A method for predicting carbonization depth of industrial building concrete

Abstract

The present invention discloses a method for predicting the carbonation depth of industrial building concrete, which obtains historical data of the carbonation depth of industrial buildings under natural environment and divides the data into a training set and a test set; the training set and the test set are adaptively decomposed by using a variational modal decomposition method to obtain modal component subsequences of the training set and the test set; the modal component subsequences of the training set are used as training prediction models to train each modal component subsequence of a gated cyclic unit, and a gated cyclic network carbonation depth prediction submodel of each modal component subsequence of the training set is established; the gated cyclic network carbonation depth prediction submodel of each modal component subsequence of the training set is used to predict the modal component subsequence of the test set, and the predicted values ​​are accumulated and compared with the true values ​​for analysis, so as to complete the prediction and evaluation of the carbonation depth of industrial building concrete. The present invention more carefully grasps the characteristics of the time series of the carbonation depth of industrial buildings, and greatly improves the accuracy of carbonation depth prediction.

Description

Technical Field

[0001] The invention belongs to the technical field of concrete carbonization depth prediction in industrial building durability research, and specifically relates to a concrete carbonization depth prediction method based on a VMD-GRU model. Background Art

[0002] As one of the pillar industries in my country, industry plays a vital role in the country's economic development, and the continuous and stable progress of industrial production activities requires the safe structural performance of industrial buildings as a guarantee. In my country's urban planning, the total land area of ​​industrial buildings accounts for 20% to 35%. At present, the existing industrial buildings in my country are mainly concrete structures with a service age of more than 20 years. They are usually in a harsh environment of high temperature and high humidity under dynamic loads. Long-term service can easily cause structural performance damage and partial function loss. Therefore, in view of the carbonization problem in the structural durability degradation of industrial buildings, the establishment of a concrete carbonization depth prediction model has important practical significance for the durability analysis of concrete structures and the prediction of the remaining life of structures.

[0003] With the rapid development of the national economy, the number of industrial buildings has increased rapidly and the scale has continued to expand. In the entire life cycle of industrial buildings, including from the design and construction stage to the service operation stage, various imperceptible tiny factors will cause different carbonation depths per unit time. The prediction of concrete carbonation depth can evaluate and predict the degree of building damage and remaining life, so the prediction of concrete carbonation depth is very necessary for industrial buildings. Effective prediction has important practical significance for the structural durability analysis and structural remaining life prediction of industrial buildings. Domestic and foreign scholars have conducted in-depth research on concrete carbonation prediction, but they are all based on diffusion theory and a combination of theoretical experience. The uncertainty and randomness of the carbonization process under the coupling of various factors are not considered, and the prediction accuracy is different from the actual situation. Therefore, it is very necessary to study effective concrete carbonation prediction methods, which will contribute to the continued health of the industry. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the carbonation depth of industrial building concrete in view of the deficiencies in the above-mentioned prior art, so as to solve the problem that there is a lack of effective method for predicting the carbonation depth during the durability degradation of existing industrial buildings.

[0005] The present invention adopts the following technical solutions:

[0006] A method for predicting the carbonation depth of industrial building concrete comprises the following steps:

[0007] S1. Obtain historical data of carbonization depth of industrial buildings in natural environment, and then divide the historical data into training set and test set;

[0008] S2, using the variational modal decomposition method to adaptively decompose the training set and the test set in step S1, extracting the detailed features of the original sequence, and then using phase space reconstruction to process and optimize the decomposed modal component data to obtain the modal component subsequences of the training set and the test set;

[0009] S3, using the training set modal component subsequence of the variational modal decomposition in step S2 as a training prediction model, adopting the backward error propagation method to train each modal component subsequence of the gated recurrent unit, using the loss function to adjust and update the weight parameters and offset parameters, completing the deep network learning, and establishing the gated recurrent network carbonization depth prediction submodel of each modal component subsequence of the training set;

[0010] S4. Use the gated recurrent network carbonization depth prediction sub-model of each modal component sub-sequence of the training set obtained in step S3 to predict the modal component sub-sequence of the test set in step S2, accumulate the predicted values ​​and compare and analyze them with the true values ​​to complete the prediction and evaluation of the carbonization depth of industrial building concrete.

