of one input layer (64 neurons), one hidden layer (32 neurons), and one output layer (16 neurons). Each neuron in the hidden layer corresponds to a Gaussian kernel function. ,in This is the encoded 64-dimensional feature vector. For the first The center vectors of the Gaussian kernels (determined from the encoded features in the training set using the k-means clustering algorithm). For the first The width parameter of each Gaussian kernel is set to 0.8 times the average distance between cluster centers. The output layer maps the 32-dimensional hidden layer output to a 16-dimensional wear resistance performance correlation vector through linear transformation. Each element in this vector corresponds to a specific dimension of wear resistance performance (such as short-term wear rate, long-term wear stability, environmental adaptability, etc.), realizing a direct mapping from microstructural features to wear resistance performance related dimensions.

[0039] Subsequently, the wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation. A dynamic weighting adjustment mechanism is designed based on key factors of the concrete construction environment (such as the annual number of freeze-thaw cycles and the annual average temperature fluctuation range). The weight calculation formula is as follows: ,in The first correlation vector for wear resistance performance The weight of each element, For the first The correlation coefficients between each element and the wear resistance performance (obtained through statistical analysis of training data) For the first The environmental adaptability coefficient of each element (dynamically adjusted based on key environmental factors, such as increasing the coefficient of elements related to freeze-thaw resistance when freeze-thaw cycles are frequent). (Value). Each element of the 16-dimensional wear resistance performance association vector is multiplied by its corresponding weight to obtain a weighted feature representation. This representation enhances the wear resistance performance dimension that matches the current construction environment, making subsequent predictions more relevant to actual application scenarios.

[0040] Finally, the weighted feature representation is calculated through the output layer of the wear resistance prediction model to obtain the predicted value of the concrete's wear resistance. The output layer adopts a two-layer fully connected network structure. The first layer is a fully connected layer with 32 neurons (activation function is ReLU), mapping the 16-dimensional weighted feature representation to a 32-dimensional space; the second layer is an output layer with 1 neuron (no activation function), directly outputting the predicted value of the wear resistance. During the calculation process, a construction environment correction term is introduced. ,in This is the predicted value for the final wear resistance performance. This is the basic calculation result of the output layer. The number of key environmental factors, For the first Correction factors for each environmental factor (e.g., let the correction factor for a freeze-thaw cycle be 1). The correction factor for temperature fluctuations is ), For the first The standardized values ​​(ranging from 0 to 1) of each environmental factor are used. This correction mechanism enables the predicted values ​​to reflect the impact of environmental factors, and the final output of the wear resistance performance prediction value can be directly used in the subsequent optimization scheme generation process.

[0041] As one example, the optimization scheme for the wear resistance performance of concrete is generated based on the predicted wear resistance value and the constraint model. The optimization scheme includes the optimal ratio of wear-resistant material and concrete material and the optimal mixing control parameters for the two, including the following sub-steps: Based on the predicted wear resistance value and the constraint model, a bi-objective optimization function is constructed with the objectives of maximizing wear resistance and minimizing material cost. The weighted summation method is then used to transform the bi-objective function into a single-objective optimization problem of concrete wear resistance. The single-objective optimization function is solved iteratively using the Bayesian optimization algorithm. A mapping relationship between the parameter space and the objective space is established through Gaussian process regression to generate a set of candidate solutions for the wear resistance of concrete. Based on the constraint model, the constraint satisfaction of the candidate scheme set is verified to obtain a feasible solution set that satisfies all the constraint conditions for optimizing the wear resistance of concrete. Multi-objective optimization solution set analysis is performed on the feasible solution set, and the Pareto front distribution characteristics of wear resistance performance and material cost are obtained based on the analysis results; Based on the Pareto front distribution characteristics, the optimal solution cluster in the feasible solution set is identified by density clustering algorithm, and the multiple similar candidate solutions contained in the optimal solution cluster are sorted according to the center position of the solution cluster to obtain the sorted optimal candidate solution set. Based on the Pareto optimality principle, the optimal solution is selected from the sorted set of optimal candidate solutions. By analyzing the parameter space mapping relationship corresponding to the optimal solution, the optimal ratio of the wear-resistant material and the concrete material and the optimal mixing control parameters of the two are obtained by inversion, and the optimized solution for the wear-resistant performance of concrete is output.

