Damage quantitative index prediction method in concrete freeze thawing-sulfate erosion process

Abstract

The invention belongs to the technical field of concrete, and provides a concrete freeze thawing-sulfate erosion process damage quantitative index prediction method in order to solve the problem that an existing method lacks comprehensive, dynamic and deep evaluation means, and the method comprises the steps that damage quantitative indexes are collected; constructing a feature data training set, a feature data test set, a label data training set and a label data test set of the damage quantitative indexes; inputting the feature data training set of the damage quantitative index into a CNN-LSTM-Attention model to obtain a predicted value of the damage quantitative index, and updating the bias and the weight of an output layer of the CNN-LSTM-Attention model by adopting an Adam optimization algorithm; and inputting the feature data test set of the damage quantitative index into the optimized CNN-LSTM-Attention model, and outputting a predicted value of the feature data test set. According to the method, rapid and accurate prediction is realized.

Description

Technical Field

[0001] The invention belongs to the field of research on the durability of concrete lining structures, and specifically relates to a method for predicting quantitative damage indicators in a concrete freeze-thaw-sulfate erosion process. Background Art

[0002] Saline soil is widely distributed in my country, and the western region is the main concentrated distribution area of ​​saline soil, which is rich in sulfate. The environment in this region is also characterized by low precipitation and high evaporation. Concrete structures are easily affected by dry-wet cycles, resulting in serious deterioration of mechanical properties. In addition, my country's cold regions cover a wide area, and seasonal cold regions account for a high proportion. Due to special climatic conditions, the climate in cold regions is extremely cold, the seasonal temperature difference is large, and concrete structures are susceptible to freeze-thaw damage. In the sulfate-saline soil environment in the cold western regions, concrete structures will not only be affected by freeze-thaw, but also by sulfate erosion, and the performance of concrete will deteriorate more rapidly. Therefore, the quantitative prediction of concrete damage has become an important and complex issue.

[0003] At present, some research methods on the durability of concrete materials mainly focus on corrosion tests and monitoring of some indicators, lacking comprehensive, dynamic and in-depth evaluation methods. In addition, the current quantitative prediction methods for concrete damage are mostly based on empirical formulas or mathematical models, which take into account many factors and require many parameters, which is not conducive to the rapid and accurate prediction of damage. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention provides a method for predicting quantitative damage indicators of concrete in the freeze-thaw-sulfate erosion process, which aims to comprehensively and accurately evaluate the performance changes of concrete in actual engineering by combining a convolutional neural network with an attention mechanism and a long short-term memory network (LSTM) through freeze-thaw sulfate dry-wet cycle test and monitoring multiple damage quantification indicators (mass loss rate, relative compressive strength, relative dynamic elastic modulus).

[0005] A method for predicting quantitative indicators of concrete damage during freeze-thaw-sulfate attack, comprising: S1: During the freeze-thaw-sulfate wet-dry cycle test of concrete specimens, the quantitative damage index of concrete specimens was collected during the test; Among them, the damage quantification indexes include mass loss rate, relative compressive strength and relative dynamic elastic modulus; S2: preprocess the damage quantification index in S1 to construct a feature data training set and a feature data test set of the damage quantification index and a label data training set and a label data test set of the damage quantification index; S3: Input the characteristic data training set of the damage quantification index of S2 into the CNN-LSTM-Attention model to obtain the predicted value based on the characteristic data training set of the damage quantification index; Adopt the Adam optimization algorithm, combine the predicted value and the label data training set of the damage quantification index to update the bias and weight of the output layer of the CNN-LSTM-Attention model, and optimize the CNN-LSTM-Attention model; Among them, the first layer of the CNN-LSTM-Attention model is the convolutional neural network layer, the second layer is the long short-term memory neural network layer, the third layer is the attention layer, and the fourth layer is the output layer; S4: The feature data test set of the damage quantification index is input into the optimized CNN-LSTM-Attention model. The output layer of the CNN-LSTM-Attention model outputs the predicted value based on the feature data test set of the damage quantification index, and the prediction result of the damage quantification index of the concrete freeze-thaw-sulfate attack process is obtained. The labeled data test set of the damage quantification index is used to verify the convergence of the CNN-LSTM-Attention model through simulation experiments.

[0006] Preferably, the process of data preprocessing of the S2 damage quantification index includes: S2.1: Calculate the mean and standard deviation of the damage quantification index, remove the outliers in the damage quantification index according to the calculated mean and standard deviation, and obtain the damage quantification index with missing data points; S2.2: fill in the missing data points of the damage quantification index in S2.1 by using the interpolation method to obtain the filled damage quantification index, and expand the filled damage quantification index to obtain the expanded damage quantification index; S2.3: normalize the damage quantification index expanded in S2.2 to obtain a normalized damage quantification index set; S2.4: Divide the damage quantification indicator set into a feature data training set and a feature data test set of the damage quantification indicator.

[0007] Preferably, S2.3 includes: S2.3.1: Normalize the mass loss rate in the damage quantification index expanded in S2.2 to obtain the normalized mass loss rate, which is calculated as follows: ; in, is the mass loss rate after expansion, is the minimum value of the mass loss rate after expansion, is the maximum value of the mass loss rate after expansion, It isj The normalized mass loss rate; Then the set of normalized mass loss rates is obtained as : ; S2.3.2: The relative compressive strength in the expanded damage quantification index of S2.2 is normalized to obtain the normalized relative compressive strength, which is calculated as follows: ; in, is the relative compressive strength after expansion, is the minimum value of the relative compressive strength after expansion, is the maximum value of the relative compressive strength after expansion, It is j The normalized relative compressive strength; Then the normalized relative compressive strength set is obtained : ; S2.3.3: Normalize the relative dynamic elastic modulus in the expanded damage quantification index of S2.2 to obtain the normalized relative dynamic elastic modulus, which is calculated as follows: ; in, is the relative dynamic elastic modulus after expansion, is the minimum value of the relative dynamic elastic modulus after expansion, is the maximum value of the relative dynamic elastic modulus after expansion, It is j The normalized relative dynamic elastic modulus; Then the normalized relative dynamic elastic modulus set is obtained : .

[0008] Preferably, S3 includes: S3.1: Input the feature data training set of the damage quantification index into the convolutional neural network layer of the CNN-LSTM-Attention model, perform convolution operation, obtain the convolved value, and then pass ReLU The activation function performs nonlinear transformation on the convolutional values ​​to obtain nonlinear transformed values, and performs pooling operation on the nonlinear transformed values ​​to obtain the output result of the convolutional neural network layer; S3.2: Input the output result of the convolutional neural network layer of S3.1 into the long short-term memory neural network layer to obtain the output result of the long short-term memory neural network layer; wherein the long short-term memory neural network layer includes an input gate, a forget gate, an output gate and a cell state; S3.3: Input the output result of the long short-term memory neural network layer in S3.2 into the attention layer to obtain the output result of the attention layer; S3.4: The output result of the attention layer in S3.3 is input into the output layer, and the output layer outputs the predicted value based on the training set of the damage quantification indicator feature data.