[0011] Specifically, in step S1, the data sequence is divided into a training set and a test set in a ratio of 7:3.

[0012] Specifically, step S3 is as follows:

[0013] S301, decompose the original data and initialize each mode u k , center frequency ω, operator λ and maximum number of iterations N;

[0014] S302, update parameter u k , ω, λ and the maximum number of iterations N;

[0015] S303, for a given discrimination accuracy e>0, if it does not meet Then stop the iteration, otherwise return to step S302;

[0016] S304, using phase space reconstruction to process and optimize the modal component data obtained by decomposition, so that the reconstructed samples can better reflect the carbonization changes of industrial buildings under natural conditions;

[0017] S305, using the modal components of the high-dimensional feature space reconstructed in step S304 as training samples to substitute into the gated recurrent network for learning and training, each modal component learns more knowledge about carbonization depth prediction from the backward error propagation update formula of the gated recurrent network, and obtains an accurate carbonization depth prediction model for each component. ( ' n) , and finally, superimpose the predicted values ​​of each component to obtain the final predicted value and output the result;

[0018] S306: Output the prediction result of the VMD-GRU model and compare it with the true value.

[0019] Furthermore, in step S301, the carbonization history data is decomposed into three amplitude-frequency modulation sub-signals using variational modes, and the frequency center and bandwidth of each decomposed component are determined by iteratively searching for the optimal solution of the variational model to achieve effective separation of each component.

[0020] Furthermore, in step S301, the input parameter quantity is 1, which is the historical carbonization data of the service life of industrial buildings in various stages under natural environment; the middle hidden layer is determined as 3 components according to variational mode decomposition; and the output parameter quantity is 1, that is, the predicted carbonization depth value.

[0021] Further, in step S302, the update is as follows:

[0022]

[0023]

[0024]

[0025] in, and Represent f(ω), μ i (ω), λ(ω), and Fourier transform of , n represents the number of iterations.

[0026] Furthermore, in step S304, a reconstruction method is used to construct a high-dimensional space modal component sequence. For the modal component sequence of the historical carbonization historical data decomposed by the variational mode, x={x i / i=1,2,3,...,N} predicts the phase space reconstruction of h during different service periods, specifically:

[0027]

[0028] Where L = N-(d-1)·τ-h, N is the total number of historical carbonization depth samples, τ and d are the delay time and embedding dimension, respectively, and X i (i=1,2,…,L) represents the i-th space vector in the phase space.

[0029] Furthermore, in step S305, the final predicted value o' is:

[0030] o'=o' (1) +o' (2) +…+o' (n)

[0031] Among them, o' (n) Predicted values ​​of carbonization depth for each modal component.

[0032] Furthermore, the back propagation update formula of the GRU network is:

[0033] z t =sigm(w xz x t +w hz x t-1 +b z )

[0034] r t =sigm(w xr x t +w hr x t-1 +b r )

[0035]

[0036]

[0037] o t =sigmoid(w o h t +b o )

[0038] Among them, z t is the update gate, which controls the extent to which the state information at time t-1 is brought into the current state, r t To reset the gate, control the degree to which the state information at time t-1 is ignored; is the candidate hidden state, h t is the output at time t, b z is the update gate offset parameter, ⊙ is the Hadamard product, that is, multiplying the corresponding elements, x t-1 is the input of the hidden state at time t-1, o t is the output value at time t, b o is the offset parameter at time t.

[0039] Compared with the prior art, the present invention has at least the following beneficial effects:

[0040] The invention provides a prediction model for carbonization of industrial building concrete. Firstly, the historical carbonization depth data under natural conditions collected in actual projects are used to obtain the carbonization history records of industrial buildings. A variational mode decomposition (VMD) method is introduced to decompose the original data sequence into K (1, 2, ... N) modal variables, and the phase space reconstruction is used to process and optimize the modal component data obtained by the decomposition, so that the reconstructed samples can better reflect the carbonization changes of industrial buildings under natural conditions. Then, the weight parameters and offset parameters are continuously adjusted and updated through the loss function and back propagation of the gated recurrent unit GRU, the knowledge of carbonization depth under natural conditions is learned, and a sub-model for predicting carbonization depth is established, so as to overcome the problems of gradient vanishing and gradient explosion of recurrent neural networks (RNNs), slow learning speed and difficulty in training, and inability to fully utilize information at each moment.

