Durability testing method of recycled aggregate hydraulic concrete based on neural network

Abstract

The present invention relates to the field of artificial intelligence and data processing technology, and specifically to a method for detecting the durability of recycled aggregate hydraulic concrete based on a neural network, which is as follows: collecting recycled aggregate hydraulic concrete data and manually labeling the recycled aggregate hydraulic concrete based on the collected data; constructing a neural network based on adaptive topological connections as a recycled aggregate hydraulic concrete durability detection model, inputting the collected data into the model to train the model until a preset stop iteration condition is met and the model training is completed; inputting new recycled aggregate hydraulic concrete data into the trained model to obtain a durability classification result. The present invention uses a recycled aggregate hydraulic concrete durability detection model constructed by a neural network based on adaptive topological connections to handle complex relationships at multiple scales and dimensions, improve the stability of the detection results, and simultaneously improve detection efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data processing technology, and in particular to a method for detecting the durability of recycled aggregate hydraulic concrete based on a neural network. Background Art

[0002] As an environmentally friendly material, recycled aggregate is widely used in concrete. However, due to the diversity of its sources and differences in its physicochemical properties, the durability of concrete exhibits complex multi-scale nonlinear relationships. Traditional concrete durability assessment methods often rely on simple linear regression or fixed feature combinations. These methods cannot effectively capture the nonlinear characteristics in the data, nor can they cope with the interactions between different aggregate types, mixing ratios, and other influencing factors. The performance of recycled aggregate hydraulic concrete is affected by many factors, including the source of the aggregate, mixing ratio, compressive strength, water absorption, chloride ion permeability, etc. These factors may exhibit different durability characteristics under different environmental conditions. For example, natural aggregate and recycled aggregate have different physical and chemical properties, which directly affect the mechanical properties and durability of concrete. At the same time, indicators such as the number of freeze-thaw cycles and carbonation depth of different concrete samples also play a vital role in the durability of concrete. However, traditional methods can only process single attributes of these data and ignore the complex interactions between them.

[0003] The defects of the existing recycled aggregate concrete durability testing technology are as follows:

[0004] Most existing neural networks use a fixed fully connected topology and are unable to dynamically adjust the connection strength, resulting in the model having difficulty capturing the complex nonlinear relationships between features for data with multi-scale and multi-dimensional features. Conventional initialization methods (such as random initialization) fail to consider the local structure of the data, which can easily cause the neural network to fall into local optimal solutions during training, reducing the stability of training. When processing large-scale data, traditional neural networks have high computational costs due to their fixed topology and a large number of redundant connections, making it difficult to improve computational efficiency while ensuring accuracy. Existing technologies usually use a single loss function in multi-objective tasks, which cannot effectively balance the conflicts between multiple tasks, resulting in poor model optimization effects. When the network topology changes, the traditional backpropagation algorithm is prone to discontinuity in the gradient propagation path, resulting in training oscillations and reduced convergence efficiency.

[0005] Therefore, the present invention proposes a method for detecting the durability of recycled aggregate hydraulic concrete based on a neural network to solve the above problems. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention develops a method for detecting the durability of recycled aggregate hydraulic concrete based on a neural network. The present invention constructs a recycled aggregate hydraulic concrete durability detection model based on a neural network with adaptive topological connections, which can handle complex relationships at multiple scales and dimensions, improve the stability of the detection results, and at the same time improve the detection efficiency and accuracy.

[0007] The technical solution to the technical problem of the present invention is a method for detecting the durability of recycled aggregate hydraulic concrete based on a neural network, comprising the following steps:

[0008] S1. Data collection: Collect recycled aggregate hydraulic concrete data and manually label the recycled aggregate hydraulic concrete based on the collected data;

[0009] S2. Constructing a neural network based on adaptive topological connections as a durability testing model for recycled aggregate hydraulic concrete, inputting the collected data into the model for training, and obtaining a durability classification result for the recycled aggregate hydraulic concrete based on the model's analysis of the recycled aggregate hydraulic concrete data, until a preset stopping condition is met to complete the model training;

[0010] S3. Input the new recycled aggregate hydraulic concrete data into the trained recycled aggregate hydraulic concrete durability performance detection model to obtain the durability performance classification result of the aggregate hydraulic concrete.

[0011] S1 is as follows:

[0012] (1) The collected data of recycled aggregate hydraulic concrete include but are not limited to aggregate source, mixing ratio, compressive strength, water absorption, chloride ion permeability, number of freeze-thaw cycles and carbonation depth;

[0013] The methods of collecting data include field collection, laboratory collection, and collection by sensors and detection systems;

[0014] (2) Manually labeled categories include: high durability, medium durability, and low durability.