[0009] Preferably, the specific calculation process of the pooling operation in S3.1 is: ; ; ; in, It is s The value after the mass loss rate pooling operation, It is s The value after the relative compressive strength pooling operation, It is s The value after the relative dynamic elastic modulus pooling operation; It is i The value of the mass loss rate after nonlinear transformation, It is i The value after nonlinear transformation of relative compressive strength, It is i The value of the relative dynamic elastic modulus after nonlinear transformation, It is i+ 1 value of mass loss rate after nonlinear conversion, It is i +1 value after nonlinear conversion of relative compressive strength, It is i +1 value after nonlinear conversion of relative dynamic elastic modulus, The function represents the maximum value.

[0010] Preferably, in S3.2, the output of the long short-term memory neural network layer is: ; ; ; in, is the hyperbolic tangent function, for t The quality loss rate of the output of the short-term memory neural network layer at the time step; for t The relative compressive strength of the output of the short-term memory neural network layer of the time step; for t The relative dynamic elastic modulus of the output of the short-term memory neural network layer at time step, yes t The quality loss rate after the output gate in the time step short-term memory neural network layer, yes t The time step relative compressive strength output gated value, yes t The time step relative dynamic elastic modulus outputs the gated value.

[0011] Preferably, S3.3 includes: S3.3.1: Use a weight matrix The results of the LSTM neural network layer output are weighted to obtain the attention score of the damage quantification index: ; ; ; in, is the hyperbolic tangent function, for The attention score of the time step quality loss rate, for The relative strength of attention of the time step, for The attention score of the time step relative to the dynamic elastic modulus; S3.3.2: Normalize the attention score of the damage quantification index in S3.3.1 through the Softmax function to obtain the attention weight of the normalized damage quantification index: ; ; ; in, for t The attention weight of the time step quality loss rate, for t The attention weights of the time steps relative to the compressive strength, for t The attention weight of the time step relative to the dynamic elastic modulus, For e The exponential function with base , is the total time step, is the current time step; S3.3.3: Based on the attention weights in S3.3.2, the damage quantification indicators output by the long short-term memory neural network layer are weighted and summed to obtain the output result of the attention layer: ; ; ; in, for t The quality loss rate of the attention layer output at the time step, for t The relative compressive strength of the attention layer output at each time step, for t The relative dynamic elastic modulus of the attention layer output at time step; is the total time step, is the current time step.

[0012] Preferably, the output result of the attention layer in S3.3 is input into the output layer, and the output layer outputs the predicted value of the feature data training set based on the damage quantification index: ; ; ; in, is the bias of the output layer, is the weight of the output layer, is the predicted value of mass loss rate, is the predicted value of relative compressive strength, Predicted value of relative dynamic elastic modulus.

[0013] Preferably, the Adam optimization algorithm is used to update the bias and weight of the output layer of the CNN-LSTM-Attention model, specifically including: (1): Determine the first-order moment estimate of the current traversal number and the second-order moment estimate of the current traversal number: ; ; in, is the gradient of the current traversal times, is the first-order moment estimate of the current traversal number, The first moment estimate of the number of previous passes, is the second-order moment estimate of the current traversal number, is the exponential decay rate of the momentum term, is the exponential decay rate of the variance term; (2): Based on (1), the deviation is corrected to obtain the first-order moment estimate and the second-order moment estimate after eliminating the deviation of the current traversal number: ; ; in, To eliminate the first-order moment estimate after the deviation of the current traversal number, To eliminate the bias of the current traversal times, the second-order moment estimate is obtained; is the exponential decay rate of the momentum term x Second power, is the exponential decay rate of the variance term x Power; (3): Based on (2), update the weight of the output layer of the CNN-LSTM-Attention model and bias , and then optimize the CNN-LSTM-Attention model: ; ; in, It is x+ The weight of the output layer in one pass, It is The neural network bias of the output layer of the traversal, It is x The neural network bias of the output layer of the traversal, is the learning rate, is a constant; (4): Calculate the loss function of the CNN-LSTM-Attention model MSE , update the weight by the number of traversals W and bias b , find the loss function that minimizes: ; ; ; in, MSE 1 is the loss function of the CNN-LSTM-Attention model based on the quality loss rate, MSE 2 is the loss function of the CNN-LSTM-Attention model based on relative compressive strength. MSE 3 is the loss function of the CNN-LSTM-Attention model based on relative dynamic elastic modulus, is the labeled data training set of mass loss rate, is a label data training set with relative compressive strength, It is the labeled data training set of relative dynamic elastic modulus. The labeled data training set of mass loss rate, the labeled data training set of relative compressive strength and the labeled data training set of relative dynamic elastic modulus all belong to the labeled data training set of damage quantification indicators.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention provides a method for predicting quantitative damage indicators in the freeze-thaw-sulfate erosion process of concrete. By detecting the dynamic concrete mass loss rate, relative compressive strength, and relative dynamic elastic modulus values, the method can be used to predict the quantitative damage indicators of the next freeze-thaw-sulfate erosion process without the need to perform a large number of experiments to obtain the corresponding parameters. The present invention does not fit the mass loss rate, relative compressive strength, and relative dynamic elastic modulus values ​​or use mathematical assumptions in the prediction process of concrete damage quantitative indicators, and has high reliability. The present invention can perform quantitative damage prediction on various types of concrete service environments, including the freeze-thaw sulfate dry-wet cycle environment in this case, as well as the corrosive environment caused by climate change, and has the characteristics of a wide range of applications. Therefore, the method of the present invention comprehensively solves a series of problems of current prediction technologies, such as narrow scope of application, complex calculations, and high prediction costs.

[0015] 2. The damage quantitative index prediction method based on freeze-thaw-sulfate erosion process of concrete lining structure provided in the present invention integrates CNN, Attention and LSTM, and the established prediction model combines the characteristics of different neural network algorithms. It avoids the problem that the existing methods are based on empirical formulas or mathematical models, which consider many factors and require many parameters, which is not conducive to the rapid and accurate prediction of damage, and provides a quick and effective method for the prediction of concrete damage quantitative indexes. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] 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.

[0017] In the attached picture: Attached Figure 1 : Flow chart of the method of the present invention.