[0041] Furthermore, the GRU network method is used to train a data-driven prediction model without accurately analyzing the physical and chemical reactions within the carbonization process. This overcomes the shortcomings of the existing carbonization depth prediction model, which is low in accuracy and easily affected by various factors. The carbonization depth results can be predicted quickly and accurately. At the same time, the VMD model is used to determine the number of modal decompositions of the data according to the actual situation of the data, to achieve effective separation of the inherent modes, and to perform adaptive decomposition. After effectively extracting the detailed features of the original sequence, the characteristics of the carbonization depth time series of industrial buildings can be grasped in more detail, the original sequence can be stabilized, and the accuracy of carbonization depth prediction can be greatly improved.

[0042] Furthermore, in step S1, the training set is used to train the carbonization depth prediction model and determine the parameters in the model, such as weight parameters and offset parameters. After the prediction model is obtained from the training set, the test set is used to perform model prediction and performance evaluation.

[0043] Furthermore, step S3 uses the gated recurrent network GRU method to train the carbonization depth prediction model because this method has fewer parameters, requires fewer samples, has a faster training speed, and avoids the problem of long-term memory gradient dissipation in traditional recurrent neural networks RNN. They can preserve information in long-term sequences and will not be cleared over time or removed due to irrelevance to the prediction, which is beneficial for processing events with very long time series delays.

[0044] Furthermore, the variational mode decomposition (VMD) method in step S301 can determine the number of modal decompositions according to the actual data conditions, and can adaptively match the optimal center frequency and bandwidth of each mode during the search and solution process, and can achieve effective separation of intrinsic mode components (IMFs) and frequency domain division of signals, thereby obtaining effective decomposition components of a non-strongly linear time series with low actual data complexity.

[0045] Furthermore, in step S304, the phase space reconstruction processes and optimizes the modal component data obtained by the decomposition, so that the reconstructed samples better reflect the carbonization changes of industrial buildings under natural conditions.

[0046] Furthermore, the backward propagation of the gated recurrent network GRU allows it to continuously adjust and update the weight parameters and offset parameters through the loss function within the allowable error range, accurately learn the relevant knowledge of carbonization depth prediction under natural conditions, and minimize the error between the predicted value and the true value.

[0047] In summary, the present invention does not accurately analyze the physicochemical reactions inside the concrete carbonization process, overcomes the shortcomings of the existing carbonization depth prediction model, which is low in accuracy and easily affected by various factors, and can quickly and accurately predict the carbonization depth result; the VMD model determines the number of modal decompositions of carbonization data according to actual conditions, realizes the effective separation of inherent modes, performs adaptive decomposition, and effectively extracts the detailed features of the original sequence, which can more finely grasp the characteristics of the industrial building carbonization depth time series, and greatly improves the accuracy of carbonization depth prediction.

[0048] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart for realizing the present invention;

[0050] Figure 2 VMD decomposition and reconstruction of carbonization depth;

[0051] Figure 3 This is a flow chart of the VMD-GRU carbonization depth prediction model;

[0052] Figure 4 This is the schematic diagram of the GRU model;

[0053] Figure 5 IMF1-IMF3 modal components after VMD decomposition of historical data of carbonization depth under natural conditions;

[0054] Figure 6 This is a comparison chart of four carbonization depth prediction methods. DETAILED DESCRIPTION

[0055] See also Figure 1 The present invention provides a method for predicting the carbonization depth of industrial building concrete, comprising the following steps:

[0056] S1. Obtain historical data on carbonization depth of industrial buildings in natural environments, and randomly divide them into training sets and test sets according to relevant proportions (this ratio is usually used when the data samples are less than 10,000), where the ratio is 7:3.

[0057] S2. Adaptively decompose the training set and test set in step S1 using the variational mode decomposition (VMD) method to effectively extract the detailed features of the original sequence, reduce the unsteadiness of the original signal, and improve the prediction accuracy; then use phase space reconstruction to process and optimize the modal component data obtained by decomposition, so that the reconstructed samples can better reflect the carbonization changes of industrial buildings under natural conditions, such as Figure 2 As shown;

[0058] Variational mode decomposition is an adaptive modal variation and signal processing method that can determine the number of modal decompositions of given signals and data according to actual conditions and achieve effective separation of inherent modes.