[0042] In this embodiment, a bi-objective optimization function is first constructed based on the predicted wear resistance performance and the constraint model, aiming to maximize wear resistance performance and minimize material cost. This function is then transformed into a single-objective optimization problem using a weighted summation method. This design considers both the need to improve wear resistance performance and cost control, avoiding the imbalance between insufficient performance or excessive cost in a single-objective-driven solution, thus achieving a reasonable balance between the two. Next, a Bayesian optimization algorithm is used to iteratively solve the single-objective optimization function. Gaussian process regression is used to establish a mapping relationship between the parameter space and the objective space to generate a set of candidate solutions. This efficient exploration of the parameter space reduces invalid computation, improving the efficiency and reliability of candidate solution generation and ensuring that the generated solutions have a certain degree of adaptability. Finally, the constraint satisfaction of the candidate solution set is verified according to the constraint model. The process begins by obtaining a feasible solution set. Solutions that do not meet the actual engineering requirements are filtered out, ensuring that all solutions analyzed subsequently satisfy the optimization constraints. Next, multi-objective optimization solution set analysis is performed on the feasible solution set to obtain the Pareto front distribution characteristics. A density clustering algorithm is then used to identify the optimal solution cluster and rank the candidate solutions. This clarifies the optimal balance range between wear resistance and cost, narrowing the range of high-quality solutions and improving the accuracy of optimal solution identification. Finally, based on the Pareto optimality principle, the optimal solution is selected from the ranked solution set. The optimal ratio of wear-resistant materials to concrete materials and the optimal mixing control parameters are obtained by analyzing the parameter space mapping relationship. This ensures that the final optimized solution not only meets the balance requirements of wear resistance and cost but also adapts to actual construction conditions, enhancing the application value of the optimized solution in engineering scenarios.

[0043] The implementation process of this embodiment is as follows: First, based on the predicted wear resistance value and the constraint condition model, a bi-objective optimization function is constructed and transformed into a single-objective optimization problem. Specifically, the objective is to "maximize the wear resistance of concrete (i.e., the predicted wear resistance value)". Minimize (reflecting the wear resistance of materials) and minimize material costs (denoted as...) With the objective of "reflecting economic feasibility", a dual-objective optimization function is constructed: ,in The decision variables include parameters related to wear resistance performance and cost, such as wear-resistant material content, concrete aggregate gradation, mixing speed, and mixing time. To adapt to the subsequent single-objective optimization algorithm, an improved weighted summation method is used to transform the bi-objective problem into a single-objective problem. The transformation formula is as follows: , The maximum value of the predicted wear resistance performance in the sample set. The maximum value of material cost in the sample set (used for...) and (Perform dimensionless transformation to eliminate dimensional differences). , The target weight can be dynamically adjusted based on the core requirements for concrete wear resistance. Simultaneously, inequality constraints from the aforementioned constraint model (such as...) are introduced. ,in The core requirement threshold for wear resistance performance, (Using the cost constraint threshold) to form a complete single-objective optimization problem.

[0044] Next, a Bayesian optimization algorithm is used to iteratively solve the single-objective optimization function, generating a set of candidate solutions for the wear resistance of concrete. The decision variable space... It covers multiple dimensions, including the dosage of wear-resistant materials and the concrete mixing speed. First, 50 initial samples were generated using Latin hypercube sampling to cover the range of decision variables. Then, an improved Gaussian process regression was used to establish a mapping relationship between the parameter space and the target space, employing an improved quadratic exponential kernel. , , These are sample points in the decision variable space; The signal variance is used to control the overall scaling of the kernel function; Dimensions of decision variables; For the first The length scale parameter of each decision variable is used to control the weight of that variable's influence on the objective function (the more sensitive a variable is to the objective function, the greater its influence). The smaller the value of ( ) This represents the noise variance, used to characterize the noise during the observation process; For the Kronecker function (when hour, (Otherwise, it is 0). Based on this Gaussian process regression model, the next sampling point is selected using the expected improvement criterion. After a predetermined number of iterations, multiple sets of candidate solutions covering the potential optimal region of the parameter space are generated.