[0018] Attached Figure 2 : Flow chart of freeze-thaw-sulfate wet-dry cycle test system.

[0019] Attached Figure 3 : Prediction of mass loss rate of concrete under freeze-thaw sulfate dry-wet cycles based on three neural network models: (a) LSTM, (b) CNN-LSTM, (c) CNN-LSTM-Attention.

[0020] Attached Figure 4 : Prediction of relative compressive strength of concrete under freeze-thaw sulfate wet-dry cycles based on three neural network models: (a) LSTM, (b) CNN-LSTM, (c) CNN-LSTM-Attention.

[0021] Attached Figure 5 : Prediction diagram of relative dynamic elastic modulus of concrete under freeze-thaw sulfate dry-wet cycles based on three neural network models: (a) LSTM, (b) CNN-LSTM, (c) CNN-LSTM-Attention. DETAILED DESCRIPTION

[0022] The following is combined with Figure 1 -Attached Figure 5 The preferred embodiments of the present invention are described. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0023] Example: like Figure 1 , a method for predicting quantitative indicators of concrete damage during freeze-thaw-sulfate attack, comprising the following steps: S1: Determine the temperature acceleration coefficient, the corrosion solution concentration acceleration coefficient and the freeze-thaw degradation proportional coefficient; determine the sulfate dry-wet cycle days according to the temperature acceleration coefficient and the corrosion solution concentration acceleration coefficient, and determine the freeze-thaw cycle number according to the freeze-thaw degradation proportional coefficient; based on the sulfate dry-wet cycle days and the freeze-thaw cycle number, complete the freeze-thaw-sulfate dry-wet cycle alternation test of concrete specimens, and collect the quantitative damage indicators of concrete specimens during the test: mass loss rate, relative compressive strength and relative dynamic elastic modulus.

[0024] S1.1: Determine the temperature acceleration factor and the corrosion solution concentration acceleration factor, and introduce the freeze-thaw degradation proportional coefficient.

[0025] S1.1.1: The average temperature of the concrete specimens during the sulfate drying and drying cycle test is given by: ; in: T 均 is the average test temperature, T 烘箱 is the oven temperature, t 1 is the duration of oven drying, T 1 is the average annual temperature of the region in Celsius, t 2 is the sulfate soaking time, t 3 is the air drying time; The expression of temperature acceleration coefficient is: ; Where: K a is the temperature acceleration coefficient; E / R is the activation energy, taken as 14242; T 1 is the average annual temperature of the region in Celsius.

[0026] S1.1.2: The acceleration factor for the concentration of the etching solution is: ; Where: is the acceleration factor of the corrosion solution concentration; is the sulfate concentration in the etching solution; is the average concentration of sulfate in the actual environment; The concentration of corrosive ions used in the accelerated test; It is the concentration of corrosive ions in the actual service environment of concrete.

[0027] S1.1.3: Introduce freeze-thaw degradation proportionality factor : The ratio between the number of freeze-thaw cycles in the test and the number of freeze-thaw cycles under natural conditions ranges from 1:10 to 1:15, with a value of 1:12, that is, one freeze-thaw cycle in the test is equivalent to 12 freeze-thaw cycles under natural conditions, so the freeze-thaw degradation proportional coefficient Take 12.

[0028] S1.2: Determine the number of days for sulfate dry-wet cycles based on the temperature acceleration factor and the corrosion solution concentration acceleration factor, and determine the number of freeze-thaw cycles based on the freeze-thaw degradation proportional coefficient.

[0029] The freeze-thaw-sulfate wet-dry cycle test system is used to simulate the actual environment for 1 year. The freeze-thaw cycle takes into account the accelerated freeze-thaw degradation, and the sulfate wet-dry cycle takes into account the accelerated temperature and the accelerated concentration of the corrosive solution.

[0030] Number of freeze-thaw cycles during the test: ; Sulfate dry-wet cycle days during the test: ; In this example, Haidong City, Qinghai Province is taken as the research object, where the annual average temperature is 7°C. 2- Content, the actual environmental corrosion ion concentration is taken as 2%.

[0031] The oven temperature is 65°C, the oven drying duration is 6 hours, the annual average temperature of the area in Celsius is 7°C, the sulfate soaking time is 16 hours, and the air drying time is 0.5 hours, then: The average test temperature is: .

[0032] Temperature acceleration factor: ; Acceleration factor of corrosion solution concentration: ; The number of freeze-thaw cycles in Haidong City is 100 to 130 times per year. This example takes the average value of 115 times, and assumes that there is one freeze-thaw cycle per day and night, so the number of days of freeze-thaw in a year is 115 days. Sulfate dry-wet cycle days = number of days in a year - number of freeze-thaw cycle days = 365-115 = 250 days.

[0033] Number of freeze-thaw cycles: 115÷12=9.6 times; number of sulfate dry-wet cycles: 250÷18.1÷2.5=5.5d.

[0034] The concrete specimens were prepared, and the actual adverse environmental conditions and the requirements for the surface flatness of the specimens in the subsequent SHPB impact test were comprehensively considered. Finally, it was determined that a freeze-thaw-sulfate wet-dry cycle test included 8 freeze-thaw cycles and 6 sulfate wet-dry cycles. The number of freeze-thaw-sulfate wet-dry cycles was recorded as N , a total of 7 times, the test system is shown in Figure 2 .

[0035] S1.3: Based on the number of sulfate dry-wet cycle days and freeze-thaw cycles, complete the freeze-thaw-sulfate dry-wet cycle alternation test and collect the quantitative damage indicators during the test: mass loss rate, relative compressive strength and relative dynamic elastic modulus.

[0036] S1.3.1: Based on step 1.2, determine the freeze-thaw-sulfate wet-dry cycle test system.

[0037] Before the first freeze-thaw cycle test, all specimens were immersed in 5% sodium sulfate solution for 4 days to make them fully saturated with water. The freeze-thaw cycle test and sulfate dry-wet cycle test were carried out alternately, and n ( n Before the freeze-thaw cycle test, the concrete specimens were immersed in a 5% sodium sulfate solution for 1 day.

[0038] After soaking, take out the concrete specimens and put them into the freeze-thaw box one by one, and put wooden strips under each concrete specimen to prevent the bottom of the concrete specimen from being affected by the freezing of the residual water at the bottom of the freeze-thaw box. Keep a certain distance between adjacent concrete specimens to avoid sticking together, so that each surface is evenly affected by the temperature changes of the freeze-thaw box.