[0059] VMD uses quadratic penalty and Lagrange multipliers to transform the above constrained problem into an unconstrained problem, and uses the alternating direction multiplier method to solve this unconstrained problem. By iteratively searching the variational model to determine the mode uk(t) and its corresponding center frequency and bandwidth, the signal is adaptively decomposed into several IMF components.

[0060] S3, using the training set modal component subsequence decomposed by VMD in step S2 to train the prediction model, using the backward error propagation method to train each modal component subsequence of the gated recurrent unit (GRU), using the loss function to adjust and update the weight parameters and offset parameters, completing deep network learning, and establishing the GRU carbonization depth prediction submodel of each modal component;

[0061] The Gated Recurrent Unit structure consists of an update gate, a reset gate, and a hidden layer. The reset gate and the update gate are sigmoid functions. It is also a variant of the long short-term memory LSTM, which alleviates the gradient disappearance or explosion phenomenon. At the same time, its structure and calculation are simpler and faster.

[0062] GRU is used to build a carbonization depth prediction model to improve the accuracy and speed of prediction. Because GRU training speed is slightly faster and requires a small amount of data to generalize, it can greatly improve the training efficiency and iteration speed. If the amount of data is sufficient, its powerful expression ability can produce better prediction results. The optimization process includes VMD decomposition, parameter update, iterative accuracy judgment, decomposition variable reconstruction, and VMD-GRU model establishment. Specifically:

[0063] S301, decomposition of original data VMD, initialization of mode u k , center frequency ω and bandwidth λ;

[0064] The carbonization historical data are decomposed into three amplitude-frequency modulated sub-signals using variational mode, and the frequency center and bandwidth of each decomposition component are determined by iteratively searching for the optimal solution of the variational model to achieve effective separation of each component.

[0065] The input parameter is 1, which is the historical carbonization data of industrial buildings at various service stages under natural environment; the middle hidden layer is determined as 3 components according to variational mode decomposition; the output parameter is 1, which is the predicted carbonization depth value.

[0066] S302: Update the parameter u according to equations (1), (2) and (3). k , ω and λ;

[0067]

[0068]

[0069]

[0070] in, and Represent f(ω), μ i (ω), λ(ω), and Fourier transform of , n represents the number of iterations.

[0071] S303, for a given discrimination accuracy e>0, if it does not meet And stop iteration, otherwise return to S302.

[0072] S304, using phase space reconstruction to process and optimize the modal component data obtained by decomposition, so that the reconstructed samples can better reflect the carbonization changes of industrial buildings under natural conditions;

[0073] S305, using the modal components of the high-dimensional feature space reconstructed in step S404 as training samples to be brought into the GRU network for learning and training, each modal component learns more knowledge about carbonization depth prediction from the backward error propagation update formula (5) of the GRU network, and obtains an accurate carbonization depth prediction model for each component. ( ' n) , and finally all component prediction values ​​are superimposed to obtain the final prediction value and the output result is as shown in formula (6);

[0074]

[0075] o'=o' (1) +o' (2) +…+o' (n) (6)

[0076] S306, output the prediction result of the VMD-GRU model and compare it with the true value;

[0077] S4. Use the VMD-GRU carbonization depth prediction sub-model obtained in step S3 to predict the test set modal component sub-sequence in step S2, accumulate the predicted values ​​of each sub-model to complete the reconstruction, and then compare and analyze them with the measured values ​​to complete the prediction of carbonization depth of industrial building concrete.

[0078] The specific construction method of VMD-GRU carbonization prediction model:

[0079] 1) VMD decomposition and reconstruction of related variables of the original data: decompose the original data into three amplitude-frequency modulation sub-signals, and determine the frequency center and bandwidth of each modal component by iteratively searching for the optimal solution of the variational model to achieve effective separation of each component; then use phase space reconstruction to process and optimize the data of each modal component obtained by decomposition, so that the reconstructed samples can better reflect the carbonization changes of industrial buildings under natural conditions;

[0080] 2) Determine the structure of the GRU prediction model: The GRU prediction model consists of three parts, namely the reset gate, the update gate, and the hidden layer; the reset gate and the update gate are sigmoid functions.