[0015] A durability testing model for recycled aggregate hydraulic concrete was constructed based on a neural network with adaptive topological connections. The durability testing model for recycled aggregate hydraulic concrete was trained. The training process is as follows:

[0016] For a neural network, the connection strength of its topological structure is calculated, the neural network is initialized, and the connection of neurons in the neural network topological structure is adjusted;

[0017] Adopting multi-task driven adaptive loss function to balance multi-objective optimization;

[0018] In the back propagation process of the neural network, the path continuity constraint optimization method is adopted;

[0019] A topology-robust sample weighting method is used to calculate sample weights, and then a topology-driven parameter sharing strategy is used to share the weights.

[0020] A progressive neural network topology freezing method is used to freeze some layers of the neural network.

[0021] The connectivity enhancement, initialization and connection adjustment of the neural network are as follows:

[0022] (1) A dynamic feature fusion topology connection mechanism is used to adjust the connection strength of the neural network topology structure. Specifically, the connection weights are dynamically adjusted by the mutual information entropy and gradient sensitivity between features to achieve adaptive feature interaction.

[0023] (2) The neural network uses a weight initialization method based on local density perception, combining the feature space cluster center and local density distribution to generate initial weights;

[0024] (3) A staged topology reconstruction strategy is adopted to dynamically add and delete neuron connections according to the feature contribution of recycled aggregate hydraulic concrete data to adjust the neuron connections of the neural network topology structure, as follows:

[0025] (3-1) Calculate the characteristic contribution of recycled aggregate hydraulic concrete data;

[0026] (3-2) Adjust the neural network topology. Each layer of the neural network is sorted according to the feature contribution, retaining the top M% of high-contribution features and removing the rest.

[0027] For the retained features with high contribution, new connections are inserted in the areas where the feature contribution is lower than the threshold to obtain a dynamic retopology structure.

[0028] The adaptive loss function is as follows:

[0029] The recycled aggregate hydraulic concrete durability performance testing model uses a multi-task driven adaptive loss function. The adaptive loss function combines the durability classification task with the index regression task and dynamically adjusts the loss weight. This allows the recycled aggregate hydraulic concrete durability performance testing model to automatically focus on difficult-to-optimize tasks and balance multi-objective optimization tasks.

[0030] Adaptive loss function The calculation formula is as follows:

[0031] ,

[0032] ,

[0033] in, is the loss weight coefficient of the classification task, To calculate the classification loss using weighted cross entropy loss, For the The weight of the samples; is the category weight, For the The samples belong to The true labels of the categories, For the The samples belong to The predicted probability of each category, is the logarithmic function with base 10, is the total number of samples, is the total number of categories, is the number of regression tasks, For the The loss weight of the iteration, is the regression loss.

[0034] The path continuity constraint optimization used in the directional propagation process of the neural network is as follows:

[0035] The topology change penalty term is used in the gradient calculation of the neural network back propagation. The calculation formula is as follows:

[0036] ,

[0037] in, is the gradient corrected by the topology change penalty term; is the penalty coefficient; is the indicator function, Indicates when the connection strength Below the preset threshold Take 1 when it is, otherwise take 0; is a symbolic function; For the neural network Layer neurons to the next layer The weights of the neuron connections.

[0038] The calculation of sample weights and weight sharing are as follows:

[0039] (1) Calculate sample weight:

[0040] (1-1) The matrix, expected value and variance of the connection strength of the sample, using the Frobenius norm, calculate the topological consistency score of the sample during the forward propagation of the neural network;

[0041] (1-2) Dynamically adjust sample weights based on their topological consistency scores;

[0042] (2) Weight sharing:

[0043] A topology-driven parameter sharing strategy is used to share weights. For connection groups with similar contributions, connections with similar contributions are forced to share parameters.

[0044] (2-1) Contribution feature construction: All connections of the layer, extract their contribution scores , construct a one-dimensional feature vector , , Represents the neural network Layer neurons to the next layer The contribution score of each neuron connection, Represents the neural network The contribution score of all connections in the layer, the lower limit of the contribution score is , the upper limit of contribution score is ;

[0045] (2-2) Contribution grouping and clustering:

[0046] Use K-means algorithm to analyze one-dimensional feature vectors Perform contribution grouping and clustering, and then determine the objective function that minimizes the intra-cluster variance, expressed as ,in, represents the objective function that minimizes the intra-cluster variance, Indicates the shared clusters, Indicates the shared clusters, Indicates the preset number of clusters, Indicates the The centroid of each cluster;

[0047] (2-3) Weight sharing:

[0048] ,

[0049] in, express The sharing weights of the sharing groups, Indicates that it belongs to connections between neurons that share a cluster.

[0050] The neural network topology freezing is as follows:

[0051] A progressive neural network topology freezing method is adopted. When the network topology structure change rate is lower than a preset threshold, the layer of the neural network is frozen. The network topology structure change rate is measured by the connection change rate or the weight change rate.

[0052] The training process of the recycled aggregate hydraulic concrete durability performance detection model is repeated until the preset stopping iteration condition is met, and a trained recycled aggregate hydraulic concrete durability performance detection model is obtained.

