High fill roadbed settlement prediction method based on neural network

Abstract

The invention discloses a high-fill roadbed settlement prediction method based on a neural network, and relates to the technical field of civil engineering construction operation and maintenance, and the method comprises the following steps: S1, constructing a relation function between roadbed filler settlement and filling height; s2, dividing roadbed filling height data and settlement time sequence data after cleaning into a training set and a test set; s3, constructing a settlement prediction model; s4, performing model training; and S5, carrying out model testing until the model testing is passed, and determining a high-fill roadbed settlement prediction result of the to-be-predicted engineering field by using the settlement prediction model passing the testing. According to the method, an original physical information neural network is improved, a network structure in which 1D convolution and causal convolution are arranged in a staggered manner is adopted, time sequence features of monitoring data are effectively extracted, the time causality of the monitoring data is ensured, and a new thought is provided for railway high-fill roadbed settlement prediction.

Description

Technical Field

[0001] The invention relates to the technical field of civil engineering construction operation and maintenance, and in particular to a high fill roadbed settlement prediction method based on a neural network. Background Art

[0002] Against the backdrop of my country's continuous socio-economic development, the development of the railway industry is also entering a new stage. High-speed railway construction technology takes the continuous improvement of operating speed as one of its development goals, and puts forward strict requirements on the stability of the roadbed. The settlement of the roadbed after construction directly affects the safety, stability and service life of train operations. Compared with general roadbeds, high-fill roadbeds are more prone to settlement deformation due to their large filling height and heavy weight, coupled with the external train load. These characteristics make high-fill roadbeds have higher filling quality requirements and stricter settlement control requirements. How to reasonably use observation data and accurately and reliably predict the settlement deformation of high-fill roadbeds has become one of the hot topics in engineering research.

[0003] The prediction of high-speed railway subgrade settlement includes three categories: traditional prediction methods, numerical analysis prediction methods, and settlement prediction methods based on measured data. When faced with the characteristics of subgrade settlement data, such as "small settlement deformation and large data fluctuation", traditional prediction methods and numerical analysis prediction methods find it difficult to consider issues such as soil compression, soil consolidation, and model parameter determination. In contrast, settlement prediction methods based on measured data are more advantageous in practice. Settlement prediction methods based on measured data include hyperbolic method, three-point method, exponential curve method, and gray system. With the development of computer technology, artificial intelligence technology has brought new ideas to solve the problem of subgrade settlement prediction. Machine learning algorithms have been applied to subgrade settlement prediction, and recurrent neural network models and their variant network models are mostly used. Traditional machine learning methods use data-driven methods to predict subgrade settlement, use measured data for training and learning, and give prediction results. This pure data-driven prediction method requires a large number of training samples, and observation bias leads to poor model generalization ability. Therefore, a dual-driven neural network that integrates physical information and data information is proposed. Considering the inherent physical mechanism, it makes up for the defect that the prediction results in pure data-driven neural networks are not highly consistent with the laws of actual physical and mechanical models. In the implementation of time series data prediction tasks, the physical information dual-driven neural network can embed the mapping relationship between high fill roadbed settlement deformation and filling height into the loss function, which helps to ensure that the prediction results conform to the physical process of roadbed settlement development and generalize unseen data to a certain extent, which is particularly important for long-term prediction of time series data. However, the traditional physical data dual-driven neural network performs poorly in processing the time dependency of time series data in time series prediction tasks, resulting in poor capture of the time series characteristics of monitoring data, affecting the reliability and accuracy of roadbed settlement prediction. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a high fill roadbed settlement prediction method based on neural network.

[0005] The technical solution of the present invention is: a method for predicting the settlement of a high fill roadbed based on a neural network comprises the following steps:

[0006] S1. According to the subgrade settlement and filling height of the subgrade filling material at various compaction degrees, a relationship function between the subgrade settlement and filling height is constructed;

[0007] S2, collecting the roadbed filling height data and settlement time series data of the project site to be predicted, and performing data cleaning on the settlement time series data, and dividing the roadbed filling height data and the cleaned settlement time series data into a training set and a test set;

[0008] S3. Construct a settlement prediction model based on the relationship function between roadbed filler settlement and filling height;

[0009] S4, using the training set as the input of the settlement prediction model to perform model training;

[0010] S5. Use the test set to perform model testing on the trained settlement prediction model until the model test passes, and use the settlement prediction model that passes the test to determine the high fill roadbed settlement prediction result of the engineering site to be predicted.