[0039] A freeze-thaw cycle includes a spraying section, a cooling section, a low-temperature constant temperature section, a heating stage, and a high-temperature constant temperature section. The duration of each stage is 5 minutes, 2 hours, 2 hours, 0.5 hours, and 1.5 hours, for a total of 6 hours and 5 minutes. In the spraying section, the concrete specimen is placed under the spray head of the freeze-thaw box for 5 minutes of water replenishment. The low-temperature constant temperature section of the freeze-thaw box is set to -20°C to ensure that the concrete specimen can complete the freezing process at this temperature. The high-temperature constant temperature section is set to +25°C. After the high-temperature constant temperature section ends, the freeze-thaw box automatically sprays three times, each time for 1 minute, so that the concrete specimen is in a water-retaining state.

[0040] The operation method of the sulfate dry-wet cycle test is as follows: the concrete specimen is immersed in a 5% sulfate solution, and the time from the beginning of the concrete specimen being placed in the solution to the end of the immersion process is 16 hours. After the immersion process is over, the concrete specimen is taken out and placed in a designated location for air drying, and the air drying time is 1 hour. After the air drying process is over, the concrete specimen is placed in a blast oven, the oven switch is turned on, and after the temperature rises to 65°C, the temperature is maintained at about 65°C and dried for 6 hours. After the drying is completed, the concrete specimen is cooled for 1 hour, and the surface temperature of the concrete specimen needs to be cooled to room temperature, and then the above-mentioned freeze-thaw cycle test and sulfate dry-wet cycle test are continued.

[0041] Since the sulfate corrosion product ettringite in the concrete specimen may decompose at 70℃, which will cause the corrosion mechanism to change, it is necessary to ensure that the drying temperature is 65℃±5℃. The drying temperature in this sulfate dry-wet cycle test is 65℃.

[0042] The air-freeze-air-thaw method was performed 8 times (about 2 days) according to the above freeze-thaw cycle system; after the freeze-thaw cycle was completed, the sulfate dry-wet cycle was performed 6 times (about 6 days) according to the above sulfate dry-wet cycle steps.

[0043] S1.3.2: Use an electronic balance to measure the mass of the concrete specimens in the freeze-thaw-sulfate wet-dry cycle test and obtain the mass loss rate : ; Where: For concrete specimens N Mass loss rate after freeze-thaw-sulfate wet-dry cycle; is the initial mass of the concrete specimen; For Experience N Second-rate( N =1, 2, 3, 4, 5, 6, 7) Mass of specimen after freeze-thaw-sulfate wet-dry cycle.

[0044] S1.3.3: Calculate the relative compressive strength of concrete specimens: ; ; In the formula, is the relative compressive strength, It is the pressure applied by the press to the concrete specimen; is the compressive area of ​​the concrete specimen; For Experience N The measured compressive strength of concrete specimens after freeze-thaw-sulfate wet-dry cycles; is the initial measured compressive strength of the concrete specimen.

[0045] S1.3.4: Measure the initial ultrasonic velocity and the ultrasonic N Second-rate( N =1, 2, 3, 4, 5, 6, 7) The ultrasonic velocity of the specimen after freeze-thaw-sulfate wet-dry cycle, and then the relative dynamic elastic modulus of concrete is obtained: ; Where: For concrete specimens N Relative dynamic elastic modulus after freeze-thaw-sulfate drying and wetting cycles; For concrete specimens N Dynamic elastic modulus after freeze-thaw-sulfate wet-dry cycles; is the dynamic elastic modulus before freeze-thaw sulfate cycle V 0 is the initial ultrasonic velocity of the concrete specimen; For Experience N Second-rate( N =1, 2, 3, 4, 5, 6, 7) Ultrasonic velocity of the specimen after freeze-thaw-sulfate wet-dry cycles.

[0046] S2: Perform data preprocessing on the damage quantification indicators in S1 to construct feature data training sets and feature data test sets of damage quantification indicators as well as label data training sets and label data test sets of damage quantification indicators. Specifically including: S2.1: In each freeze-thaw-sulfate wet-dry cycle test, six specimens are used to measure the mass loss rate, relative compressive strength and relative dynamic elastic modulus, calculate the mean and standard deviation of the mass loss rate, relative compressive strength and relative dynamic elastic modulus, and remove the outliers in the damage quantification index to obtain the damage quantification index with missing data points.

[0047] S2.1.1: Calculate the mean values ​​of mass loss rate, relative compressive strength, and relative dynamic elastic modulus respectively: ; ; ; in, represents the mean value of mass loss rate, Represents the mean value of relative compressive strength, represents the mean value of relative dynamic elastic modulus; Indicates i The mass loss rate, Indicates i Relative compressive strength; Indicates i Relative dynamic elastic modulus; S2.1.2: Calculate the standard deviation of mass loss rate, relative compressive strength, and relative dynamic elastic modulus respectively: ; ; ; in, represents the standard deviation of the mass loss rate, represents the standard deviation of relative compressive strength, represents the standard deviation of the relative dynamic elastic modulus; S2.1.3: Based on Principle, the range of mass loss rate is determined as , the relative compressive strength range is , the range of relative dynamic elastic modulus is , remove the outliers outside the range, obtain the normal damage quantification index, and calculate the arithmetic mean of the normal damage quantification index to obtain the average quality loss rate , average relative compressive strength , average value of relative dynamic elastic modulus ; S2.1.4: Based on S2.1.3, the set of arithmetic means of the damage quantification indicators after seven freeze-thaw-sulfate drying and wetting cycles is: Quality loss rate: , relative compressive strength: , relative dynamic elastic modulus: ; S2.2: fill in the missing data points of the damage quantification index in S2.1 by using the interpolation method to obtain the filled damage quantification index, and expand the filled damage quantification index to obtain the expanded damage quantification index; S2.2.1: Use interpolation to fill in missing data points and expand the data for mass loss rate: ; in, represents the number of freeze-thaw-sulfate drying and wetting cycles, , is the mass loss rate after interpolation filling, and is and The known mass loss rate at is to estimate the number of cycles corresponding to the mass loss rate, and All are the freeze-thaw-sulfate dry-wet cycles corresponding to known mass loss rates.

[0048] S2.2.2: Interpolate missing data points for relative compressive strength: ; in, represents the number of freeze-thaw-sulfate drying and wetting cycles, , is the relative compressive strength after interpolation filling, and is and The known relative compressive strength at is to estimate the number of cycles corresponding to the relative compressive strength, and All are the numbers of freeze-thaw-sulfate drying and wetting cycles corresponding to known relative compressive strength.