[0045] Then, the constraint satisfaction of the candidate solution set is verified based on the constraint model to obtain a feasible solution set that satisfies all constraints for optimizing concrete wear resistance. Specifically, decision variables from each candidate solution are extracted and substituted into the constructed constraint model to check whether all constraints are satisfied (including wear resistance prediction constraints, material cost constraints, and microstructure parameter constraints, such as "interfacial bond strength"). ",in (This is the threshold for interfacial bonding strength in the constraint model). For candidate solutions that do not meet the constraints, we first try to make small adjustments to the key parameters (such as the amount of wear-resistant material and the stirring time) based on the original solution (for example, increasing the amount of wear-resistant material by 1%-2% and appropriately extending the stirring time). If the constraints are still not met after the adjustments, the solution is eliminated. Finally, we screen out the feasible solution set that meets all constraints to ensure that the solutions analyzed in the subsequent analysis all meet the actual requirements of the engineering.

[0046] Subsequently, multi-objective optimization solution set analysis was performed on the feasible solution set to obtain the Pareto front distribution characteristics of wear resistance performance and material cost. Specifically, a non-dominated sorting algorithm was used to rank the schemes in the feasible solution set: if the scheme... Predicted values ​​of wear resistance And material costs Then it is called a scheme "Domination" scheme Through multiple rounds of non-dominated sorting, all "non-dominated solutions" (i.e., solutions not dominated by any other solution) are selected, and these non-dominated solutions constitute the Pareto front. The distribution characteristics of the Pareto front show a trade-off between "improved wear resistance and increased material cost." Different solutions on the front correspond to different trade-off strategies such as "high wear resistance - high cost" and "low wear resistance - low cost," providing a basis for the selection of the optimal solution in the subsequent stages.

[0047] Next, based on the Pareto front distribution characteristics, a density clustering algorithm is used to identify the optimal solution clusters in the feasible solution set. Then, based on the center position of each cluster, multiple similar candidate solutions contained within the optimal solution cluster are ranked to obtain the ranked set of optimal candidate solutions. Specifically, an improved DBSCAN clustering algorithm is used, with a set neighborhood radius. (Based on the standardized distance between abrasion resistance and material cost, to determine the clustering granularity to suit engineering requirements) and the minimum sample size (e.g., set to 3 to ensure cluster effectiveness), the non-dominated clusters on the Pareto front are de-clustered into multiple clusters. According to the core requirements of concrete abrasion resistance (e.g., "high abrasion resistance preferred"), the cluster with superior abrasion resistance is selected as the optimal cluster; subsequently, the center of this optimal cluster is calculated (denoted as ). (i.e., the statistical center of wear resistance performance and cost of intra-cluster solutions), and expressed by the distance formula: The candidate solutions within the optimal solution cluster are sorted (where, , Considering the wear resistance and material cost of the candidate solutions, , Distance weighting can be adjusted based on core requirements, such as setting it for "high wear resistance priority". ), thus obtaining the sorted set of optimal candidate solutions.

[0048] Finally, based on the Pareto optimality algorithm principle of "no room for improvement," the solution ranked highest is selected as the optimal solution (this solution is closest to the cluster center within the optimal solution cluster and best meets the core requirements of wear resistance performance). Since the Gaussian process regression of the Bayesian optimization above has established a positive mapping model between "decision variables (wear-resistant material dosage, stirring speed, etc.) and the predicted value of wear resistance performance / cost," the optimal solution now corresponds to a specific value in the target space (such as wear resistance performance). ,cost To invert decision variables, the posterior probability distribution of the Gaussian process needs to be utilized: based on the trained Gaussian process model, the target space is locked... , ( , To define the model's prediction error tolerance, the noise variance of the Gaussian process is used. The region is determined, and then a solution is searched in the parameter space that satisfies the condition that "when the parameter is input, the output of the Gaussian process falls into the target region". Specifically, through conditional sampling, under the posterior distribution of the Gaussian process, a solution is generated that satisfies the condition. ( The decision variable sample (corresponding to the optimal solution) is then used for screening in conjunction with constraint models (such as the range of wear-resistant material dosage and engineering limits for mixing speed). Finally, the values ​​of the decision variables that make the Gaussian process output closest to the optimal objective are determined, namely: the optimal ratio of wear-resistant material to concrete material (such as the proportion of wear-resistant material to the total mass of concrete), and the optimal mixing control parameters for both (such as mixing speed, mixing time, and mixing water temperature). These parameters are integrated into an optimized scheme for the wear-resistant performance of concrete and output to the concrete construction control system, thereby enabling adjustments to the wear-resistant material admixture and mixing process during actual construction.