[0053] S3 is as follows:

[0054] The new recycled aggregate hydraulic concrete data is input into the trained recycled aggregate hydraulic concrete durability performance detection model. In the forward propagation stage, the input data passes through each layer of the neural network in turn, and after calculation in each layer, it passes through the output layer composed of the Softmax activation function, and finally outputs the probability value of each durability performance classification. The category with the highest probability value is the durability performance category of the recycled aggregate hydraulic concrete. The durability performance categories include high durability, medium durability and low durability.

[0055] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0056] (1) The present invention proposes a dynamic feature fusion topology connection mechanism, which dynamically adjusts the connection weights by the mutual information entropy and gradient sensitivity between features, so that the neural network can adaptively enhance the interaction between key features and suppress irrelevant connections, solving the problem that the traditional neural network fixed fully connected topology cannot effectively handle multi-scale nonlinear relationships, and improving the network's modeling ability for complex data relationships; The present invention proposes a weight initialization method based on local density perception, which combines the clustering center of the feature space with the local density distribution to generate initial weights, so that the initial state of the network is biased towards the high-density cluster area, thereby enhancing the sensitivity to local clustering features, avoiding the local optimal problem caused by the traditional initialization method, and improving the stability and efficiency of the model in the early stage of training; The present invention proposes a staged topology reconstruction strategy, dynamically adding and deleting neuronal connections to adapt to the characteristics of feature importance changing with the training stage. In this way, the model can focus on key feature paths while reducing computational costs. While ensuring model accuracy, the amount of computation is greatly reduced by pruning redundant connections;

[0057] (2) This paper proposes a multi-task driven adaptive loss function, which combines the durability classification task with the index regression task and dynamically adjusts the loss weight of each task, thereby improving the collaborative optimization capability of multi-objective tasks, effectively alleviating the conflict problem between multi-objective tasks, and improving the model's optimization capability for different tasks;

[0058] (3) In the back-propagation process, the present invention adopts path continuity constraint optimization and corrects the gradient through the topology change penalty term to avoid the gradient oscillation problem caused by topology adjustment, reduce the training oscillation caused by the dynamic change of the topology structure, and improve the stability and convergence of network training;

[0059] (4) The present invention calculates the topological consistency score of the samples and weights the samples to prevent abnormal samples in extreme environments from overfitting the model training, effectively avoiding the interference of local abnormal samples in the data on the training, and improving the generalization ability and accuracy of the model; the present invention proposes a topology-driven parameter sharing strategy, which reduces redundant parameters and the complexity of the model by sharing parameters of connection groups with similar contribution, making the model lightweight and suitable for large-scale data training, while ensuring the expression ability of important connections.

[0060] In summary, the present invention systematically solves the problems of insufficient modeling capabilities, low training efficiency, and multi-task conflicts in traditional neural networks through innovative mechanisms such as dynamic topology adjustment, enhanced training stability, and multi-task collaborative optimization. It achieves lightweight design while improving model accuracy and is suitable for industrial-grade application needs in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0062] Figure 1 Schematic diagram of the method of the present invention.

[0063] Figure 2 A comparison chart of the classification performance images of the dynamic linking mechanism used in the present invention and conventional technology.

[0064] Figure 3 This is a performance stability comparison chart of the local density-aware initialization method used in the present invention and conventional technology.

[0065] Figure 4 Schematic diagram of the impact of the topology adjustment strategy used in the present invention on computing efficiency.

[0066] Figure 5 A three-dimensional surface diagram of the multi-scale fusion effect of the adaptive topology mechanism used in the present invention and the conventional technology based on the material mechanics theoretical model. DETAILED DESCRIPTION

[0067] To clearly illustrate the technical features of this solution, the present invention is described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and configurations of specific examples are described below.

[0068] Example 1

[0069] like Figures 1 to 5 As shown, a method for detecting the durability of recycled aggregate hydraulic concrete based on a neural network includes the following steps:

[0070] S1. Data collection: Collect recycled aggregate hydraulic concrete data and manually label the recycled aggregate hydraulic concrete based on the collected data;

[0071] S2. Constructing a neural network based on adaptive topological connections as a durability testing model for recycled aggregate hydraulic concrete, inputting the collected data into the model for training, and obtaining a durability classification result for the recycled aggregate hydraulic concrete based on the model's analysis of the recycled aggregate hydraulic concrete data, until a preset stopping condition is met to complete the model training;

[0072] S3. Input the new recycled aggregate hydraulic concrete data into the trained recycled aggregate hydraulic concrete durability performance detection model to obtain the durability performance classification result of the aggregate hydraulic concrete.

