Freeze-thaw concrete strength evaluation method and device based on interpretable neural network

Abstract

The application discloses a kind of based on explainable neural network's frozen and thawed concrete strength evaluation method and device, obtain frozen and thawed concrete component test parameter and evaluation parameter;Freeze-thawed concrete sample data set is divided into training set and test set;Establish explainable neural network machine learning model, the input of model is the dosage of various raw material components in each group of freeze-thawed concrete sample, and the output is frozen and thawed concrete strength;Real-time feature data is brought into the explainable neural network machine learning model established, and the strength of freeze-thawed concrete is predicted.The application replaces the traditional test method to predict the strength of freeze-thawed concrete, saves a lot of time and cost;Compared with traditional machine learning algorithm, the prediction model established by the generalized additive explainable neural network machine learning algorithm used in the application provides accurate and explainable prediction results.

Description

Technical Field

[0001] This invention relates to the field of freeze-thaw concrete strength evaluation and testing technology, specifically a method and apparatus for evaluating freeze-thaw concrete strength based on interpretable neural networks. Background Technology

[0002] Concrete strength has always been one of the important indicators for evaluating concrete durability. Concrete components in cold regions often develop cracks due to the destructive effects of load stress, which leads to a decrease in durability or even failure. This affects the comfort and safety of buildings to a certain extent. Therefore, obtaining freeze-thaw concrete strength data quickly, reasonably and effectively is of great guiding significance for judging structural safety, thereby maximizing the service life of concrete.

[0003] Typically, the raw materials for concrete production include cement, blast furnace slag, fly ash, water, air-entraining agents, superplasticizers, fine aggregates, and coarse aggregates. Currently, most traditional direct determinations of concrete strength require significant time and labor costs, and related empirical models also suffer from insufficient accuracy. Therefore, a method for rapidly predicting the strength of freeze-thawed concrete based on concrete raw material composition information and curing conditions is of great significance for optimizing the original concrete mix proportions.

[0004] Currently, in the prediction and analysis of the performance of freeze-thaw concrete, traditional machine learning models lack interpretability. The models trained by the algorithms are regarded as black boxes. As the complexity of machine learning models increases, people often find it difficult to understand and accept their prediction results. Therefore, machine learning models based on generalized additive interpretable neural networks can achieve better results in prediction problems. Summary of the Invention

[0005] To overcome the problems of long test cycles, complex experiments, and lack of interpretability of ordinary machine learning models in traditional concrete durability testing, this invention provides a method and apparatus for evaluating the strength of freeze-thaw concrete based on interpretable neural networks.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network, comprising the following steps:

[0007] (1) Obtain test data of freeze-thawed concrete components, including concrete characteristic data and compressive strength evaluation data;

[0008] (2) Based on step (1), construct the dataset and divide the dataset into training set and test set after random arrangement;

[0009] (3) Establish an interpretable neural network machine learning model and train the model based on the training set in step (2). The input of the model is the feature data of each concrete sample, namely the amount of raw materials and the curing conditions, and the output is the evaluation data of the concrete, namely the compressive strength. The interpretable neural network machine learning model is an interpretable neural network xNN structure, which includes a fully connected multilayer perceptron decomposed into a projection layer, followed by multiple subnetworks, where each subnetwork represents a nonlinear shape function. xNN uses neural network parameterization for both main effects and pairwise interactions. Each main effect or pairwise interaction is simulated by a fully connected subnetwork consisting of one or two input nodes. These subnetworks are added together to form the final output.

[0010] The training process of the interpretable neural network machine learning model is a three-stage adaptive training: First, the subnetworks of the main effects are trained and pruned for fitting; second, important pairwise interactions are selected, fitted and pruned, and the residuals are fitted under genetic constraints; finally, all important main effects and pairwise interactions are collectively fine-tuned.