[0049] S2.2.3: Interpolate missing data points for the relative dynamic elastic modulus: ; in, represents the number of freeze-thaw-sulfate drying and wetting cycles, , is the relative dynamic elastic modulus after interpolation, and is and The known relative dynamic elastic modulus at is to estimate the number of cycles corresponding to the relative dynamic elastic modulus, and All are the freeze-thaw-sulfate dry-wet cycles corresponding to the known relative dynamic elastic modulus.

[0050] S2.2.4: After the damage quantification index data is interpolated, it is expanded (using the interpolation method) to obtain the final set of interpolated mass loss rates: , relative compressive strength , and the set of relative dynamic elastic moduli is ; Indicates j The mass loss rate, Indicates j The relative compressive strength, Indicates j A relative dynamic elastic modulus, in this application, Take 70.

[0051] S2.3: normalize the damage quantification index expanded in S2.2 and summarize them to obtain a normalized damage quantification index set; S2.3.1: Normalize the mass loss rate in the damage quantification index expanded in S2.2 to obtain the normalized mass loss rate, which is calculated as follows: ; in, is the mass loss rate after expansion, is the minimum value of the mass loss rate after expansion, is the maximum value of the mass loss rate after expansion, It is j The normalized mass loss rate; Then the set of normalized mass loss rates is obtained as : ; S2.3.2: The relative compressive strength in the expanded damage quantification index of S2.2 is normalized to obtain the normalized relative compressive strength, which is calculated as follows: ; in, is the relative compressive strength after expansion, is the minimum value of the relative compressive strength after expansion, is the maximum value of the relative compressive strength after expansion, It is j The normalized relative compressive strength; Then the normalized relative compressive strength set is obtained : ; S2.3.3: Normalize the relative dynamic elastic modulus in the expanded damage quantification index of S2.2 to obtain the normalized relative dynamic elastic modulus, which is calculated as follows: ; in, is the relative dynamic elastic modulus after expansion, is the minimum value of the relative dynamic elastic modulus after expansion, is the maximum value of the relative dynamic elastic modulus after expansion, It is j The normalized relative dynamic elastic modulus; Then the normalized relative dynamic elastic modulus set is obtained : .

[0052] S2.4: Divide the damage quantification index set normalized in S2.3 into a training set and a test set.

[0053] S2.4.1: Divide the normalized damage quantification index into groups of five into segments of fixed length; Mass loss rate segmentation for: ;in, n is a positive integer, , For the n The mass loss rate; The mass loss rate feature data set for: ; Relative compressive strength of the segmented fragments for: ,in, n Is a positive integer , For the n The normalized relative compressive strength; The relative compressive strength characteristic data set F for: ; Relative dynamic elastic modulus segmentation for: ; in, n is a positive integer, ; For the n The normalized relative elastic modulus; Then the relative dynamic elastic modulus characteristic data set G for: ; S2.4.2: Use the next value of each damage quantification index segment in S2.4.1 as a label to construct a label data set of damage quantification index; Mass loss rate label dataset : ; in, , The predicted label representing the mass loss rate; Relative compressive strength label dataset : ; in, , Prediction label representing relative compressive strength; Relative dynamic elastic modulus label dataset : ; in, , Prediction label representing the relative dynamic elastic modulus; in, n is a positive integer, .

[0054] S2.4.3: Characteristic datasets of damage quantification indicators V, F, G The label data set of the damage quantification index is divided into a feature data training set and a feature data test set according to the ratio of 8:2. , , Divide them into labeled data training set and labeled data test set in 8:2 ratio respectively; The characteristic data training set of mass loss rate is expressed as , the characteristic data training set of relative compressive strength is expressed as , the characteristic data training set of relative dynamic elastic modulus is expressed as ; The characteristic data test set of mass loss rate is expressed as , the characteristic data test set of relative compressive strength is expressed as , the characteristic data test set of relative dynamic elastic modulus is expressed as ; The labeled data training set of mass loss rate is expressed as , the label data training set of relative compressive strength is expressed as , the label data training set of relative dynamic elastic modulus is expressed as ; The labeled data test set of quality loss rate is expressed as , the label data test set of relative compressive strength is expressed as , the labeled data test set of relative dynamic elastic modulus is expressed as .

[0055] S3: Input the characteristic data training set of the damage quantification index of S2 into the CNN-LSTM-Attention model to obtain the predicted value based on the characteristic data training set of the damage quantification index; Adopt the Adam optimization algorithm, combine the predicted value and the label data training set of the damage quantification index to update the bias and weight of the output layer of the CNN-LSTM-Attention model, and optimize the CNN-LSTM-Attention model; Among them, the first layer of the CNN-LSTM-Attention model is the convolutional neural network layer, the second layer is the long short-term memory neural network layer, the third layer is the attention layer, and the fourth layer is the output layer.

[0056] S3.1: Input the feature data training set of the damage quantification index into the convolutional neural network layer of the CNN-LSTM-Attention model, perform convolution operation, obtain the convolved value, and then pass ReLU The activation function performs nonlinear transformation on the convolutional values ​​to obtain nonlinear transformed values, and performs pooling operation on the nonlinear transformed values ​​to obtain the output result of the convolutional neural network layer; S3.1.1: Characteristic data training of damage quantification indicators V 0.8 , F 0.8 and G 0.8 Enter the convolutional neural network layer and perform convolution operation: ; ; ; in, It is i The value after convolution of the mass loss rate, It is i The value after convolution of relative compressive strength, It is i The value after convolution of the relative dynamic elastic modulus, is the weight of the convolution kernel, b is the bias of the convolution kernel, k is the size of the convolution kernel, j From 1 to k , for V 0.8 , F 0.8 , G 0.8 Middle elements, They are The individual data; S3.1.2: After the convolution operation is completed, ReLU The activation function performs a nonlinear transformation on the convolutional values: ; ; ; in, It is i The value of the mass loss rate after nonlinear transformation, It is i The value after nonlinear transformation of relative compressive strength, It is i The value of the relative dynamic elastic modulus after nonlinear transformation, The function represents the maximum value; S3.1.3: Perform a pooling operation on the mass loss rate value after nonlinear conversion in S3.1.2. The specific calculation process of the pooling operation is as follows: ; ; ; in, It is s The value after the mass loss rate pooling operation, It is s The value after the relative compressive strength pooling operation, It is s The value after the relative dynamic elastic modulus pooling operation; It is i The value of the mass loss rate after nonlinear transformation, It is i The value after nonlinear transformation of relative compressive strength, It is i The value of the relative dynamic elastic modulus after nonlinear transformation, It is i+ 1 The value of mass loss rate after nonlinear conversion, It is i +1 value after nonlinear conversion of relative compressive strength, It is i +1 value after nonlinear conversion of relative dynamic elastic modulus, The function represents the maximum value.