[0073] Specific implementation method: S1 is as follows:

[0074] (1) The collected data of recycled aggregate hydraulic concrete include but are not limited to aggregate source, mixing ratio, compressive strength, water absorption, chloride ion permeability, number of freeze-thaw cycles and carbonation depth;

[0075] Among them, the source of aggregate is categorical data, indicating whether the aggregate used is natural aggregate or recycled aggregate; the mixing ratio is continuous data, indicating the mixing ratio of different aggregates with cement, water and other components; the compressive strength is continuous data, indicating the strength of concrete in a compression test, usually expressed as MPa value, reflecting the basic mechanical properties of concrete; the water absorption rate is continuous data, indicating the water absorption capacity of concrete; the chloride ion permeability is continuous data, indicating the ability of concrete to penetrate chloride ions; the number of freeze-thaw cycles is continuous data, indicating the durability of concrete under freeze-thaw conditions, that is, the number of freeze-thaw cycles that concrete can withstand; the carbonation depth is continuous data, indicating the depth of carbon dioxide erosion on the concrete surface;

[0076] In addition, the data of recycled aggregate hydraulic concrete may also include particle size distribution, aggregate density, aggregate shape, water-binder ratio, cement type, curing time, temperature conditions, relative humidity, alkali content, chemical additive type and dosage, etc.

[0077] Among them, particle size distribution is continuous data, indicating the size distribution of aggregate particles; aggregate density is continuous data, indicating that the density of aggregate is related to the overall quality and strength of concrete; aggregate shape is categorical data, indicating the shape of aggregate; water-cement ratio is continuous data, indicating the mass ratio of water to cement; cement type is categorical data, indicating the type of cement used; curing time is continuous data, indicating the curing time of concrete; temperature condition is continuous data, indicating the curing effect and later performance of concrete under different temperature conditions; relative humidity is continuous data, indicating the ambient humidity during the concrete curing process; alkali content is continuous data, indicating the alkali content in cement and aggregate; chemical additive type and dosage are categorical and continuous data, including the type and dosage of chemical additives such as air entraining agent, water reducer, and retarder;

[0078] It should be noted that this embodiment is only intended to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0079] The methods of collecting data include field collection, laboratory collection, and collection by sensors and detection systems;

[0080] The experimental data collected on-site are obtained through actual sampling, including test samples of concrete at the construction site or in laboratory conditions; laboratory collection uses specialized instruments and equipment (such as compression testing machines, water absorption meters, chloride ion permeameters, etc.) to conduct comprehensive performance tests on concrete samples; sensors and monitoring systems use sensors (such as vibration sensors, temperature sensors, etc.) and online monitoring systems to collect dynamic data of concrete samples in actual use.

[0081] (2) Manually labeled categories include: high durability, medium durability, and low durability.

[0082] Among them, high durability is suitable for concrete that is used for a long time and exposed to extreme environments, medium durability is suitable for use in general environments, and low durability is suitable for use in harsher environmental conditions.

[0083] In a specific embodiment, a recycled aggregate hydraulic concrete durability performance detection model is constructed based on a neural network with adaptive topological connections, and the recycled aggregate hydraulic concrete durability performance detection model is trained. The training process is as follows:

[0084] For a neural network, the connection strength of its topological structure is calculated, the neural network is initialized, and the connection of neurons in the neural network topological structure is adjusted;

[0085] Adopting multi-task driven adaptive loss function to balance multi-objective optimization;

[0086] In the back propagation process of the neural network, the path continuity constraint optimization method is adopted;

[0087] A topology-robust sample weighting method is used to calculate sample weights, and then a topology-driven parameter sharing strategy is used to share the weights.

[0088] A progressive neural network topology freezing method is used to freeze some layers of the neural network.

[0089] In a specific embodiment, the connectivity enhancement, initialization and connection adjustment of the neural network are as follows:

[0090] (1) A dynamic feature fusion topological connection mechanism is used to adjust the connection strength of the neural network topology. Specifically, the connection weights are dynamically adjusted by the mutual information entropy and gradient sensitivity between features to perform adaptive feature interaction. This can enable the recycled aggregate hydraulic concrete durability performance detection model to adaptively enhance key feature interactions, suppress irrelevant connections, and solve the problem of insufficient multi-scale nonlinear relationship modeling of recycled aggregate hydraulic concrete data.

[0091] The calculation formula is as follows:

[0092] ,

[0093] in, For the neural network Layer neurons to the next layer The connection strength of neurons is used to characterize the dynamic interaction between features and to determine whether to apply gradient penalty terms during back propagation.

[0094] is the mutual information entropy, and the features are calculated and The degree of statistical dependence during training, for multi-source heterogeneous data, such as categorical aggregate sources and continuous compressive strength, is captured through mutual information entropy to capture the dynamic correlation between cross-scale features, addressing the shortcomings of traditional linear correlation metrics;

[0095] For the neural network Layer The feature input of each neuron;

[0096] For the neural network Layer The feature input of each neuron;

[0097] is the variance function, which measures the dispersion of the characteristic distribution and reflects the variability of the characteristic. It is applicable to the distribution differences of different aggregate characteristics in the data.