[0049] See Figure 2 This is a schematic diagram of a deep learning-based intelligent optimization system for concrete wear resistance, provided in an embodiment of the present invention. The deep learning-based intelligent optimization system for concrete wear resistance includes: The acquisition module 10 is used to acquire microscopic images of concrete construction material samples, abrasion resistance requirements parameters for concrete construction, and environmental parameters for concrete construction. Extraction module 11 is used to perform cross-scale feature fusion on the microscopic image data through a deep learning network to extract the material microstructure features of concrete. Module 12 is used to construct a constraint model for optimizing the wear resistance performance of concrete based on the wear resistance requirement parameters and the environmental parameters. The prediction module 13 is used to determine the predicted value of the wear resistance of concrete based on the microstructure characteristics of the material and through a pre-trained wear resistance prediction model. The optimization module 14 is used to generate an optimization scheme for the wear resistance performance of concrete based on the predicted wear resistance performance value and the constraint condition model. The optimization scheme includes the optimal ratio of wear resistance material and concrete material and the optimal mixing control parameters of the two.

[0050] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention acquires microscopic images of concrete construction material samples, along with abrasion resistance requirements and environmental parameters. A deep learning network is used to perform cross-scale feature fusion on the microscopic image data, extracting the microstructural features of the concrete material. This feature fusion improves the comprehensiveness and accuracy of the material's microstructural characterization by integrating microstructural information at different scales. A constraint model for optimizing concrete abrasion resistance is constructed based on the abrasion resistance requirements and environmental parameters. This model correlates abrasion resistance requirements with environmental factors, adapting the constraints to actual engineering scenarios. Based on the material's microstructural features, a pre-trained abrasion resistance prediction model determines the predicted abrasion resistance value of the concrete. This process leverages the inherent correlation between microstructural features and abrasion resistance to improve prediction accuracy. Based on the predicted abrasion resistance value and the constraint model, an optimized abrasion resistance performance scheme is generated, including the optimal ratio of abrasion-resistant materials to concrete materials and optimal mixing control parameters, achieving targeted adjustments to abrasion resistance performance. This embodiment of the invention, by deeply integrating material microstructural information and combining the construction environment and abrasion resistance requirements to construct a constraint model and perform performance prediction, solves the problem of insufficient optimization effects in existing technologies, enabling effective optimization of concrete abrasion resistance performance.

[0051] As one example, the extraction module is specifically used for: Multi-scale structural semantic segmentation processing is performed on the microscopic image data to generate a three-level structural feature map of aggregate region, interface transition region and pore region based on the physical scale characteristics of material components. The three-level structural feature maps are processed by parallel multi-scale feature extraction branches in the deep learning network to obtain local feature sets at each scale. Specifically, the geometric morphology and spatial distribution features of the aggregate region are extracted by the first feature extraction branch of the deep learning network, the micro-topology and bond strength features of the interface transition zone are extracted by the second feature extraction branch of the deep learning network, and the size distribution and connectivity features of the pore region are extracted by the third feature extraction branch of the deep learning network. The local feature set is input into the graph neural network fusion unit in the deep learning network to construct a material structure relationship graph with aggregate, interface and pore as nodes, and cross-scale physical interaction is modeled by edge weights to obtain the fused cross-scale features. The fused cross-scale features are subjected to multi-task decoupling processing, and the aggregate-slurry interface topological features and porosity features of concrete materials are separated by an attention mechanism. The topological and porosity characteristics of the separated aggregate-slurry interface are optimized in an anti-wear manner to generate material microstructure characteristics that are geared towards wear resistance. The material microstructure characteristics include the optimized aggregate-slurry interface topological and porosity characteristics.

[0052] As one example, the building module is specifically used for: The wear resistance requirements parameters are quantified in multiple dimensions to identify the core requirements indicators for the wear resistance performance of concrete. Environmental sensitivity analysis was performed on the environmental parameters to identify the key environmental factors affecting the wear resistance of concrete. Based on the core requirement indicators and key environmental factors, a constraint rule base for optimizing the wear resistance of concrete is established. The constraint rule base is transformed into quantifiable constraint functions to construct the constraint model for optimizing the wear resistance of concrete.