[0057] S3.2: Input the output result of the convolutional neural network layer of S3.1 into the long short-term memory neural network layer to obtain the output result of the long short-term memory neural network layer; wherein the long short-term memory neural network layer includes an input gate, a forget gate, an output gate and a cell state.

[0058] S3.2.1: Input the output of the S3.1 convolutional neural network layer into the forget gate of the long short-term memory neural network layer: ; ; ; yes t The time step quality loss rate is the value after entering the forget gate, yes t The relative compressive strength of the time step is the value after the forget gate is input, yes t The time step relative dynamic elastic modulus is the value after entering the forget gate, is the Sigmoid function, is the weight in the forget gate, are the mass loss rate, compressive strength, and relative dynamic elastic modulus output by the short-term memory neural network layer at the previous time step, is the bias coefficient in the forget gate; S3.2.2: Input the output of the S3.1 convolutional neural network layer into the input gate of the long short-term memory neural network layer: ; ; ; in, yes t The value of the time step mass loss rate after passing through the input gate, yes t The relative compressive strength of the time step after passing through the input gate, yes t The value of the relative dynamic elastic modulus of the time step after passing through the input gate, yes t The weights in the time-step forget gate, for t Bias coefficient in the time step forget gate; S3.2.3: Input the output of the convolutional neural network layer of S3.1 into the candidate memory unit of the cell state: ; ; ; yes t The value of the mass loss rate output by the candidate memory unit at time step, yes t The relative compressive strength value of the candidate memory unit output at the time step, yes t The value of the relative elastic modulus output by the candidate memory unit at time step, is the weight in the forget gate, is the bias coefficient in the forget gate, is the hyperbolic tangent function; S3.2.4: Combine S3.2.1-S3.2.3 and update the damage quantification index based on cell state: ; ; ; in, yes t The time step is based on the mass loss rate after the cell state is updated, yes t The time step is based on the relative compressive strength of the cell state after updating, yes t The time step is based on the relative dynamic elastic modulus after the cell state is updated; S3.2.5: Input the output of the convolutional neural network layer in S3.1 into the output gate: ; ; ; in, yes t The time step mass loss rate is the value after the input and output gate, yes t The time step relative compressive strength input and output gate values, yes t The time step is relative to the dynamic elastic modulus input and output gate value, is the weight in the forget gate, is the bias coefficient in the forget gate; S3.2.6: Based on S3.2.4 and S3.2.5, the output of the long short-term memory neural network layer is: ; ; ; in, is the hyperbolic tangent function, for t The quality loss rate of the output of the short-term memory neural network layer at the time step; for t The relative compressive strength of the output of the short-term memory neural network layer of the time step; for t The relative dynamic elastic modulus of the output of the short-term memory neural network layer at time step, yes t The quality loss rate after the output gate in the time step short-term memory neural network layer, yes t The time step relative compressive strength output gated value, yes t The time step relative dynamic elastic modulus outputs the gated value.

[0059] S3.3: Input the output result of the long short-term memory neural network layer in S3.2 into the attention layer to obtain the output result of the attention layer; S3.3.1: Use a weight matrix The results of the LSTM neural network layer output are weighted to obtain the attention score of the damage quantification index: ; ; ; in, is the hyperbolic tangent function, for The attention score of the time step quality loss rate, for The relative strength of attention of the time step, for The attention score of the time step relative to the dynamic elastic modulus; S3.3.2: Normalize the attention score of the damage quantification index in S3.3.1 through the Softmax function to obtain the attention weight of the normalized damage quantification index: ; ; ; in, for t The attention weight of the time step quality loss rate, for t The attention weights of the time steps relative to the compressive strength, for t The attention weight of the time step relative to the dynamic elastic modulus, For e The exponential function with base , is the total time step, is the current time step; S3.3.3: Based on the attention weights in S3.3.2, the damage quantification indicators output by the long short-term memory neural network layer are weighted and summed to obtain the output result of the attention layer: ; ; ; in, for t The quality loss rate of the attention layer output at the time step, for t The relative compressive strength of the attention layer output at each time step, for t The relative dynamic elastic modulus of the attention layer output at time step; is the total time step, is the current time step.

[0060] S3.4: Input the output result of the attention layer in S3.3 to the output layer, and the output layer outputs the predicted value based on the training set of damage quantification indicator feature data: ; ; ; in, is the bias of the output layer, is the weight of the output layer, is the predicted value of mass loss rate, is the predicted value of relative compressive strength, Predicted value of relative dynamic elastic modulus.

[0061] The Adam optimization algorithm is used to update the bias and weight of the output layer of the CNN-LSTM-Attention model. Specifically, the following steps are performed: (1): Determine the first-order moment estimate of the current traversal number and the second-order moment estimate of the current traversal number: ; ; in, is the gradient of the current traversal times, is the first-order moment estimate of the current traversal number, The first moment estimate of the number of previous passes, is the second-order moment estimate of the current traversal number, is the exponential decay rate of the momentum term, is the exponential decay rate of the variance term; (2): Based on (1), the deviation is corrected to obtain the first-order moment estimate and the second-order moment estimate after eliminating the deviation of the current traversal number: ; ; in, To eliminate the first-order moment estimate after the deviation of the current traversal number, To eliminate the bias of the current traversal times, the second-order moment estimate is obtained; is the exponential decay rate of the momentum term x Second power, is the exponential decay rate of the variance term x Power; (3): Based on (2), update the weight of the output layer of the CNN-LSTM-Attention model and bias , and then optimize the CNN-LSTM-Attention model: ; ; in, It is The weight of the output layer during the traversal, It is The neural network bias of the output layer of the traversal, It is x The neural network bias of the output layer of the traversal, is the learning rate, is a constant used to prevent division by zero errors; (4): Calculate the loss function of the CNN-LSTM-Attention model MSE , update the weight by the number of traversals W and bias b , find the loss function that minimizes: ; ; ; in, MSE 1 is the loss function of the CNN-LSTM-Attention model based on the quality loss rate, MSE 2 is the loss function of the CNN-LSTM-Attention model based on relative compressive strength. MSE 3 is the loss function of the CNN-LSTM-Attention model based on relative dynamic elastic modulus, is the labeled data training set of mass loss rate, is a label data training set with relative compressive strength, is the label data training set of relative dynamic elastic modulus. The label data training set of mass loss rate, the label data training set of relative compressive strength and the label data training set of relative dynamic elastic modulus all belong to the label data training set of damage quantification index. n is the number of elements in the feature data set training set, and the traversal is set to 80 times. The weights and biases finally optimized are assigned to the weights and biases of the output layer of S3.4 respectively.