[0011] Furthermore, in S1, the relationship function f(ξ,h) between the settlement of roadbed fill and the filling height is expressed as: f(ξ,h)=ue -h / t +ξ i ; In the formula, ξ i represents the first coefficient, u represents the second coefficient, t represents the third coefficient, h represents the filling height, and e represents the exponent.

[0012] Furthermore, S2 includes the following sub-steps:

[0013] S21, collecting roadbed filling height data and settlement time series data at the project site to be predicted;

[0014] S22, performing singular value test, missing value interpolation, data noise reduction and data standardization on the sedimentation time series data in sequence;

[0015] S23, dividing the roadbed filling height data and the settlement time series data after cleaning into a training set and a test set.

[0016] Furthermore, in S22, the calculation formula for data noise reduction processing is:

[0017] s(t)=s d (t)+s k (t);

[0018] In the formula, s(t) represents the sedimentation time series data after data noise reduction, s d (t) represents the actual deformation signal, s k (t) represents deterministic noise.

[0019] Further, the sedimentation prediction model includes an input layer, a first 1D convolutional layer, a first ReLU activation layer, a first causal convolutional layer, a second ReLU activation layer, a second 1D convolutional layer, a third ReLU activation layer, a second causal convolutional layer, a fourth ReLU activation layer, a linear layer, and an output layer;

[0020] The input end of the input layer serves as the input end of the settlement prediction model; the output end of the input layer, the first 1D convolution layer and the input end of the first ReLU activation layer are connected in sequence; the first output end of the first ReLU activation layer, the first causal convolution layer and the input end of the second ReLU activation layer are connected in sequence; the second output end of the first ReLU activation layer and the first output end of the second ReLU activation layer are both connected to the input end of the second 1D convolution layer; the output end of the second 1D convolution layer is connected to the input end of the third ReLU activation layer; the second output end of the second ReLU activation layer and the first output end of the third ReLU activation layer are both connected to the input end of the second causal convolution layer; the output end of the second causal convolution layer is connected to the input end of the fourth ReLUc activation layer; the second output end of the third ReLU activation layer and the output end of the fourth ReLU activation layer are both connected to the input end of the linear layer; the output end of the linear layer is connected to the input end of the output layer; the output end of the output layer serves as the output end of the settlement prediction model.

[0021] Furthermore, the expression between the input layer and the output layer is:

[0022] Y = N(X; θ);

[0023] Where θ = [w, b], X represents the input variable of the settlement prediction model, Y represents the output variable of the settlement prediction model, θ represents the parameter to be identified of the settlement prediction model, N(·) represents the neural network operator of the parameter to be identified, w represents the weight between different network layers, and b represents the bias of the neural network.

[0024] Furthermore, the expression of the loss function loss_g of the settlement prediction model is:

[0025] loss_g=λ1loss_g u +λ2loss_g f ;

[0026]

[0027] Where loss_g uRepresents the data-driven term, loss_g f represents the residual term of the physical equation, λ1 represents the first weight coefficient, λ2 represents the second weight coefficient, N represents the total number of samples, s' i represents the observed data, represents the network prediction value, h i represents the filling height of the roadbed, and f(ξ,h) represents the relationship function between the settlement of the roadbed fill and the filling height.

[0028] Furthermore, the method for determining the first weight coefficient and the second weight coefficient in the loss function of the settlement prediction model is: constructing an adaptive optimization function, and inputting the current first weight coefficient and the current second weight coefficient into the adaptive optimization function; if the adaptive optimization function value is greater than 0.05, then increasing the current first weight coefficient or the current second weight coefficient until the adaptive optimization function value is less than or equal to 0.05; otherwise, using the current first weight coefficient and the current second weight coefficient.

[0029] Furthermore, the expression of the adaptive optimization function A is:

[0030] And λ1+λ2=1;

[0031] Where loss_g u Represents the data-driven term, loss_g f represents the residual term of the physical equation, λ1 represents the first weight coefficient, λ2 represents the second weight coefficient, max(·) represents the maximum value function, and min(·) represents the minimum value function.