[0011] The interpretable neural network machine learning model is described as follows:

[0012]

[0013] Where E(y|x) represents the expected value of the compressive strength y given feature x, S1 and S2 represent the sets of main effects and pairwise interactions, respectively, g is a function of the model, and h j and f jk It is the shape function corresponding to the main effect and the pairwise interaction, μ is the intercept, j represents the j-th main effect, and jk represents the pairwise interaction; x j and x k These are the feature data of the j-th and k-th concrete samples, respectively, and are used as the input to the model.

[0014] Each main effect and pairwise interaction is assumed to have a zero mean, i.e.

[0015]

[0016] Where F(x) j ) and F(x j x k ) represents the cumulative distribution function of the corresponding j-th main effect and jk-th pairwise interaction;

[0017] (4) Use the test set in step (2) to test the prediction performance of the trained model. If the prediction accuracy does not meet the requirements, adjust the parameters of the interpretable neural network machine learning model until the required prediction accuracy is achieved.

[0018] (5) Input the real-time collected feature data of freeze-thaw concrete into an interpretable neural network machine learning model that meets the requirements to predict the compressive strength of freeze-thaw concrete.

[0019] Furthermore, the training model is constrained by sparsity, heritability, and marginal clarity.

[0020] Furthermore, model training will stop when the maximum training duration is reached or when the validation performance has not improved within a certain duration. Each main effect or pairwise interaction is normalized to have a zero mean, the bias nodes of the output layer represent the overall mean, and the subnetworks are pruned according to sparsity constraints.

[0021] Furthermore, model training follows marginal clear regularization, in which all main effects and pairwise interactions are renormalized during the fine-tuning phase.

[0022] Furthermore, the model possesses inherent interpretability, examining the contribution of each individual variable to the overall prediction. The importance ratio (IR) of each main effect is quantitatively measured in the following ways:

[0023] IR(j)=D(h j ) / T

[0024] IR(j,k)=D(f jk ) / T

[0025] Among them, D(h) j ) and D(f jk ) represents the variance of the shape function corresponding to the main effect and the pairwise interaction, and T represents D(h) j ) and D(f jk The sum of ), where j represents the j-th main effect, and jk represents the pairwise interaction.

[0026] Furthermore, the evaluation method for the established interpretable neural network machine learning model is to calculate the root mean square error and correlation coefficient between the target value on the test set and the predicted target value:

[0027]

[0028]

[0029] Where, N i and P i These are the i-th test target value and the i-th predicted target value in the test set, respectively; N is the average test target value; and n is the number of samples.

[0030] Secondly, the present invention also provides a freeze-thaw concrete strength evaluation device based on an interpretable neural network, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the freeze-thaw concrete strength evaluation method based on an interpretable neural network.

[0031] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the aforementioned method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network.

[0032] The beneficial effects of the present invention are as follows: The present invention provides a method and apparatus for evaluating the strength of freeze-thaw concrete based on interpretable neural networks, which replaces the traditional test method for predicting the strength of freeze-thaw concrete, saving a lot of time and cost; compared with traditional machine learning algorithms, the prediction model established by the generalized additive interpretable neural network machine learning algorithm used in the present invention provides accurate and interpretable prediction results. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This invention provides a flowchart illustrating a method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network.

[0035] Figure 2 This is a schematic diagram illustrating the changes in model parameters during the machine learning training phase in Embodiment 1 of the present invention;

[0036] Figure 3 This is a global interpretation of the prediction results in Embodiment 1 of the present invention;

[0037] Figure 4 This is a partial interpretation of the prediction results in Embodiment 1 of the present invention.

[0038] Figure 5 This is a structural diagram of a freeze-thaw concrete strength evaluation device based on an interpretable neural network according to the present invention. Detailed Implementation

[0039] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0040] This invention provides a method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network, such as... Figure 1 The diagram shown is a flowchart of the freeze-thaw concrete strength evaluation method based on an interpretable neural network, which specifically includes:

[0041] S1: Obtain test data for freeze-thawed concrete components, including characteristic data and strength evaluation data of freeze-thawed concrete such as cement, blast furnace slag, fly ash, water, air-entraining agent, superplasticizer, fine aggregate, coarse aggregate, curing period and number of freeze-thaw cycles;

[0042] S2: Construct the dataset by randomly dividing it into a training set and a test set. When splitting all experimental data into training and test sets, the test set accounts for 20% of the total number of samples. In this first embodiment, the test set accounts for 20% of the total number of samples.