[0053] As one example, the prediction module is specifically used for: Feature importance analysis was performed on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete; The key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding to obtain the encoded material microstructure features; The coded material microstructure features are transformed into wear resistance performance correlation vectors through the feature mapping mechanism in the wear resistance performance prediction model. The wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation; The weighted feature representation is calculated using the output layer of the wear resistance prediction model to obtain the predicted value of the wear resistance of concrete.

[0054] As one example, the optimization module is specifically used for: Based on the predicted wear resistance value and the constraint model, a bi-objective optimization function is constructed with the objectives of maximizing wear resistance and minimizing material cost. The weighted summation method is then used to transform the bi-objective function into a single-objective optimization problem of concrete wear resistance. The single-objective optimization function is solved iteratively using the Bayesian optimization algorithm. A mapping relationship between the parameter space and the objective space is established through Gaussian process regression to generate a set of candidate solutions for the wear resistance of concrete. Based on the constraint model, the constraint satisfaction of the candidate scheme set is verified to obtain a feasible solution set that satisfies all the constraint conditions for optimizing the wear resistance of concrete. Multi-objective optimization solution set analysis is performed on the feasible solution set, and the Pareto front distribution characteristics of wear resistance performance and material cost are obtained based on the analysis results; Based on the Pareto front distribution characteristics, the optimal solution cluster in the feasible solution set is identified by density clustering algorithm, and the multiple similar candidate solutions contained in the optimal solution cluster are sorted according to the center position of the solution cluster to obtain the sorted optimal candidate solution set. Based on the Pareto optimality principle, the optimal solution is selected from the sorted set of optimal candidate solutions. By analyzing the parameter space mapping relationship corresponding to the optimal solution, the optimal ratio of the wear-resistant material and the concrete material and the optimal mixing control parameters of the two are obtained by inversion, and the optimized solution for the wear-resistant performance of concrete is output.

[0055] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0056] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization of concrete wear resistance based on deep learning, characterized in that, include: To obtain microscopic images of concrete construction material samples, abrasion resistance requirements for concrete construction, and environmental parameters for concrete construction; The microscopic image data is fused across scales using a deep learning network to extract the microscopic structural features of concrete material. Based on the wear resistance requirement parameters and the environmental parameters, a constraint condition model for optimizing the wear resistance performance of concrete is constructed. Based on the microstructure characteristics of the material, the predicted value of the wear resistance of concrete is determined by a pre-trained wear resistance prediction model. Based on the predicted wear resistance value and the constraint model, an optimization scheme for the wear resistance of concrete is generated. The optimization scheme includes the optimal ratio of wear-resistant material and concrete material and the optimal mixing control parameters of the two. The step of performing cross-scale feature fusion on the microscopic image data using a deep learning network to extract the microstructural features of concrete material includes the following sub-steps: Multi-scale structural semantic segmentation processing is performed on the microscopic image data to generate a three-level structural feature map of aggregate region, interface transition region and pore region based on the physical scale characteristics of material components. The three-level structural feature maps are processed by parallel multi-scale feature extraction branches in the deep learning network to obtain local feature sets at each scale. Specifically, the geometric morphology and spatial distribution features of the aggregate region are extracted by the first feature extraction branch of the deep learning network, the micro-topology and bond strength features of the interface transition zone are extracted by the second feature extraction branch of the deep learning network, and the size distribution and connectivity features of the pore region are extracted by the third feature extraction branch of the deep learning network. The local feature set is input into the graph neural network fusion unit of the deep learning network to construct a material structure relationship graph with aggregate, interface, and pore as nodes. Cross-scale physical interactions are modeled through edge weights to obtain the fused cross-scale features. Specifically, the local feature set is input into the graph neural network (GNN) fusion unit, initially using the local feature vectors of aggregate, interface transition zone, and pore as the three types of nodes in the graph, denoted as... , , ,in This is a 128-dimensional aggregate node feature vector. It is a 64-dimensional interface node feature vector. The feature vectors for 64-dimensional pore nodes are used; then, node connection edges are constructed based on the spatial relationships of each region in the microscopic image; the following formula is used to model cross-scale physical interactions: , For nodes With nodes The edge weight is used to characterize the strength of the association between two nodes; , These are the weighting coefficients; The cosine similarity between the features of two nodes reflects the degree of correlation at the feature level. The spatial distance decay function, This represents the shortest spatial distance between two regions in the microscopic image. Based on concrete material component diffusion experiments, the attenuation effect of simulated distance on physical interaction is determined to be a constant; the features of each node are updated through the GNN message passing mechanism, and the fused cross-scale features are finally output. The fused cross-scale features are subjected to multi-task decoupling processing, and the aggregate-slurry interface topological features and porosity features of concrete materials are separated through an attention mechanism. Specifically, the cross-scale features are input into two parallel decoupled branches: interface topological feature extraction and porosity feature extraction. An attention module is set before each branch to enhance the feature channels relevant to the target task. The attention weight calculation formula is as follows: , For the first The attention weights for each feature channel, with values ​​ranging from [0,1]; For the first The correlation score between each channel feature and the target task is calculated through the mutual information between the feature and the task label. The interface topology task label is the measured value of concrete interface roughness, and the porosity task label is the measured value of concrete pore volume fraction. The sum of the exponential scores of all feature channels is used to normalize the weights. Through this attention mechanism, the interface topology branch strengthens the feature channels related to texture and thickness, and the porosity branch strengthens the feature channels related to the number and size of pores, respectively outputting the aggregate-slurry interface topology features and porosity features. The topological and porosity characteristics of the separated aggregate-slurry interface are subjected to adversarial optimization to generate material microstructure features oriented towards wear resistance. These material microstructure features include the optimized aggregate-slurry interface topological and porosity characteristics. An adversarial optimization network is constructed, comprising a generator. With discriminator Generator Using the separated interface topology and porosity features as input, the optimized features are output through three fully connected layers. Discriminator enter Microstructural characteristics of truly wear-resistant concrete Output the similarity score between the two; design the loss function: , For expectation operators; This represents the probability that the discriminator recognizes truly superior features. The discriminator's recognition probability of the optimized features output by the generator; The regularization coefficient is used. To optimize the L2 norm of the features before and after training, the optimization magnitude is constrained; the generator is trained alternately. With discriminator The process continues until the preset loss function converges, and finally outputs the material microstructure features, including the optimized aggregate-slurry interface topology and porosity features.