[0032] The beneficial effects of the present invention are as follows: the present invention is different from the previous roadbed settlement prediction model. It considers the evolution of mechanical characteristics under the load of high-fill roadbed filling material, obtains the physical and mechanical model of roadbed filling material through experimental fitting, uses data information as the driving force of the settlement prediction model, and uses the physical information control equation as the constraint of the settlement prediction model to construct a settlement prediction model, which reduces the excessive dependence of the settlement prediction model on the amount of training data and reduces the data collection cost; effectively corrects the deviation of the settlement prediction model in the prediction of roadbed settlement, improves the prediction accuracy of the settlement prediction model, and makes the model prediction results conform to the physical and mechanical change law of the roadbed structure. The present invention improves the original physical information neural network, adopts a network structure with 1D convolution and causal convolution interlaced, effectively extracts the time series characteristics of the monitoring data, and ensures its time causality. At the same time, the present invention provides a new idea for the prediction of railway high-fill roadbed settlement, which has certain significance for railway roadbed operation and maintenance and the prevention and control of uneven settlement diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of the high fill roadbed settlement prediction method based on neural network;

[0034] Figure 2 It is the structural diagram of the settlement prediction model;

[0035] Figure 3 Compare the data cleaning charts;

[0036] Figure 4 Schematic diagram of the prediction results of the settlement prediction model. DETAILED DESCRIPTION

[0037] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0038] like Figure 1 As shown, the present invention provides a high fill roadbed settlement prediction method based on neural network, comprising the following steps:

[0039] S1. According to the subgrade settlement and filling height of the subgrade filling material at various compaction degrees, a relationship function between the subgrade settlement and filling height is constructed;

[0040] S2, collecting the roadbed filling height data and settlement time series data of the project site to be predicted, and performing data cleaning on the settlement time series data, and dividing the roadbed filling height data and the cleaned settlement time series data into a training set and a test set;

[0041] S3. Construct a settlement prediction model based on the relationship function between roadbed filler settlement and filling height;

[0042] S4, using the training set as the input of the settlement prediction model to perform model training;

[0043] S5. Use the test set to perform model testing on the trained settlement prediction model until the model test passes, and use the settlement prediction model that passes the test to determine the high fill roadbed settlement prediction result of the engineering site to be predicted.

[0044] In S4-S5, the test set after data cleaning is used as the input of the physical information neural network to obtain the roadbed settlement prediction results, and the accuracy of the prediction results is evaluated. The evaluation indicators include mean square error, root mean square error, mean absolute error and mean absolute percentage error. The accuracy evaluation value of the settlement prediction model is tested to determine whether it meets the engineering prediction accuracy requirements. If the mean square error value obtained is greater than the prediction accuracy requirements of the engineering object, the constructed settlement prediction model is readjusted; if the mean square error value obtained is less than or equal to the prediction accuracy requirements of the engineering object, the prediction results and settlement prediction model are saved.

[0045] In the embodiment of the present invention, in S1, the expression of the relationship function f(ξ,h) between the settlement of the roadbed filler and the filling height is: f(ξ,h)=ue -h / t+ξ i ; In the formula, ξ i represents the first coefficient, u represents the second coefficient, t represents the third coefficient, h represents the filling height, and e represents the exponent.

[0046] Select a representative filling area at the project site for filling sampling, conduct settlement deformation tests on the roadbed filling at different compaction degrees ξ1, ξ2... and different filling heights h1, h2..., obtain the correlation curve between roadbed settlement S and filling height h at different compaction degrees, and obtain the roadbed filling settlement-filling height relationship equation by fitting. S is the roadbed deformation, in m; the second coefficient u and the third coefficient t are obtained by fitting; the first coefficient ξ i Take values ​​according to different compaction degrees.

[0047] In this embodiment of the present invention, S2 includes the following sub-steps:

[0048] S21, collecting roadbed filling height data and settlement time series data at the project site to be predicted;

[0049] S22, performing singular value test, missing value interpolation, data noise reduction and data standardization on the sedimentation time series data in sequence;

[0050] S23, dividing the roadbed filling height data and the settlement time series data after cleaning into a training set and a test set.