[0062] In the present application, 8 neurons are set in the output layer of the long short-term memory neural network layer, corresponding to 8 initial output layer weights [0.5159404], [0.2228024], [0.00190622], [0.2992239], [0.5869981], [0.31229842], [0.60286933], [0.64401984]; after 80 times of traversal, the weights of the output layer are [0.5159404], [0.2228024], [0.00190622], [0.29266465], [0.6093647], [0.35189116], [0.61722505], [0.63815206]; The output layer is set with 1 neuron, and the initial bias is [0.00199428]. After 80 traversals, the bias of the output layer is [0.04527584].

[0063] S4: The optimization of the CNN-LSTM-Attention model is completed through the above operations. At this time, the weights and biases of the output layer in the CNN-LSTM-Attention model are optimized and determined. The characteristic data test set of the damage quantification index is input into the optimized CNN-LSTM-Attention model. The output layer of the CNN-LSTM-Attention model outputs the predicted value based on the characteristic data test set of the damage quantification index, and the prediction result of the damage quantification index of the concrete freeze-thaw-sulfate attack process is obtained; the convergence of the CNN-LSTM-Attention model is verified through simulation experiments using the labeled data test set of the damage quantification index.

[0064] Specifically, the characteristic data test set of mass loss rate , characteristic data test set of relative compressive strength , characteristic data test set of relative dynamic elastic modulus The optimized CNN-LSTM-Attention model is input, and the output layer of the CNN-LSTM-Attention model outputs the predicted value of the feature data test set of the damage quantification index.

[0065] After each prediction is completed, the predicted value of the characteristic data test set of the damage quantification index output by the output layer is added to the characteristic data test set of the mass loss rate, and the prediction of the damage quantification index after the next freeze-thaw-sulfate dry-wet cycle is started.

[0066] Test the label data of mass loss rate , a test set of labeled data with relative compressive strength , a test set of labeled data for relative dynamic elastic modulus Different models (LSTM prediction, CNN-LSTM, CNN-LSTM-Attention) were introduced for simulation experiments to verify the convergence of the CNN-LSTM-Attention model. The smaller the calculated MSE, the smaller the convergence and the higher the prediction accuracy.

[0067] Figure 3 , Figure 4 , Figure 5 The comparison between the test values ​​of concrete durability damage index and the true values ​​of the test sample set on the three neural network models (LSTM prediction, CNN-LSTM, and CNN-LSTM-Attention) for the labeled data test set of damage quantification index. Figure 3 In the figure, (a) is LSTM, (b) is CNN-LSTM, and (c) is CNN-LSTM-Attention; Figure 4 In the figure, (a) is LSTM, (b) is CNN-LSTM, and (c) is CNN-LSTM-Attention; Figure 5 In the figure, (a) is LSTM, (b) is CNN-LSTM, and (c) is CNN-LSTM-Attention. The blue line is the actual value of the actual damage index, and the red line is the damage index value predicted by the corresponding model. The following Table 1 shows the actual value and predicted value of the loss quantification index.

[0068] Table 1 The actual and predicted values ​​of loss quantification indicators: ; Based on the predicted values ​​and true values ​​of the loss quantification indicators in Table 1, the MSE of the predicted values ​​and the true values ​​is calculated.

[0069] The quality loss rate is based on the MSE of LSTM prediction = 0.0116, the MSE of CNN-LSTM prediction = 0.00143, and the MSE of CNN-LSTM-Attention prediction = 0.000874.

[0070] The relative compressive strength is based on the MSE of LSTM prediction = 0.0358, the MSE of CNN-LSTM prediction = 0.00408, and the MSE of CNN-LSTM-Attention prediction = 0.00145.

[0071] The relative dynamic elastic modulus is predicted by LSTM with MSE=0.0524, CNN-LSTM with MSE=0.00267, and CNN-LSTM-Attention with MSE=0.001364. The image intuitively shows the advancement of the CNN-LSTM-Attention model over other models.

[0072] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for predicting quantitative indicators of concrete damage during freeze-thaw-sulfate attack, characterized in that: include: S1: During the freeze-thaw-sulfate wet-dry cycle test of concrete specimens, the quantitative damage index of concrete specimens was collected during the test; Among them, the damage quantification indexes include mass loss rate, relative compressive strength and relative dynamic elastic modulus; S2: preprocess the damage quantification index in S1 to construct a feature data training set and a feature data test set of the damage quantification index and a label data training set and a label data test set of the damage quantification index; S3: Input the characteristic data training set of the damage quantification index of S2 into the CNN-LSTM-Attention model to obtain the predicted value based on the characteristic data training set of the damage quantification index; Adopt the Adam optimization algorithm, combine the predicted value and the label data training set of the damage quantification index to update the bias and weight of the output layer of the CNN-LSTM-Attention model, and optimize the CNN-LSTM-Attention model; Among them, the first layer of the CNN-LSTM-Attention model is the convolutional neural network layer, the second layer is the long short-term memory neural network layer, the third layer is the attention layer, and the fourth layer is the output layer; S4: The feature data test set of the damage quantification index is input into the optimized CNN-LSTM-Attention model. The output layer of the CNN-LSTM-Attention model outputs the predicted value based on the feature data test set of the damage quantification index, and the prediction result of the damage quantification index of the concrete freeze-thaw-sulfate attack process is obtained. The labeled data test set of the damage quantification index is used to verify the convergence of the CNN-LSTM-Attention model through simulation experiments.

2. The method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 1, characterized in that: The process of data preprocessing for S2 damage quantification index includes: S2.1: Calculate the mean and standard deviation of the damage quantification index, remove the outliers in the damage quantification index according to the calculated mean and standard deviation, and obtain the damage quantification index with missing data points; S2.2: fill in the missing data points of the damage quantification index in S2.1 by using the interpolation method to obtain the filled damage quantification index, and expand the filled damage quantification index to obtain the expanded damage quantification index; S2.3: normalize the damage quantification index expanded in S2.2 to obtain a normalized damage quantification index set; S2.4: Divide the damage quantification indicator set into a feature data training set and a feature data test set of the damage quantification indicator.