[0043] S3: Establish an interpretable neural network machine learning model. The input of the model is the feature data of each group of freeze-thaw concrete samples, namely the amount of raw materials and the test conditions. The output is the evaluation data of freeze-thaw concrete.

[0044] S4: Train an interpretable neural network machine learning model and store the trained interpretable neural network machine learning model on the device;

[0045] S5: Use a test set to test the prediction performance of the trained model. If the prediction accuracy does not meet the requirements, adjust the parameters of the interpretable neural network machine learning model until a suitable prediction accuracy is achieved.

[0046] S6: Input real-time feature data, namely the amount of various raw materials used in freeze-thaw concrete and curing conditions, into the established interpretable neural network machine learning model to predict the strength of freeze-thaw concrete.

[0047] Specifically, in step S3, when building an interpretable neural network machine learning model, the machine learning framework used is TensorFlow.

[0048] The established interpretable neural network machine learning model is constrained by sparsity, heritability, and marginal clarity.

[0049] The interpretable neural network machine learning model trained in step S4 includes:

[0050] S41: A three-stage adaptive training algorithm. First, the subnetworks with main effects are trained and pruned for fitting. Second, important pairwise interactions are selected, fitted, and pruned, and the residuals are fitted under genetic constraints. Finally, all important main effects and pairwise interactions are collectively fine-tuned.

[0051] S42: Adaptive training algorithms include:

[0052] The interpretable neural network machine learning model is described as follows:

[0053]

[0054] Where E(y|x) represents the expected value of the compressive strength y given feature x, S1 and S2 represent the sets of main effects and pairwise interactions, respectively, μ is the intercept, g is a function of the model, and h j and d jk It is the shape function corresponding to the main effect and the pairwise interaction, where j represents the j-th main effect and jk represents the pairwise interaction; x j and x k These are the feature data of the j-th and k-th concrete samples, respectively, and are used as the input to the model.

[0055] Each main effect and pairwise interaction is assumed to have a zero mean, i.e.,

[0056]

[0057] Where, F(x) j ) and F(x j x k ) represents the cumulative distribution function of the corresponding j-th main effect and jk-th pairwise interaction.

[0058] S43: A novel interpretable neural network (xNN) architecture, comprising:

[0059] A fully connected multilayer perceptron is decomposed into a projection layer, followed by multiple subnetworks, each representing a nonlinear shape function. xNN employs neural network parameterization for both main effects and pairwise interactions.

[0060] The maximum number of pairwise interactions is set to K=20. Each subnetwork is configured with 5 ReLU hidden layers, each with 40 nodes. The weights of the subnetworks are initialized using a Gaussian orthogonal initializer. The initial learning rate of the Adam optimizer is set to 0.0001. The number of training epochs for the three training phases are set to 5000, 5000, and 500, respectively. 20% of the validation set is used for early stopping, with an early stopping threshold of 50 epochs. The tolerance threshold η is set to 1% of the minimum validation loss. The marginal clarity regularization strength can be selected empirically from 0.0001 to 1.

[0061] Each main effect or pairwise interaction is simulated using a fully connected subnetwork consisting of one or two input nodes. These subnetworks are then combined to form the final output.

[0062] S44: Model training will stop when the maximum training duration is reached or when validation performance has not improved within a certain duration. Each main effect or pairwise interaction is normalized to have a zero mean, so that the bias nodes of the output layer represent the overall mean. Trivial subnetworks are pruned according to sparsity constraints.

[0063] Model training follows marginal clear regularization, in which all main effects and pairwise interactions are renormalized during the fine-tuning phase.