[0051] Extract roadbed filling height data H = {h1, h2, ..., h k The unit is m (meter). The roadbed settlement S = {s1, s2, ..., s k}, unit: m (meter).

[0052] Sedimentation time series data S = {s1, s2, ..., s k The singular value test specifically involves the following equation:

[0053] d i =2s i -(s i+1 +s i-1 ), i = 2, 3, ..., k-1;

[0054]

[0055] Among them, s i represents the ith settlement data, s i+1 represents the i+1th sedimentation data, s i-1 represents the i-1th sedimentation data, d i The sedimentation amount S = {s1, s2, ..., s k}The changing characteristics of represents the average value of the sedimentation sequence variation, represents the mean square error of the sedimentation sequence change, q i Indicates the change feature d i The absolute value of the deviation from the mean and the mean square error than.

[0056] Sedimentation time series data S = {s1, s2, ..., s k Missing value interpolation specifically involves the following equation:

[0057]

[0058] Among them, s i and i+1 They represent the monitoring values ​​before and after the interpolated monitoring value, s is the interpolated monitoring value, t and t i ,t i+1 is the corresponding monitoring time.

[0059] Sedimentation time series data S = {s1, s2, ..., s k} The noise reduction adopts the wavelet transform method, which specifically includes selecting the wavelet basis function, determining the decomposition level J required, decomposing the original signal into J layers through the wavelet decomposition algorithm, obtaining the high-frequency coefficients of each layer and the low-frequency coefficients of the last layer, and setting a threshold for the high-frequency coefficients during wavelet decomposition to reduce the possibility of misjudgment in noise elimination. Finally, after the high-frequency coefficients of each layer are processed by threshold quantization, they are reconstructed with the low-frequency coefficients of the last layer to obtain the denoised settlement monitoring sequence. Specifically, the following equations are involved:

[0060] Sedimentation time series data S = {s1, s2, ..., s k Normalization is to convert the indicators to the same scale to evaluate the model performance, which can simplify the model calculation and improve the operation efficiency. Data normalization methods include Min-Max Normalization, Z-Score normalization and Logarithmic normalization, etc. After the data is normalized, the data is used as the input of the physical information neural network. The specific formula of data normalization of the present invention involves:

[0061]

[0062] In the formula, represents the normalized sedimentation value, s i Indicates the i-th sedimentation value to be normalized, i = 1, 2, 3...; mean_s is the mean sedimentation of each sample, max_s is the maximum sedimentation value in the sample, and min_s is the minimum sedimentation value in the sample.

[0063] The sedimentation monitoring time series S' after data cleaning is S' = {s'1, s'2, ..., s' k} and roadbed settlement data H = {h1, h2, ..., h k} Pairwise division into training set S' x ={s' x1 、s' x2 ,……,s' xk}、H={h x1 、h x2 ,……,h xk} and test set S' c ={s' c1 、s' c2 ,……,s' ck}、H={h c1 、h c2 ,……,h ck}, the division ratio is 3:1.

[0064] In the embodiment of the present invention, in S22, the calculation formula for performing data noise reduction processing is:

[0065] s(t)=s d (t)+s k (t);

[0066] In the formula, s(t) represents the sedimentation time series data after data noise reduction, s d (t) represents the actual deformation signal, s k (t) represents deterministic noise, 2 -J (i-1)f~2 -J if, j = 1, 2, 3, ..., J is the scale parameter of the decomposition, 2 -J is the number of frequency bands, i=1,2,…,2 J is a frequency band sequence of the decomposed signal.