3. The method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 2, characterized in that: S2.3 includes: S2.3.1: Normalize the mass loss rate in the damage quantification index expanded in S2.2 to obtain the normalized mass loss rate, which is calculated as follows: ； in, is the mass loss rate after expansion, is the minimum value of the mass loss rate after expansion, is the maximum value of the mass loss rate after expansion, It is j The normalized mass loss rate; Then the set of normalized mass loss rates is obtained as : ； S2.3.2: The relative compressive strength in the expanded damage quantification index of S2.2 is normalized to obtain the normalized relative compressive strength, which is calculated as follows: ； in, is the relative compressive strength after expansion, is the minimum value of the relative compressive strength after expansion, is the maximum value of the relative compressive strength after expansion, It is j The normalized relative compressive strength; Then the normalized relative compressive strength set is obtained : ； S2.3.3: Normalize the relative dynamic elastic modulus in the expanded damage quantification index of S2.2 to obtain the normalized relative dynamic elastic modulus, which is calculated as follows: ； in, is the relative dynamic elastic modulus after expansion, is the minimum value of the relative dynamic elastic modulus after expansion, is the maximum value of the relative dynamic elastic modulus after expansion, It is j The normalized relative dynamic elastic modulus; Then the normalized relative dynamic elastic modulus set is obtained : 。 4. The method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 3, characterized in that: S3 includes: S3.1: Input the feature data training set of the damage quantification index into the convolutional neural network layer of the CNN-LSTM-Attention model, perform convolution operation, obtain the convolved value, and then pass ReLU The activation function performs nonlinear transformation on the convolutional values ​​to obtain nonlinear transformed values, and performs pooling operation on the nonlinear transformed values ​​to obtain the output result of the convolutional neural network layer; S3.2: Input the output result of the convolutional neural network layer of S3.1 into the long short-term memory neural network layer to obtain the output result of the long short-term memory neural network layer; wherein the long short-term memory neural network layer includes an input gate, a forget gate, an output gate and a cell state; S3.3: Input the output result of the long short-term memory neural network layer in S3.2 into the attention layer to obtain the output result of the attention layer; S3.4: The output result of the attention layer in S3.3 is input into the output layer, and the output layer outputs the predicted value based on the training set of the damage quantification indicator feature data.

5. A method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 4, characterized in that: The specific calculation process of S3.1 pooling operation is: ； ； ； in, It is s The value after the mass loss rate pooling operation, It is s The value after the relative compressive strength pooling operation, It is s The value after the relative dynamic elastic modulus pooling operation; It is i The value of the mass loss rate after nonlinear transformation, It is i The value after nonlinear transformation of relative compressive strength, It is i The value of the relative dynamic elastic modulus after nonlinear transformation, It is i+ 1 The value of mass loss rate after nonlinear conversion, It is i +1 value after nonlinear conversion of relative compressive strength, It is i +1 value after nonlinear conversion of relative dynamic elastic modulus, The function represents the maximum value.

6. A method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 5, characterized in that: In S3.2, the output of the long short-term memory neural network layer is: ； ； ； in, is the hyperbolic tangent function, for t The quality loss rate of the output of the short-term memory neural network layer at the time step; for t The relative compressive strength of the output of the short-term memory neural network layer of the time step; for t The relative dynamic elastic modulus of the output of the short-term memory neural network layer at time step, yes t The quality loss rate after the output gate in the time step short-term memory neural network layer, yes t The time step relative compressive strength output gated value, yes t The time step relative dynamic elastic modulus outputs the gated value.

7. A method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 6, characterized in that: S3.3 includes: S3.3.1: Use a weight matrix The results of the LSTM neural network layer output are weighted to obtain the attention score of the damage quantification index: ； ； ； in, is the hyperbolic tangent function, for The attention score of the time step quality loss rate, for The relative strength of attention of the time step, for The attention score of the time step relative to the dynamic elastic modulus; S3.3.2: Normalize the attention score of the damage quantification index in S3.3.1 through the Softmax function to obtain the attention weight of the normalized damage quantification index: ； ； ； in, for t The attention weight of the time step quality loss rate, for t The attention weights of the time steps relative to the compressive strength, for t The attention weight of the time step relative to the dynamic elastic modulus, For e The exponential function with base , is the total time step, is the current time step; S3.3.3: Based on the attention weights in S3.3.2, the damage quantification indicators output by the long short-term memory neural network layer are weighted and summed to obtain the output result of the attention layer: ； ； ； in, for t The quality loss rate of the attention layer output at the time step, for t The relative compressive strength of the attention layer output at each time step, for t The relative dynamic elastic modulus of the attention layer output at time step; is the total time step, is the current time step.

8. The method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 7, characterized in that: The output result of the S3.3 attention layer is input into the output layer, which outputs the predicted value of the feature data training set based on the damage quantification index: ； ； ； in, is the bias of the output layer, is the weight of the output layer, is the predicted value of mass loss rate, is the predicted value of relative compressive strength, Predicted value of relative dynamic elastic modulus.

9. A method for predicting quantitative indicators of concrete freeze-thaw-sulfate attack damage according to claim 8, characterized in that: The Adam optimization algorithm is used to update the bias and weight of the output layer of the CNN-LSTM-Attention model. Specifically, the following steps are performed: (1): Determine the first-order moment estimate of the current traversal number and the second-order moment estimate of the current traversal number: ； ； in, is the gradient of the current traversal times, is the first-order moment estimate of the current traversal number, The first moment estimate of the number of previous passes, is the second-order moment estimate of the current traversal number, is the exponential decay rate of the momentum term, is the exponential decay rate of the variance term; (2): Based on (1), the deviation is corrected to obtain the first-order moment estimate and the second-order moment estimate after eliminating the deviation of the current traversal number: ； ； in, To eliminate the first-order moment estimate after the deviation of the current traversal number, To eliminate the bias of the current traversal times, the second-order moment estimate is obtained; is the exponential decay rate of the momentum term x Second power, is the exponential decay rate of the variance term x Power; (3): Based on (2), update the weight of the output layer of the CNN-LSTM-Attention model and bias , and then optimize the CNN-LSTM-Attention model: ； ； in, It is x+ The weight of the output layer in one pass, It is The neural network bias of the output layer of the traversal, It is x The neural network bias of the output layer of the traversal, is the learning rate, is a constant; (4): Calculate the loss function of the CNN-LSTM-Attention model MSE , update the weight by the number of traversals W and bias b , find the loss function that minimizes: ； ； ； in, MSE 1 is the loss function of the CNN-LSTM-Attention model based on the quality loss rate, MSE 2 is the loss function of the CNN-LSTM-Attention model based on relative compressive strength. MSE 3 is the loss function of the CNN-LSTM-Attention model based on relative dynamic elastic modulus, is the labeled data training set of mass loss rate, is a label data training set with relative compressive strength, It is the labeled data training set of relative dynamic elastic modulus. The labeled data training set of mass loss rate, the labeled data training set of relative compressive strength and the labeled data training set of relative dynamic elastic modulus all belong to the labeled data training set of damage quantification indicators.

Images