[0081] 3) Determination of VMD-GRU prediction model parameters: The input parameter is 1, namely the historical carbonization data under the natural environment; the middle hidden layer is determined as 3 modal components according to VMD decomposition; the output parameter is 1, namely the predicted carbonization depth value.

[0082] S5. Compare and analyze the carbonation depth superposition prediction value obtained in step S4 with the measured value to complete the comparison of the carbonation depth prediction results of industrial building concrete.

[0083] Step S4 puts each modal component subsequence in the test set into the optimal prediction model of the gated recurrent unit GRU, which has trained various offset parameters and weight parameters using the backward error propagation algorithm and loss function, and uses each component prediction model to obtain the predicted value of each component of the test set. Finally, the predicted values ​​are superimposed and compared with the true values.

[0084] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can usually be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0085] See also Figure 6 By using GRU, ELM, VMD-ELM network prediction model and VMD-GRU prediction model for comparative analysis, Figure 6 It can be seen that the VMD-GRU prediction model grasps the law of carbonization depth in a shorter time than other prediction models, and the predicted results are more closely aligned with the true values.

[0086] The prediction effect of VMD-GRU model on the test set is compared with that of single model ELM and GRU model. Figure 6 As shown in Table 1, the proposed VMD-GRU model shows obvious advantages over the single ELM and GRU models in predicting the carbonization depth with strong non-stationarity and volatility. In addition, the evaluation indicators such as MAE, RMSE and MAPE of the single model vary greatly, and some values ​​differ by more than 100. Figure 5 The goodness of fit test of the single model did not reach half, but the errors of each indicator of the proposed model were small.

[0087] Table 1 is the prediction effect evaluation table of each model test set

[0088]

[0089] like Figure 5 As shown in Table 1, the prediction effect of the GRU model is slightly better than that of the ELM model, thanks to the GRU complex gating structure that effectively captures the dependencies with large intervals in the middle of the time series data; at the same time, as shown in Table 2, the prediction performance of the VMD-ELM model obtained after variational mode decomposition data preprocessing and phase space reconstruction is much higher than that of the ELM model, which verifies the high efficiency of the preprocessed data.

[0090] In summary, the method for predicting carbonization depth of industrial building concrete of the present invention has the following specific features:

[0091] The GRU network method is used to establish a data-driven carbonization prediction model. It does not accurately analyze the physical and chemical reactions in the carbonization process, overcomes the shortcomings of the existing carbonization depth prediction model, which is not accurate and easily affected by various factors, and can quickly and accurately predict the carbonization depth results; at the same time, the VMD method can be used to adaptively decompose the original carbonization data and reconstruct the phase space of each modal component, so that the characteristics of the industrial building carbonization depth time series can be more carefully grasped, the original series can be stabilized, and the accuracy of carbonization depth prediction can be improved. Finally, the predicted values ​​of each component prediction model are superimposed to obtain the final predicted value, and the test set is used to test the obtained results and compare them with the measured values.

[0092] In view of the fact that the carbonization process in the natural environment is affected by multiple factors and the time series of carbonization depth is usually non-stationary and nonlinear, the present invention adopts variational mode decomposition (VMD) to adaptively decompose the original carbonization data and then reconstruct it using phase space, which can more finely grasp the time series characteristics of carbonization depth and achieve data series stabilization, so that the reconstructed samples can better reflect the actual carbonization changes, so as to prepare the GRU input; combined with the characteristics of gated recurrent unit (GRU) training requiring less data and fast training speed; the VMD-GRU carbonization depth prediction model improves the accuracy and speed of carbonization prediction.

[0093] The use of VMD can avoid the modal aliasing phenomenon in the data processing process, meet the actual needs of the project, and has important practical significance for the structural durability analysis and structural remaining life prediction of industrial buildings.