[0067] In the embodiment of the present invention, Figure 2 As shown, the settlement prediction model includes an input layer, a first 1D convolution layer, a first ReLU activation layer, a first causal convolution layer, a second ReLU activation layer, a second 1D convolution layer, a third ReLU activation layer, a second causal convolution layer, a fourth ReLU activation layer, a linear layer, and an output layer;

[0068] The input end of the input layer serves as the input end of the settlement prediction model; the output end of the input layer, the first 1D convolution layer and the input end of the first ReLU activation layer are connected in sequence; the first output end of the first ReLU activation layer, the first causal convolution layer and the input end of the second ReLU activation layer are connected in sequence; the second output end of the first ReLU activation layer and the first output end of the second ReLU activation layer are both connected to the input end of the second 1D convolution layer; the output end of the second 1D convolution layer is connected to the input end of the third ReLU activation layer; the second output end of the second ReLU activation layer and the first output end of the third ReLU activation layer are both connected to the input end of the second causal convolution layer; the output end of the second causal convolution layer is connected to the input end of the fourth ReLUc activation layer; the second output end of the third ReLU activation layer and the output end of the fourth ReLU activation layer are both connected to the input end of the linear layer; the output end of the linear layer is connected to the input end of the output layer; the output end of the output layer serves as the output end of the settlement prediction model.

[0069] Different network layers implement different functions, specifically: the input layer is used to input data; the 1D convolution layer is used to use 1D convolution with different convolution kernel sizes to capture different time scale features of time series data; the ReLU activation layer is used to increase the nonlinearity of the network and capture more complex data patterns; the residual connection is used to use the residual connection to add the output of the previous layer to the output of the current layer to alleviate the disappearance of the network gradient and enable the network to be trained at a deeper level; the causal convolution layer ensures the causal relationship of the data in the time dimension and only uses past data information when predicting future values. The above network structure is connected alternately through the above network layers to deeply explore the time series characteristics of the roadbed settlement data and retain the causality of the time series data.

[0070] In the embodiment of the present invention, the expression between the input layer and the output layer is:

[0071] Y = N(X; θ);

[0072] Where θ = [w, b], X represents the input variable of the settlement prediction model, Y represents the output variable of the settlement prediction model, θ represents the parameter to be identified of the settlement prediction model, N(·) represents the neural network operator of the parameter to be identified, w represents the weight between different network layers, and b represents the bias of the neural network.

[0073] In the embodiment of the present invention, for the construction of the network loss function, the present invention improves the traditional pure data-driven loss function to include the data-driven part loss_g u and the physical information driven part loss_g f. In the data-driven part, the constraints at the data information level are realized by the difference between the predicted value output by the network model and the actual observed value. In the physical information-driven part, the mapping function of the filling height and settlement value of the high fill roadbed obtained from the experiment is used as part of the loss function, and the physical constraint is realized by the difference between the predicted outputs of the physical model and the network model. In the loss function, adaptive weights are set to balance the influence of data information and physical information on the prediction results. After multiple iterations, the physical data dual-constraint loss function converges, and the model training is considered to be completed after the engineering requirements are met. The physical data dual-constraint loss function constructed by the present invention: According to the relationship function between the settlement of roadbed filler and the filling height, the physical data dual-constraint loss function of the time series prediction model is constructed, and it is embedded in the time series prediction model to realize the physical relationship constraints in the time series prediction model. The physical data dual-constraint loss function should include data-driven terms and physical equation residual terms. The expression of the settlement prediction model loss function loss_g is:

[0074] loss_g=λ1loss_g u +λ2loss_g f ;

[0075]

[0076] Where loss_g u Represents the data-driven term, loss_g f represents the residual term of the physical equation, λ1 represents the first weight coefficient, λ2 represents the second weight coefficient, N represents the total number of samples, s' i represents the observed data, represents the network prediction value, h i represents the filling height of the roadbed, and f(ξ,h) represents the relationship function between the settlement of the roadbed fill and the filling height.

[0077] In an embodiment of the present invention, the method for determining the first weight coefficient and the second weight coefficient in the loss function of the settlement prediction model is as follows: construct an adaptive optimization function, and input the current first weight coefficient and the current second weight coefficient into the adaptive optimization function. If the adaptive optimization function value is greater than 0.05, increase the current first weight coefficient or the current second weight coefficient until the adaptive optimization function value is less than or equal to 0.05, otherwise use the current first weight coefficient and the current second weight coefficient. loss_g is the constructed settlement prediction model loss function. The smaller the value is during training, the better the model training effect is; the first weight coefficient and the second weight coefficient are used to balance the importance of the two parts of the loss during training; the data-driven term defines the average difference between the network output and the actual observation data; the physical equation residual term defines the average difference between the physical constraint loss and the network output.