[0064] The model is inherently interpretable. Examining the contribution of each individual variable to the overall prediction, the importance ratio (IR) of each main effect can be quantitatively measured in the following ways:

[0065] IR(j)=D(h j ) / T

[0066] IR(j,k)=D(f jk ) / T

[0067] Among them, D(h) j ) and D(f jk Let T represent the variance of the shape functions corresponding to the main effects and paired interactions, and let T denote the variance of D(n). j ) and D(f jk The sum of ), where j represents the j-th main effect, and jk represents the pairwise interaction.

[0068] In step S5, the evaluation method for verifying the established interpretable neural network machine learning model is to calculate the root mean square error and correlation coefficient between the test target value and the predicted target value on the test set.

[0069]

[0070]

[0071] Where Ti and Pi are the i-th test target value and the i-th predicted target value of the test set, respectively; N is the average test target value; and n is the number of samples.

[0072] The freeze-thaw concrete data collected in Embodiment 1 of this invention were derived from relevant articles on freeze-thaw concrete component tests both domestically and internationally. After training the prediction model, the RMSE between the predicted and actual values ​​of the sample data in the test set was 0.045, and the correlation coefficient R0 was [missing value]. 2 It is 0.952.

[0073] This invention provides a schematic diagram of model parameter changes during the machine learning training phase, see [link / reference]. Figure 2This invention provides the changes in training loss and validation loss in the three stages of the aforementioned model training as the training duration increases. As can be seen from the figure, with the addition of main effects and pairwise interactions in the model, the loss decreases accordingly and the prediction accuracy improves.

[0074] This invention provides a global interpretation of machine learning prediction results; see [link to relevant documentation]. Figure 3 For main effects, a line graph is used to represent the input-output relationship of each variable. For pairwise interactions, a two-dimensional graph is used to visualize the joint effect of the two base variables. The percentage of each feature variable represents its importance ratio. In the one-dimensional graph, the vertical axis represents the contribution of the feature variable to the prediction, the horizontal axis represents the input value of the feature variable, and the bar chart shows the distribution of the feature variable. The trend of the line graph can be used to judge the impact of the main effect on the prediction result. In the two-dimensional graph, the horizontal axis represents the input value of the first feature in the pairwise interaction, and the vertical axis represents the input value of the second feature in the pairwise interaction. The depth of the graph can be used to judge the impact of the pairwise interaction on the prediction result. As shown in the figure, water content is negatively correlated with concrete strength, while cement content is positively correlated with concrete strength.

[0075] This invention provides a partial interpretation of machine learning prediction results; see [link to relevant documentation]. Figure 4 This invention provides the decision-making process of a machine learning model for predicting the outcome of a single sample. The horizontal axis represents main effects or pairwise interactions, where the intercept represents the average of all sample predictions, and the vertical axis represents the contribution of each main effect or pairwise interaction to the prediction of that sample. The data in the sample is the result after regularization; see [link to relevant documentation]. Figure 4 The actual strength of this sample is 0.5052, and the predicted strength is 0.5139. In this sample, water has a positive effect on the prediction result, while cement has a negative effect on the prediction result.

[0076] Compared with traditional machine learning algorithms, the prediction model established by the generalized additive interpretable neural network machine learning algorithm used in this invention provides accurate and interpretable prediction results.

[0077] Corresponding to the aforementioned embodiment of a method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network, the present invention also provides an embodiment of a device for evaluating the strength of freeze-thaw concrete based on an interpretable neural network.

[0078] See Figure 5 The present invention provides a freeze-thaw concrete strength evaluation device based on an interpretable neural network, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a freeze-thaw concrete strength evaluation method based on an interpretable neural network as described in the above embodiment.

[0079] The embodiment of the freeze-thaw concrete strength evaluation device based on an interpretable neural network provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, including the freeze-thaw concrete strength evaluation device based on an interpretable neural network provided by this invention. (Except for...) Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0080] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0081] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device 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 the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0082] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a freeze-thaw concrete strength evaluation method based on an interpretable neural network as described in the above embodiments.