[0094] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting carbonization depth of industrial building concrete, characterized in that: The following steps are involved: S1. Obtain historical data of carbonization depth of industrial buildings in natural environment, and then divide the historical data into training set and test set; S2, using the variational modal decomposition method to adaptively decompose the training set and the test set in step S1, extracting the detailed features of the original sequence, and then using phase space reconstruction to process and optimize the decomposed modal component data to obtain the modal component subsequences of the training set and the test set; S3, using the training set modal component subsequence of the variational modal decomposition in step S2 as a training prediction model, adopting the backward error propagation method to train each modal component subsequence of the gated recurrent unit, using the loss function to adjust and update the weight parameters and offset parameters, completing the deep network learning, and establishing the gated recurrent network carbonization depth prediction submodel of each modal component subsequence of the training set; S4. Use the gated recurrent network carbonization depth prediction sub-model of each modal component sub-sequence of the training set obtained in step S3 to predict the modal component sub-sequence of the test set in step S2, accumulate the predicted values ​​and compare and analyze them with the true values ​​to complete the prediction and evaluation of the carbonization depth of industrial building concrete.

2. The method for predicting carbonization depth of industrial building concrete according to claim 1, characterized in that: In step S1, the data sequence is divided into a training set and a test set in a ratio of 7:

3.

3. The method for predicting carbonization depth of industrial building concrete according to claim 1, characterized in that: Step S3 is specifically as follows: S301, decompose the original data and initialize each mode u k , center frequency ω, operator λ and maximum number of iterations N; S302, update parameter u k , ω, λ and the maximum number of iterations N; S303, for a given discrimination accuracy e>0, if it does not satisfy Then stop the iteration, otherwise return to step S302; S304, using phase space reconstruction to process and optimize the modal component data obtained by decomposition, so that the reconstructed samples can better reflect the carbonization changes of industrial buildings under natural conditions; S305, using the modal components of the high-dimensional feature space reconstructed in step S304 as training samples to substitute into the gated recurrent network for learning and training, each modal component learns more knowledge about carbonization depth prediction from the backward error propagation update formula of the gated recurrent network, and obtains an accurate carbonization depth prediction model for each component. ( ' n) , and finally, superimpose the predicted values ​​of each component to obtain the final predicted value and output the result; S306: Output the prediction result of the VMD-GRU model and compare it with the true value.

4. The method for predicting carbonization depth of industrial building concrete according to claim 3, characterized in that: In step S301, the carbonization history data is decomposed into three amplitude-frequency modulation sub-signals using variational mode, and the frequency center and bandwidth of each decomposition component are determined by iteratively searching for the optimal solution of the variational model to achieve effective separation of each component.

5. The method for predicting carbonization depth of industrial building concrete according to claim 3, characterized in that: In step S301, the input parameter quantity is 1, which is the historical carbonization data of industrial buildings in various service stages under natural environment; the middle hidden layer is determined as 3 components according to variational mode decomposition; the output parameter quantity is 1, that is, the predicted carbonization depth value.

6. The method for predicting carbonization depth of industrial building concrete according to claim 3, characterized in that: In step S302, the update is as follows: in, and Represent f(ω), μ i (ω), λ(ω), and Fourier transform of , n represents the number of iterations.

7. The method for predicting carbonization depth of industrial building concrete according to claim 3, characterized in that: In step S304, a reconstruction method is used to construct a high-dimensional space modal component sequence. For the modal component sequence of the historical carbonization historical data decomposed by the variational mode, x = {x i / i=1,2,3,…,N} predicts the phase space reconstruction of h during different service periods, specifically: Where L = N-(d-1)·τ-h, N is the total number of historical carbonization depth samples, τ and d are the delay time and embedding dimension, respectively, and X i Represents the i-th space vector in the phase space, i=1,2,…,L.

8. The method for predicting carbonization depth of industrial building concrete according to claim 3, characterized in that: In step S305, the final predicted value o' is: o'=o'(1)+o'(2)+...+o'(n) Among them, ( ' n) Predicted values ​​of carbonization depth for each modal component.

9. The method for predicting carbonization depth of industrial building concrete according to claim 8, characterized in that: The back propagation update formula of the GRU network is: z t =sigm(w xz x t +w hz x t-1 +b z ) r t =sigm(w xr x t +w hr x t-1 +b r ) o t =sigmoid(w o h t +b o ) Among them, z t is the update gate, which controls the extent to which the state information at time t-1 is brought into the current state, r t To reset the gate, control the degree to which the state information at time t-1 is ignored; is the candidate hidden state, h t is the output at time t, b z is the update gate offset parameter, ⊙ is the Hadamard product, that is, multiplying the corresponding elements, x t-1 is the input of the hidden state at time t-1, o t is the output value at time t, b o is the offset parameter at time t.

Images