[0078] In the embodiment of the present invention, the expression of the adaptive optimization function A is:

[0079] And λ1+λ2=1;

[0080] Where loss_g u Represents the data-driven term, loss_g f represents the residual term of the physical equation, λ1 represents the first weight coefficient, λ2 represents the second weight coefficient, max(·) represents the maximum value function, and min(·) represents the minimum value function. If the calculated value of A is greater than 5%, it is considered that the data constraints and physical constraints are unbalanced and need to be adjusted. The coefficient corresponding to the larger value is increased and adjusted, and the overall loss also increases accordingly, that is, a larger penalty is imposed on the larger loss.

[0081] The following describes the invention in conjunction with specific embodiments.

[0082] In S1, a field survey is conducted, and representative sections are selected for filler sampling. Based on the filler, basic physical and mechanical tests are carried out on the roadbed filler to obtain the physical and mechanical parameters of the roadbed filler. Correlation curves between different filling heights h1, h2... of the roadbed and settlement deformations s1, s2... under different compaction degrees y1, y2... are drawn, and curve fitting is performed to obtain the relationship function S = f(y, h) between the settlement of the roadbed filler and the filling height.

[0083] The compaction degrees used in this embodiment are y1=0.87, y2=0.90 and y3=0.93. Through experiments, the settlement and deformation of the roadbed filler at different filling heights are shown in Table 1.

[0084] Table 1

[0085]

[0086]

[0087] According to the above test results, the roadbed filler-filling height relationship equation is fitted to obtain:

[0088] S(h,t)=ue -h / t +ξ i ;

[0089] Where S is the roadbed deformation, unit: m; u is the coefficient 1, the fitting value is 0.07925; t is the coefficient 2, the fitting value is -24.91303; ξ i is the coefficient 3, and has different values ​​according to different compaction degrees. In this embodiment, ξ1=-0.08498, ξ2=-0.08441, ξ3=-0.0877; h is the filling height, unit: m.

[0090] In step S2, the high fill roadbed construction data at the project site is collected to extract the roadbed filling height data H = {h1, h2, ..., h k The roadbed settlement S = {s1, s2, ..., s k According to the sensing equipment burial drawings and roadbed filling construction records, the filling height and settlement data are sorted out according to time information and location information. The roadbed filling data is used as auxiliary information for learning the law of roadbed settlement data.

[0091] In S3, the data S obtained from the construction site is k}Perform data preprocessing, including singular value test, missing value interpolation, data noise reduction and data standardization, so as to improve data quality and enhance the reliability of time series data.

[0092] First, a singular value test is performed to remove abnormal data points that are significantly different from other observations due to equipment failure, external environment, natural disasters, data entry errors, etc. After singular value removal, the settlement monitoring data at a certain moment is empty. When there are many singular values, the settlement time series data will be "broken" linearly, destroying the true characteristics of settlement deformation. The missing data is interpolated and improved by linear interpolation to maintain the continuity of the monitoring data.

[0093] Then the settlement information is denoised. Due to environmental interference, the roadbed settlement sensing equipment has complex original settlement time series data. The roadbed settlement deformation sequence is a weak signal with the characteristic of small deformation in a short time. Error interference manifests itself as a strong noise. The two are mixed in the original settlement time series data. Therefore, the wavelet denoising method is used to achieve denoising of the settlement time series data.

[0094] The specific implementation steps of wavelet denoising include: selecting a suitable wavelet basis function according to the actual problem, this embodiment selects dbN wavelet as the wavelet basis function, determining the tomography J to be decomposed, and this embodiment determines the decomposition layer number to be 1. The original signal is decomposed using the Mallat algorithm.

[0095] Determine the threshold determination method in threshold denoising. This embodiment adopts the soft threshold Sqtwolog method. The specific mathematical expression is as follows:

[0096]

[0097] In the formula, λ represents the threshold, N represents the sampling length, σ μ Represents variance.

[0098] Perform threshold quantization on each high-frequency coefficient:

[0099]

[0100] Where d represents the high frequency coefficient; η λ (d) represents the high-frequency coefficient threshold quantization value.