[0098] The representation variance is normalized to eliminate the feature scale differences and improve the comparability of cross-scale feature interactions;

[0099] is the Sigmoid function, which constrains the connection strength to the interval (0,1);

[0100] is the gradient sensitivity adjustment factor, which controls the weight of the influence of gradient change on connection strength. Set to 0.01;

[0101] is the loss function of the neural network;

[0102] For the neural network Layer neurons to the next layer The connection weights of neurons, Characterize the importance of the connection in backpropagation and dynamically enhance the modeling capabilities of key feature pathways in the current training phase;

[0103] (2) The neural network adopts a weight initialization method based on local density perception, combining the feature space cluster center and the local density distribution to generate the initial weight, which can make the weight biased towards the high-density cluster area, improve the sensitivity to the clustering characteristics, and retain randomness to avoid falling into the local optimum, thereby enhancing the model's ability to capture the local structure of the data;

[0104] The calculation formula is as follows:

[0105] ,

[0106] in, For the neural network Layer neurons to the next layer The initial connection weights of neurons;

[0107] is the number of feature clusters obtained by the DBSCAN clustering algorithm;

[0108] For the The local density of a cluster is calculated as follows:

[0109] ,

[0110] in, For the The sample feature vector of each cluster, is the L2 norm, For the The cluster centers of the clusters, In order to control the Gaussian kernel width of local density calculation and adapt to the sparsity and local aggregation characteristics of feature distribution in recycled aggregate data, Set to 0.1;

[0111] The cluster density weight term is used to enhance the weight distribution of high-density cluster areas and improve the sensitivity to local clustering features;

[0112] is the maximum local density value among all clusters, is the local density of each cluster;

[0113] For the The cluster in The mean of each feature dimension can make the initial state of the model biased towards the true distribution of the data;

[0114] is a small random perturbation term, subject to Uniform distribution of intervals;

[0115] (3) A staged topology reconstruction strategy is adopted to dynamically add and delete neuron connections according to the feature contribution of recycled aggregate hydraulic concrete data to adjust the neuron connections of the neural network topology structure, as follows:

[0116] (3-1) Calculate the characteristic contribution of recycled aggregate hydraulic concrete data;

[0117] The feature contribution is calculated as follows:

[0118] ,

[0119] in, For the neural network Layer neurons to the next layer The contribution score of each neuron connection combines the absolute value of the weight and the expected value of the gradient to reflect the importance of the connection;

[0120] It is the moving average expectation calculation;

[0121] is the gradient expectation term, For the neural network Layer neurons to the next layer The absolute value of the weight of the neuron connection is used to measure the static importance of the connection weight and reflect the inherent strength of the feature interaction. is the loss function of the neural network Weight The gradient of the loss function Specifically, it is an adaptive loss function;

[0122] (3-2) Adjust the neural network topology. Each layer of the neural network is sorted according to the feature contribution, retaining the top M% of high-contribution features and removing the rest.

[0123] For the retained high-contribution features, new connections are inserted in the areas where the feature contribution is lower than the threshold to obtain a dynamic retopology structure;

[0124] The calculation formula for the new connection inserted is as follows:

[0125] ,

[0126] in, For the neural network The weights of the newly inserted connections in the layer, For the The set of high-contributing connections in the layer, for Middle High contribution connections, is the scaling factor for the new connection weight.

[0127] In a specific implementation, the adaptive loss function is as follows:

[0128] The recycled aggregate hydraulic concrete durability performance testing model uses a multi-task driven adaptive loss function. The adaptive loss function combines the durability classification task with the index regression task and dynamically adjusts the loss weight. This allows the recycled aggregate hydraulic concrete durability performance testing model to automatically focus on difficult-to-optimize tasks and balance multi-objective optimization tasks.

[0129] Adaptive loss function The calculation formula is as follows:

[0130] ,

[0131] ,

[0132] in, is the loss weight coefficient of the classification task, Set to 0.2, To calculate the classification loss using weighted cross entropy loss, For the The weight of the samples, is the category weight, For the The samples belong to The true labels of the categories, For the The samples belong to The predicted probability of each category, is the logarithmic function with base 10, is the total number of samples, is the total number of categories, is the number of regression tasks, For the The loss weight of the iteration, It can be dynamically adjusted with the progress of training, automatically assign weights according to task difficulty, and alleviate optimization conflicts caused by multi-objective coupling. is the regression loss;

[0133] No. The loss weight of the iteration Dynamic adjustment with training progress and automatic weight allocation according to task difficulty can alleviate the optimization conflict caused by multi-objective coupling. The calculation formula is as follows:

[0134] ,

[0135] in, For the Return tasks in front The root mean square error of the iterations, For the Return tasks in front The root mean square error of the iterations, is the temperature coefficient, Set to 0.2;

[0136] For the The regression task is calculated using Huber loss, and the calculation formula is as follows:

[0137] ,

[0138] in, is the prediction error, , For the The true value of the regression task, For the The predicted value of a regression task; is the robustness threshold, Set to 0.2.