[0083] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0084] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for evaluating the strength of freeze-thaw concrete based on interpretable neural networks, characterized in that, Includes the following steps: (1) Obtain test data of freeze-thawed concrete components, including concrete characteristic data and compressive strength evaluation data; (2) Based on step (1), construct the dataset and divide the dataset into training set and test set after random arrangement; (3) Establish an interpretable neural network machine learning model and train the model based on the training set in step (2). The input of the model is the feature data of each concrete sample, namely the amount of raw materials and the curing conditions, and the output is the evaluation data of the concrete, namely the compressive strength. The interpretable neural network machine learning model is an interpretable neural network xNN structure, which includes a fully connected multilayer perceptron decomposed into a projection layer, followed by multiple subnetworks, where each subnetwork represents a nonlinear shape function. xNN uses neural network parameterization for both main effects and pairwise interactions. Each main effect or pairwise interaction is simulated by a fully connected subnetwork consisting of one or two input nodes. These subnetworks are added together to form the final output. The training process of the interpretable neural network machine learning model is a three-stage adaptive training: First, the subnetworks of the main effects are trained and pruned for fitting; second, important pairwise interactions are selected, fitted and pruned, and the residuals are fitted under genetic constraints; finally, all important main effects and pairwise interactions are collectively fine-tuned. The interpretable neural network machine learning model is described as follows: Where E(y|x) represents the expected compressive strength y given feature x, S1 and S2 represent the sets of main effects and pairwise interactions, respectively, g is a function of the model, and h j and f jk It is the shape function corresponding to the main effect and the pairwise interaction, μ is the intercept, j represents the j-th main effect, and jk represents the pairwise interaction; x j and x k These are the feature data of the j-th and k-th concrete samples, respectively, and are used as the input to the model. Each main effect and pairwise interaction is assumed to have a zero mean, i.e. Where F(x) j ) and F(x j x k ) represents the cumulative distribution function of the corresponding j-th main effect and jk-th pairwise interaction; (4) Use the test set in step (2) to test the prediction performance of the trained model. If the prediction accuracy does not meet the requirements, adjust the parameters of the interpretable neural network machine learning model until the required prediction accuracy is achieved. (5) Input the real-time collected feature data of freeze-thaw concrete into an interpretable neural network machine learning model that meets the requirements to predict the compressive strength of freeze-thaw concrete.

2. The method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network according to claim 1, characterized in that: The training model is constrained by sparsity, heritability, and marginal clarity.

3. The method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network according to claim 2, characterized in that: Model training will stop when the maximum training duration is reached or when the validation performance does not improve within a certain duration. Each main effect or pairwise interaction is normalized to have a zero mean. The bias nodes of the output layer represent the overall mean. Subnetworks are pruned according to sparsity constraints.

4. The method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network according to claim 1, characterized in that: Model training follows marginal clear regularization, in which all main effects and pairwise interactions are renormalized during the fine-tuning phase.

5. The method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network according to claim 1, characterized in that: The model is inherently interpretable. The contribution of each individual variable to the overall prediction is examined, and the importance ratio (IR) of each main effect is quantitatively measured as follows: IR(j)=D(h j ) / T IR(j,k)=D(f jk ) / T Among them, D(h) j ) and D(f jk ) represents the variance of the shape function corresponding to the main effect and the pairwise interaction, and T represents D(h) j ) and D(f jk The sum of ), where j represents the j-th main effect, and jk represents the pairwise interaction.

6. The method for evaluating the strength of freeze-thaw concrete based on an interpretable neural network according to claim 1, characterized in that: The evaluation method for the established interpretable neural network machine learning model is to calculate the root mean square error and correlation coefficient between the target value on the test set and the predicted target value. Where, N i and P i These are the i-th test target value and the i-th predicted target value in the test set, respectively; N is the average test target value; and n is the number of samples.

7. A device for evaluating the strength of freeze-thaw concrete based on an interpretable neural network, characterized in that: It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the freeze-thaw concrete strength evaluation method based on an interpretable neural network as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the freeze-thaw concrete strength evaluation method based on an interpretable neural network as described in any one of claims 1 to 6.

Images