2. The intelligent optimization method for concrete wear resistance based on deep learning as described in claim 1, characterized in that, The process of constructing a constraint model for optimizing the wear resistance performance of concrete based on the wear resistance requirement parameters and the environmental parameters includes the following sub-steps: The wear resistance requirements parameters are quantified in multiple dimensions to identify the core requirements indicators for the wear resistance performance of concrete. Environmental sensitivity analysis was performed on the environmental parameters to identify the key environmental factors affecting the wear resistance of concrete. Based on the core requirement indicators and key environmental factors, a constraint rule base for optimizing the wear resistance of concrete is established. The constraint rule base is transformed into quantifiable constraint functions to construct the constraint model for optimizing the wear resistance of concrete.

3. The intelligent optimization method for concrete wear resistance based on deep learning as described in claim 2, characterized in that, The step of determining the predicted abrasion resistance of concrete based on the microstructure characteristics of the material using a pre-trained abrasion resistance prediction model includes the following sub-steps: Feature importance analysis was performed on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete; The key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding to obtain the encoded material microstructure features; The coded material microstructure features are transformed into wear resistance performance correlation vectors through the feature mapping mechanism in the wear resistance performance prediction model. The wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation; The weighted feature representation is calculated using the output layer of the wear resistance prediction model to obtain the predicted value of the wear resistance of concrete.