[0139] In a specific embodiment, the path continuity constraint optimization used in the directional propagation process of the neural network is as follows:

[0140] The topology change penalty term is used in the gradient calculation of the neural network back propagation. The calculation formula is as follows:

[0141] ,

[0142] in, is the gradient corrected by the topology change penalty term; is the penalty coefficient, Set to 0.2; is the indicator function, Indicates when the connection strength Below the preset threshold Take 1 when it is, otherwise take 0; is a symbolic function; For the neural network Layer neurons to the next layer The weights of the connections between neurons; The overall correction term suppresses the parameter update amplitude of the connection to be removed and reduces the gradient oscillation caused by topological mutation.

[0143] In a specific implementation, the calculation of sample weights and weight sharing are as follows:

[0144] (1) Calculate sample weight:

[0145] (1-1) The matrix, expected value and variance of the connection strength of the sample, using the Frobenius norm, calculate the topological consistency score of the sample during the forward propagation of the neural network;

[0146] The formula for calculating the consistency score of a sample is as follows:

[0147] ,

[0148] in, For the The topological consistency score of the samples, is the number of layers involved in topological consistency calculation, For the neural network Layer to layer The connection strength matrix of samples, For the neural network The expected value of the layer connection strength, is the Frobenius norm, For the neural network the variance of layer connection strengths;

[0149] (1-2) Dynamically adjust sample weights based on their topological consistency scores;

[0150] The calculation formula of sample weight is as follows:

[0151] ,

[0152] in, For the The weight of the samples, is the sensitivity coefficient, Set to 0.3, For the The topological consistency score of the samples, is the total number of samples.

[0153] (2) Weight sharing:

[0154] A topology-driven parameter sharing strategy is used to share weights. For connection groups with similar contributions, similar contribution connections are forced to share parameters, which can reduce redundant parameters while maintaining the expressiveness of key connections.

[0155] (2-1) Contribution feature construction: All connections of the layer, extract their contribution scores , construct a one-dimensional feature vector , , Represents the neural network Layer neurons to the next layer The contribution score of each neuron connection, Represents the neural network The contribution score of all connections in the layer, the lower limit of the contribution score is , the upper limit of contribution score is ;

[0156] (2-2) Contribution grouping and clustering:

[0157] Use K-means algorithm to analyze one-dimensional feature vectors Perform contribution grouping and clustering, and then determine the objective function that minimizes the intra-cluster variance, expressed as ,in, represents the objective function that minimizes the intra-cluster variance, Indicates the shared clusters, Indicates the shared clusters, Indicates the preset number of clusters, Indicates the The centroid of each cluster;

[0158] (2-3) Weight sharing:

[0159] ,

[0160] in, express The sharing weight of the sharing group, Indicates that it belongs to connections between neurons that share a cluster;

[0161] The upper and lower limits of the contribution score are calculated as follows:

[0162] ,

[0163] ,

[0164] ,

[0165] ,

[0166] in, Indicates the The centroid of the cluster, represents the within-cluster standard deviation.

[0167] In a specific implementation, the neural network topology freezing is specifically as follows:

[0168] A progressive neural network topology freezing method is used. When the network topology change rate is lower than a preset threshold, the layer of the neural network is frozen. The network topology change rate is measured by the connection change rate or the weight change rate. Freezing the neural network topology can prevent over-adjustment.

[0169] The connection change rate is used to measure the rate of change of the network topology. The connection change rate is determined by measuring the increase or decrease in the number of connections between neurons in each layer of the neural network. For example, the network topology may change the number of connections due to dynamic adjustments. The change rate can calculate the change in the number of connections in each layer during the iteration process.

[0170] The rate of change of network topology is measured by the weight change rate. The weight change rate is determined by measuring the degree of change of the weight of the neural network over a period of time. If the connection between neurons is changed, the corresponding weight will also be updated. By calculating the update amplitude of the weight and summing them, the rate of change of the network structure is obtained.

[0171] In a specific embodiment, the training process of the recycled aggregate hydraulic concrete durability performance detection model is repeated until a preset stop iteration condition is met, thereby obtaining a trained recycled aggregate hydraulic concrete durability performance detection model;

[0172] The preset stopping condition for iteration is reaching the preset maximum number of iterations, which is set to 1000 times.

[0173] In a specific implementation, S3 is as follows:

[0174] The new recycled aggregate hydraulic concrete data is input into the trained recycled aggregate hydraulic concrete durability performance detection model. In the forward propagation stage, the input data passes through each layer of the neural network in turn, and after calculation in each layer, it passes through the output layer composed of the Softmax activation function, and finally outputs the probability value of each durability performance classification. The category with the highest probability value is the durability performance category of the recycled aggregate hydraulic concrete. The durability performance categories include high durability, medium durability and low durability.