[0101] After the high-frequency coefficients of each layer are processed by threshold quantization, they are reconstructed with the low-frequency coefficients of the last layer to complete the wavelet denoising of the sedimentation time series data. Finally, the data is normalized in the range of [-1,1] by the Min-Max Scaling method to complete the data normalization process. Based on the above process, the data cleaning of the sedimentation time series data is completed. Please refer to the comparison chart of the data cleaning results Figure 3 .

[0102] The sedimentation time series data S' after data cleaning is obtained = {s'1, s'2, ..., s' k}, and then divide the data set according to the ratio of 3:1 between the training set and the test set to obtain the test set S' x ={s' x1 、s' x2 ,……,s' xk} and training set S' c ={s' c1 、s' c2 ,……,s' ck}.

[0103] In S4, the settlement prediction model structure can be found in Figure 2 , including 2 layers of 1D convolution layers and 2 layers of causal convolution layers. 1D convolution and causal convolution are designed alternately to extract the temporal features of the sedimentation data and preserve the temporal causality of the sedimentation data. Residual connections are used between convolution layers to alleviate the problem of gradient vanishing in the deep layers of the network and increase the scalability of the network in depth.

[0104] This embodiment uses the ReLU activation function as follows:

[0105] ReLU(x)=max(0,x).

[0106] In S4, the specific implementation of λ1 and λ2 adjustment is that after each epoch training is completed, the network model passes the predicted value into the loss function to calculate the data loss and physical loss, and calculates the evaluation index A for the data loss and physical loss. When the value of the evaluation index A is greater than 5%, it means that a certain item in the network has a greater influence in the learning process, and the influence of data information and physical information is unbalanced, then the coefficient of the item with greater influence is increased, that is, the penalty corresponding to the poor learning effect is increased, so as to achieve a balance between data influence and physical influence. For example, if the performance is better in data fitting, that is, the data constraint loss is smaller, and the performance is poor in physical constraints, then the weight coefficient λ2 of the physical constraint is increased to increase the overall loss and continue the network model learning and training, otherwise the weight coefficient λ1 is increased, and λ1+λ2=1 is always maintained during the adjustment of the two coefficients.

[0107] In step S5, the test set S' after data cleaning is c ={s' c1 、s' c2 ,……,s' ck The data is used as the input of the settlement prediction model to obtain the roadbed settlement prediction result S p ={s p1 、s p2 , ...}. Select the model evaluation index. This implementation uses the mean square error as the model prediction accuracy evaluation index. Calculate the mean square error value val_MSE of the test set:

[0108]

[0109] In the formula, val_MSE represents the mean square error value of the evaluation index of the model performance in the test set, n represents the number of predicted values, and s' i Indicates the actual value of settlement, s pi Represents the predicted settlement value.

[0110] The test set mean square error val_MSE calculated by the settlement prediction model is judged. The larger the test set mean square error val_MSE is, the larger the model error is. If the mean square error value obtained is greater than the prediction accuracy requirement of the engineering object, repeat steps S1 to S4; if the mean square error value obtained is less than or equal to the prediction accuracy requirement of the engineering object, save the prediction result S p ={s p1 、s p2 , ...} and prediction model. Please refer to the model prediction results Figure 4 .

[0111] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for predicting high fill roadbed settlement based on neural network, It is characterized in that The following steps are involved: S1. According to the subgrade settlement and filling height of the subgrade filling material at various compaction degrees, a relationship function between the subgrade settlement and filling height is constructed; S2, collecting the roadbed filling height data and settlement time series data of the project site to be predicted, and performing data cleaning on the settlement time series data, and dividing the roadbed filling height data and the cleaned settlement time series data into a training set and a test set; S3. Construct a settlement prediction model based on the relationship function between roadbed filler settlement and filling height; S4, using the training set as the input of the settlement prediction model to perform model training; S5. Use the test set to perform model testing on the trained settlement prediction model until the model test passes, and use the settlement prediction model that passes the test to determine the high fill roadbed settlement prediction result of the engineering site to be predicted.

2. The high fill roadbed settlement prediction method based on neural network according to claim 1, It is characterized in that In S1, the relationship function f(ξ,h) between the settlement of roadbed fill and the filling height is expressed as: f(ξ,h)=ue -h / t +ξ i ; In the formula, ξ i represents the first coefficient, u represents the second coefficient, t represents the third coefficient, h represents the filling height, and e represents the exponent.