4. The intelligent optimization method for concrete wear resistance based on deep learning as described in claim 3, characterized in that, Based on the predicted wear resistance value and the constraint model, an optimized scheme for the wear resistance of concrete is generated. The optimized scheme includes the optimal ratio of wear-resistant material to concrete material and the optimal mixing control parameters for both, and includes the following sub-steps: Based on the predicted wear resistance value and the constraint model, a bi-objective optimization function is constructed with the objectives of maximizing wear resistance and minimizing material cost. The weighted summation method is then used to transform the bi-objective function into a single-objective optimization problem of concrete wear resistance. The Bayesian optimization algorithm is used to iteratively solve the single-objective optimization function of the single-objective optimization problem. A mapping relationship between the parameter space and the objective space is established through Gaussian process regression to generate a set of candidate solutions for the wear resistance performance of concrete. Based on the constraint model, the constraint satisfaction of the candidate scheme set is verified to obtain a feasible solution set that satisfies all the constraint conditions for optimizing the wear resistance of concrete. Multi-objective optimization solution set analysis is performed on the feasible solution set, and the Pareto front distribution characteristics of wear resistance performance and material cost are obtained based on the analysis results; Based on the Pareto front distribution characteristics, the optimal solution cluster in the feasible solution set is identified by density clustering algorithm, and the multiple similar candidate solutions contained in the optimal solution cluster are sorted according to the center position of the solution cluster to obtain the sorted optimal candidate solution set. Based on the Pareto optimality principle, the optimal solution is selected from the sorted set of optimal candidate solutions. By analyzing the parameter space mapping relationship corresponding to the optimal solution, the optimal ratio of the wear-resistant material and the concrete material and the optimal mixing control parameters of the two are obtained by inversion, and the optimized solution for the wear-resistant performance of concrete is output.

5. A deep learning-based intelligent optimization system for the wear resistance of concrete, characterized in that, include: The acquisition module is used to acquire microscopic images of concrete construction material samples, abrasion resistance requirements parameters for concrete construction, and environmental parameters for concrete construction. The extraction module is used to perform cross-scale feature fusion on the microscopic image data through a deep learning network to extract the material microstructure features of concrete. A construction module is used to construct a constraint model for optimizing the wear resistance performance of concrete based on the wear resistance requirement parameters and the environmental parameters. The prediction module is used to determine the predicted value of the wear resistance of concrete based on the microstructure characteristics of the material and through a pre-trained wear resistance prediction model. An optimization module is used to generate an optimization scheme for the wear resistance performance of concrete based on the predicted wear resistance performance value and the constraint condition model. The optimization scheme includes the optimal ratio of wear resistance material and concrete material and the optimal mixing control parameters of the two. The extraction module is specifically used for: Multi-scale structural semantic segmentation processing is performed on the microscopic image data to generate a three-level structural feature map of aggregate region, interface transition region and pore region based on the physical scale characteristics of material components. The three-level structural feature maps are processed by parallel multi-scale feature extraction branches in the deep learning network to obtain local feature sets at each scale. Specifically, the geometric morphology and spatial distribution features of the aggregate region are extracted by the first feature extraction branch of the deep learning network, the micro-topology and bond strength features of the interface transition zone are extracted by the second feature extraction branch of the deep learning network, and the size distribution and connectivity features of the pore region are extracted by the third feature extraction branch of the deep learning network. The local feature set is input into the graph neural network fusion unit of the deep learning network to construct a material structure relationship graph with aggregate, interface, and pore as nodes. Cross-scale physical interactions are modeled through edge weights to obtain the fused cross-scale features. Specifically, the local feature set is input into the graph neural network (GNN) fusion unit, initially using the local feature vectors of aggregate, interface transition zone, and pore as the three types of nodes in the graph, denoted as... , , ,in This is a 128-dimensional aggregate node feature vector. It is a 64-dimensional interface node feature vector. The feature vectors for 64-dimensional pore nodes are used; then, node connection edges are constructed based on the spatial relationships of each region in the microscopic image; the following formula is used to model cross-scale physical interactions: , For nodes With nodes The edge weight is used to characterize the strength of the association between two nodes; , These are the weighting coefficients; The cosine similarity between the features of two nodes reflects the degree of correlation at the feature level. The spatial distance decay function, This represents the shortest spatial distance between two regions in the microscopic image. Based on concrete material component diffusion experiments, the attenuation effect of simulated distance on physical interaction is determined to be a constant; the features of each node are updated through the GNN message passing mechanism, and the fused cross-scale features are finally output. The fused cross-scale features are subjected to multi-task decoupling processing, and the aggregate-slurry interface topological features and porosity features of concrete materials are separated through an attention mechanism. Specifically, the cross-scale features are input into two parallel decoupled branches: interface topological feature extraction and porosity feature extraction. An attention module is set before each branch to enhance the feature channels relevant to the target task. The attention weight calculation formula is as follows: , For the first The attention weights for each feature channel, with values ​​ranging from [0,1]; For the first The correlation score between each channel feature and the target task is calculated through the mutual information between the feature and the task label. The interface topology task label is the measured value of concrete interface roughness, and the porosity task label is the measured value of concrete pore volume fraction. The sum of the exponential scores of all feature channels is used to normalize the weights. Through this attention mechanism, the interface topology branch strengthens the feature channels related to texture and thickness, and the porosity branch strengthens the feature channels related to the number and size of pores, respectively outputting the aggregate-slurry interface topology features and porosity features. The topological and porosity characteristics of the separated aggregate-slurry interface are subjected to adversarial optimization to generate material microstructure features oriented towards wear resistance. These material microstructure features include the optimized aggregate-slurry interface topological and porosity characteristics. An adversarial optimization network is constructed, comprising a generator. With discriminator Generator Using the separated interface topology and porosity features as input, the optimized features are output through three fully connected layers. Discriminator enter Microstructural characteristics of truly wear-resistant concrete Output the similarity score between the two; design the loss function: , For expectation operators; This represents the probability that the discriminator recognizes truly superior features. The discriminator's recognition probability of the optimized features output by the generator; The regularization coefficient is used. To optimize the L2 norm of the features before and after training, the optimization magnitude is constrained; the generator is trained alternately. With discriminator The process continues until the preset loss function converges, and finally outputs the material microstructure features, including the optimized aggregate-slurry interface topology and porosity features.