[0175] The obtained durability performance classification results can be applied to actual engineering decisions. For example, staff can determine whether to use this type of recycled aggregate hydraulic concrete based on the durability performance classification results, or determine the performance of this type of recycled aggregate hydraulic concrete based on the durability performance classification and make adjustments, such as adjusting the mix ratio and curing conditions of the recycled aggregate hydraulic concrete to adjust the durability performance.

[0176] Example 2

[0177] like Figure 2 As shown in the figure, in order to verify that the dynamic feature fusion topology connection mechanism adopted by the present invention improves the classification performance compared with the fixed full connection structure, Figure 2 It can be seen that the method converges quickly in the early stage of training (<200 rounds), and the final accuracy is about 8% higher than that of the fixed connection. This proves that the dynamic topology mechanism adopted by the present invention can significantly improve the model's modeling ability for recycled aggregate data by enhancing the interaction of key features.

[0178] Example 3

[0179] like Figure 3 As shown, in order to verify that the local density perception initialization method adopted by the present invention is more stable than the conventional technology (Xavier initialization and random initialization), Figure 3 It can be seen that the accuracy of the method of the present invention reaches 82%, which is significantly higher than the existing conventional technology. This shows that the local density-aware initialization used in the present invention has a smaller fluctuation range, is better at capturing the local structure of the data, can avoid local optimality, and can improve the stability of the model.

[0180] Example 4

[0181] like Figure 4 As shown in the figure, the topology adjustment strategy used in the invention affects the computational efficiency. By comparing the computational efficiency of the staged topology reconstruction and the fixed structure, it is verified whether the dynamic topology reconstruction can reduce the computational cost while ensuring the accuracy. Figure 4 As can be seen, as the training progresses, the computational complexity decreases by 60% while the accuracy continues to increase, which shows that the staged reconstruction strategy achieves dual optimization of efficiency and accuracy by pruning redundant connections.

[0182] Example 5

[0183] like Figure 5 As shown in the figure, the adaptive topology mechanism used in the present invention and the conventional technology are based on the material mechanics theory model to achieve a three-dimensional surface diagram of the multi-scale fusion effect. The three-dimensional surface reconstruction is used to compare the accuracy differences of different methods in simulating the variation law of concrete compressive strength. Figure 5 It can be seen that conventional technologies have difficulty coordinating the nonlinear effects of multi-scale parameters such as aggregate particle size and curing age due to the fixed connection structure, resulting in excessive smoothness and local distortion of the prediction surface. However, the present invention dynamically integrates feature information of different dimensions to generate a strength prediction surface that is highly consistent with the material mechanics model, especially in the area of aggregate gradation mutation, and can still maintain the physical rationality of the prediction curve. The adaptive topology mechanism effectively solves the scale mismatch problem of multi-source data, enabling the model to accurately capture the complex changes in material properties.

[0184] In addition, as can be seen from Table 1, through the three evaluation indicators of RMSE (MPa), surface curvature similarity and extreme point error, it can be shown that the present invention has made significant improvements on the basis of conventional technology.

[0185] Table 1 Performance comparison between the present invention and conventional technology

[0186]

[0187] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for testing the durability of recycled aggregate hydraulic concrete based on a neural network, characterized in that: The following steps are involved: S1. Data collection: Collect recycled aggregate hydraulic concrete data and manually label the recycled aggregate hydraulic concrete based on the collected data; S2. Constructing a neural network based on adaptive topological connections as a durability testing model for recycled aggregate hydraulic concrete, inputting the collected data into the model for training, and obtaining a durability classification result for the recycled aggregate hydraulic concrete based on the model's analysis of the recycled aggregate hydraulic concrete data, until a preset stopping condition is met to complete the model training; The durability test model of recycled aggregate hydraulic concrete is trained. The training process is as follows: For a neural network, the connection strength of its topological structure is calculated, the neural network is initialized, and the connection of neurons in the neural network topological structure is adjusted; Adopting multi-task driven adaptive loss function to balance multi-objective optimization; In the back propagation process of the neural network, the path continuity constraint optimization method is adopted; A topology-robust sample weighting method is used to calculate sample weights, and then a topology-driven parameter sharing strategy is used to share the weights. A progressive neural network topology freezing method is used to freeze some layers of the neural network; S3. Input the new recycled aggregate hydraulic concrete data into the trained recycled aggregate hydraulic concrete durability performance detection model to obtain the durability performance classification result of the aggregate hydraulic concrete.

2. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 1, characterized in that: S1 is as follows: (1) The collected data of recycled aggregate hydraulic concrete include but are not limited to aggregate source, mixing ratio, compressive strength, water absorption, chloride ion permeability, number of freeze-thaw cycles and carbonation depth; The methods of collecting data include field collection, laboratory collection, and collection by sensors and detection systems; (2) Manually labeled categories include: high durability, medium durability, and low durability.

3. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 2, characterized in that: The connectivity enhancement, initialization and connection adjustment of the neural network are as follows: (1) A dynamic feature fusion topology connection mechanism is used to adjust the connection strength of the neural network topology structure. Specifically, the connection weights are dynamically adjusted by the mutual information entropy and gradient sensitivity between features to achieve adaptive feature interaction. (2) The neural network uses a weight initialization method based on local density perception, combining the feature space cluster center and local density distribution to generate initial weights; (3) A staged topology reconstruction strategy is adopted to dynamically add and delete neuron connections according to the feature contribution of recycled aggregate hydraulic concrete data to adjust the neuron connections of the neural network topology structure, as follows: (3-1) Calculate the characteristic contribution of recycled aggregate hydraulic concrete data; (3-2) Adjust the neural network topology. Each layer of the neural network is sorted according to the feature contribution, retaining the top M% of high-contribution features and removing the rest. For the retained features with high contribution, new connections are inserted in the areas where the feature contribution is lower than the threshold to obtain a dynamic retopology structure.

4. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 3 is characterized in that: The loss function is as follows: The recycled aggregate hydraulic concrete durability performance testing model uses a multi-task driven adaptive loss function. The adaptive loss function combines the durability classification task with the index regression task and dynamically adjusts the loss weight. This allows the recycled aggregate hydraulic concrete durability performance testing model to automatically focus on difficult-to-optimize tasks and balance multi-objective optimization tasks. Adaptive loss function The calculation formula is as follows: ， ， in, is the loss weight coefficient of the classification task, To calculate the classification loss using weighted cross entropy loss, For the The weight of the samples; is the category weight, For the The samples belong to The true labels of the categories, For the The samples belong to The predicted probability of each category, is the logarithmic function with base 10, is the total number of samples, is the total number of categories, is the number of regression tasks, For the The loss weight of the iteration, is the regression loss.

5. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 4, characterized in that: The path continuity constraint optimization used in the directional propagation process of the neural network is as follows: The topology change penalty term is used in the gradient calculation of the neural network back propagation. The calculation formula is as follows: ， in, is the gradient corrected by the topology change penalty term; is the penalty coefficient; is the indicator function, Indicates when the connection strength Below the preset threshold Take 1 when it is, otherwise take 0; is a symbolic function; For the neural network Layer neurons to the next layer The weights of the neuron connections.

6. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 5, characterized in that: The calculation of sample weights and weight sharing are as follows: (1) Calculate sample weight: (1-1) The matrix, expected value and variance of the connection strength of the sample, using the Frobenius norm, calculate the topological consistency score of the sample during the forward propagation of the neural network; (1-2) Dynamically adjust sample weights based on their topological consistency scores; (2) Weight sharing: A topology-driven parameter sharing strategy is used to share weights. For connection groups with similar contributions, connections with similar contributions are forced to share parameters. (2-1) Contribution feature construction: All connections of the layer, extract their contribution scores , construct a one-dimensional feature vector , , Represents the neural network Layer neurons to the next layer The contribution score of each neuron connection, Represents the neural network The contribution score of all connections in the layer, the lower limit of the contribution score is , the upper limit of contribution score is ; (2-2) Contribution grouping and clustering: Use K-means algorithm to analyze one-dimensional feature vectors Perform contribution grouping and clustering, and then determine the objective function that minimizes the intra-cluster variance, expressed as ,in, represents the objective function that minimizes the intra-cluster variance, Indicates the shared clusters, Indicates the shared clusters, Indicates the preset number of clusters, Indicates the The centroid of each cluster; (2-3) Weight sharing: ， in, express The sharing weights of the sharing groups, Indicates that it belongs to connections between neurons that share a cluster.

7. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 6, characterized in that: The neural network topology freezing is as follows: A progressive neural network topology freezing method is adopted. When the network topology structure change rate is lower than a preset threshold, the layer of the neural network is frozen. The network topology structure change rate is measured by the connection change rate or the weight change rate.

8. The method for detecting durability of recycled aggregate hydraulic concrete based on a neural network according to claim 7, characterized in that: The training process of the recycled aggregate hydraulic concrete durability performance detection model is repeated until the preset stopping iteration condition is met, and a trained recycled aggregate hydraulic concrete durability performance detection model is obtained.

9. The method for detecting durability of recycled aggregate hydraulic concrete based on neural network according to claim 8, characterized in that: S3 is as follows: The new recycled aggregate hydraulic concrete data is input into the trained recycled aggregate hydraulic concrete durability performance detection model. In the forward propagation stage, the input data passes through each layer of the neural network in turn, and after calculation in each layer, it passes through the output layer composed of the Softmax activation function, and finally outputs the probability value of each durability performance classification. The category with the highest probability value is the durability performance category of the recycled aggregate hydraulic concrete. The durability performance categories include high durability, medium durability and low durability.

Images