3. The high fill roadbed settlement prediction method based on neural network according to claim 1, It is characterized in that The S2 comprises the following sub-steps: S21, collecting roadbed filling height data and settlement time series data at the project site to be predicted; S22, performing singular value test, missing value interpolation, data noise reduction and data standardization on the sedimentation time series data in sequence; S23, dividing the roadbed filling height data and the settlement time series data after cleaning into a training set and a test set.

4. The high fill roadbed settlement prediction method based on neural network according to claim 3, It is characterized in that In S22, the calculation formula for data noise reduction is: s(t)=s d (t)+s k (t); In the formula, s(t) represents the sedimentation time series data after data noise reduction, s d (t) represents the actual deformation signal, s k (t) represents deterministic noise.

5. The high fill roadbed settlement prediction method based on neural network according to claim 1, It is characterized in that The settlement prediction model includes an input layer, a first 1D convolution layer, a first ReLU activation layer, a first causal convolution layer, a second ReLU activation layer, a second 1D convolution layer, a third ReLU activation layer, a second causal convolution layer, a fourth ReLU activation layer, a linear layer and an output layer; The input end of the input layer is used as the input end of the settlement prediction model; the output end of the input layer, the first 1D convolutional layer and the input end of the first ReLU activation layer are connected in sequence; the first output end of the first ReLU activation layer, the first causal convolutional layer and the input end of the second ReLU activation layer are connected in sequence; the second output end of the first ReLU activation layer and the first output end of the second ReLU activation layer are both connected to the input end of the second 1D convolutional layer; The output end of the second 1D convolutional layer is connected to the input end of the third ReLU activation layer; the second output end of the second ReLU activation layer and the first output end of the third ReLU activation layer are both connected to the input end of the second causal convolutional layer; the output end of the second causal convolutional layer is connected to the input end of the fourth ReLUc activation layer; the second output end of the third ReLU activation layer and the output end of the fourth ReLU activation layer are both connected to the input end of the linear layer; the output end of the linear layer is connected to the input end of the output layer; the output end of the output layer serves as the output end of the settlement prediction model.

6. The high fill roadbed settlement prediction method based on neural network according to claim 5, It is characterized in that The expression between the input layer and the output layer is: Y = N(X; θ); Where θ = [w, b], X represents the input variable of the settlement prediction model, Y represents the output variable of the settlement prediction model, θ represents the parameter to be identified of the settlement prediction model, N(·) represents the neural network operator of the parameter to be identified, w represents the weight between different network layers, and b represents the bias of the neural network.

7. The high fill roadbed settlement prediction method based on neural network according to claim 5, It is characterized in that The expression of the loss function loss_g of the settlement prediction model is: loss_g=λ 1 loss_g u +λ 2 loss_g f ; Where loss_g u Represents the data-driven term, loss_g f represents the residual term of the physical equation, λ 1 represents the first weight coefficient, λ 2 represents the second weight coefficient, N represents the total number of samples, s' i represents the observed data, represents the network prediction value, h i represents the filling height of the roadbed, and f(ξ,h) represents the relationship function between the settlement of the roadbed fill and the filling height.

8. The high fill roadbed settlement prediction method based on neural network according to claim 7, It is characterized in that The method for determining the first weight coefficient and the second weight coefficient in the loss function of the settlement prediction model is: constructing an adaptive optimization function, and inputting the current first weight coefficient and the current second weight coefficient into the adaptive optimization function; if the adaptive optimization function value is greater than 0.05, then increasing the current first weight coefficient or the current second weight coefficient until the adaptive optimization function value is less than or equal to 0.05; otherwise, using the current first weight coefficient and the current second weight coefficient.

9. The high fill roadbed settlement prediction method based on neural network according to claim 8, It is characterized in that The expression of the adaptive optimization function A is: and λ 1 + λ 2 = 1; Where loss_g u Represents the data-driven term, loss_g f represents the residual term of the physical equation, λ 1 represents the first weight coefficient, λ 2 represents the second weight coefficient, max(·) represents the maximum value function, and min(·) represents the minimum value function.

Images