6. The intelligent optimization system for concrete wear resistance based on deep learning as described in claim 5, characterized in that, The building module is specifically used for: The wear resistance requirements parameters are quantified in multiple dimensions to identify the core requirements indicators for the wear resistance performance of concrete. Environmental sensitivity analysis was performed on the environmental parameters to identify the key environmental factors affecting the wear resistance of concrete. Based on the core requirement indicators and key environmental factors, a constraint rule base for optimizing the wear resistance of concrete is established. The constraint rule base is transformed into quantifiable constraint functions to construct the constraint model for optimizing the wear resistance of concrete.

7. The intelligent optimization system for concrete wear resistance based on deep learning as described in claim 6, characterized in that, The prediction module is specifically used for: Feature importance analysis was performed on the microstructure characteristics of the material to screen out the key feature subset that dominates the wear resistance of concrete; The key feature subset is input into the feature encoding layer of the pre-trained wear resistance prediction model for feature encoding to obtain the encoded material microstructure features; The coded material microstructure features are transformed into wear resistance performance correlation vectors through the feature mapping mechanism in the wear resistance performance prediction model. The wear resistance performance correlation vector is dynamically weighted to generate a weighted feature representation; The weighted feature representation is calculated using the output layer of the wear resistance prediction model to obtain the predicted value of the wear resistance of concrete.

8. The intelligent optimization system for concrete wear resistance based on deep learning as described in claim 7, characterized in that, The optimization module is specifically used for: Based on the predicted wear resistance value and the constraint model, a bi-objective optimization function is constructed with the objectives of maximizing wear resistance and minimizing material cost. The weighted summation method is then used to transform the bi-objective function into a single-objective optimization problem of concrete wear resistance. The Bayesian optimization algorithm is used to iteratively solve the single-objective optimization function of the single-objective optimization problem. A mapping relationship between the parameter space and the objective space is established through Gaussian process regression to generate a set of candidate solutions for the wear resistance performance of concrete. Based on the constraint model, the constraint satisfaction of the candidate scheme set is verified to obtain a feasible solution set that satisfies all the constraint conditions for optimizing the wear resistance of concrete. Multi-objective optimization solution set analysis is performed on the feasible solution set, and the Pareto front distribution characteristics of wear resistance performance and material cost are obtained based on the analysis results; Based on the Pareto front distribution characteristics, the optimal solution cluster in the feasible solution set is identified by density clustering algorithm, and the multiple similar candidate solutions contained in the optimal solution cluster are sorted according to the center position of the solution cluster to obtain the sorted optimal candidate solution set. Based on the Pareto optimality principle, the optimal solution is selected from the sorted set of optimal candidate solutions. By analyzing the parameter space mapping relationship corresponding to the optimal solution, the optimal ratio of the wear-resistant material and the concrete material and the optimal mixing control parameters of the two are obtained by inversion, and the optimized solution for the wear-resistant performance of concrete is output.

Images