CN115419120A - Highway subgrade settlement monitoring and predicting method - Google Patents

Abstract

The invention discloses a highway subgrade settlement monitoring and predicting method. According to the invention, the inside and outside combined monitoring system of the highway subgrade is constructed through the intelligent smart geobelt and the Beidou satellite for subgrade deformation monitoring, and the lightweight neural network model is constructed to realize high-efficiency and high-precision prediction of highway subgrade settlement, so that the system is not only suitable for common highway subgrades, but also suitable for widening the highway subgrade, has wide application scenes, can realize full-section full-life distributed monitoring of subgrade deformation, and can realize accurate and high-efficiency prediction of highway subgrade settlement.

Description

A method for monitoring and predicting highway subgrade settlement

technical field

The invention relates to the technical field of highway engineering, in particular to a method for predicting highway subgrade settlement.

Background technique

The settlement and deformation of the subgrade shows a highly nonlinear characteristic in the change trend, especially in road reconstruction and expansion projects, the construction time of the newly widened subgrade and the original old subgrade is usually long, and factors such as filling materials, construction quality, and construction parameters All will have an impact on the settlement deformation of the subgrade. Therefore, monitoring and predicting the deformation of the subgrade is conducive to timely taking targeted measures to post-process the subgrade with excessive deformation, and can also realize effective safety warning of subgrade disasters.

Patent Application No. CN201110257149.4 discloses a roadbed settlement prediction method based on static penetrating sounding and BP neural network, which collects site data and settlement observation data and then predicts ordinary roadbed settlement. However, in the process of solving the ordinary BP neural network, with the expansion of the input data set scale, the increase of the number of layers, the complexity of the structure or connection form and other factors, it is easy to appear underfitting, slow convergence or overfitting, etc. All kinds of problems.

It involves the specific acquisition method of subgrade deformation observation data, including application numbers CN201611126688.3, CN201911367265.4 and other patents, all of which use inclinometer tubes to monitor lateral deformation, but the actual installation costs are high, the measurement steps are cumbersome, and the life of the monitoring equipment is short , cannot carry out permanent monitoring along with the service life of the subgrade, and cannot obtain the full-section horizontal displacement.

Therefore, how to develop a road subgrade settlement monitoring and prediction method based on neural network, with a full-section and full-life distributed deformation monitoring system, not only applicable to ordinary road subgrades, but also to achieve accurate and efficient prediction of widened subgrade subsidence settlement is a technology in this field Technical problems that personnel urgently need to solve.

Contents of the invention

Aiming at the above-mentioned deficiencies in the prior art, the present invention provides a highway subgrade settlement monitoring and prediction method, which has a wide range of application scenarios, can realize distributed monitoring of subgrade deformation for full section and full life, and simultaneously predicts subgrade settlement with high precision and high prediction efficiency.

To achieve the above object, the invention provides the following technical solutions:

The roadbed subsidence settlement monitoring and prediction method starts with the basic feedforward neural network, and establishes a method for subgrade by optimizing the topology structure of the basic feedforward neural network in form and adding penalty items to the learning parameters of the basic neural network in the algorithm. A lightweight neural network model for settlement prediction. In order to further accelerate the calculation of the neural network model and improve the over-fitting problem, the algorithm of the lightweight neural network abandons the current popular gradient descent and other algorithms that easily lead to local optimality, and adopts a new quasi-Newton method to complete the algorithm. The training of the neural network enables the lightweight neural network to improve its computational efficiency while ensuring the prediction accuracy, thereby avoiding the occurrence of over-fitting problems. After the design of the lightweight neural network model is completed, the subgrade displacement monitoring is carried out by using the joint internal and external monitoring method. The obtained measured data is preprocessed to the data set, and the training process of the lightweight neural network model is completed, and it is used for the monitoring and prediction of roadbed subsidence. A comprehensive evaluation of the prediction ability, computational efficiency and optimal structure of the lightweight neural network was carried out, and a prediction method for highway subgrade settlement based on the lightweight neural network was proposed.

Further, the basic feedforward neural network is a simple 3-layer feedforward neural network, including 1 input layer, 1 hidden layer and 1 output layer, wherein the hidden layer contains j neurons. For this basic feed _- forward neural network, the input quantity In enters the neural network from the input layer and then enters the hidden layer. In the middle hidden layer, there are j neurons, which are finally output through the output layer.

Further, the formal optimization of the topology of the basic neural network is to construct a complex neural network with more hidden layers based on the basic feedforward neural network, then the net input in the above learning process will be used as " The new input "is fed into the neurons of the next hidden layer with new weights and biases. Similarly, the neurons in this layer will also convert these "new inputs" again through the activation function, and then input the converted results to the next hidden layer, and so on, until they are passed to the output layer to complete the basic neural network. Model topology optimization.

Further, adding a penalty term to the learning parameters of the basic neural network algorithmically is to use the least square error function as a loss function to train network model parameters. By using the L ₁ norm penalty on the loss function to obtain a relatively sparse structure of its parameters in advance, the lightweight construction of the neural network is completed.

Further, the novel quasi-Newton method is a novel quasi-Newton method for constructing a search direction and a search step size based on subgradient and active set methods.

Further, the internal and external joint monitoring method is to set a reference point at a specified position on the external slope of the monitoring section, and use the Beidou monitoring system to conduct regular observations of the vertical displacement of the subgrade surface. At the same time, the soil deformation monitoring system and method based on conductive polymers are used to embed smart conductive polymers into the interior of the roadbed when widening the roadbed to observe the horizontal displacement of the full section.

Furthermore, the Beidou monitoring system uses three satellites of the second generation of Beidou for positioning, and uses the relative positioning method to monitor the vertical displacement of the surface of the subgrade section through the global satellite system. Internal horizontal displacement. The Beidou monitoring system is composed of Beidou receiving antenna, fixed power supply or solar panel, receiver, wireless transmission module, etc. Among them, the Beidou receiving antenna is connected to the reference point of the specified position through a connecting support, and the reference points are respectively set at the toe of the new subgrade slope of the test section, the shoulder of the road, and the top of the junction of the old and new subgrades. The receiver, wireless transmission module, etc. are placed in an equipment protection box. The three satellites of the second generation of Beidou are used for precise positioning, and the relative positioning method is used to obtain the relative position relationship between the point to be measured and the reference point through the baseline solution of the global satellite system.

Further, the soil deformation monitoring system based on conductive polymers uses smart conductive polymers as soil internal monitoring components. Connect cables at different measuring points on the strip-shaped smart conductive polymer, and the cables of many measuring points protrude from the end of the belt, and then prepare the smart smart geotechnical belt by packing the insulating protective sleeve with the belt. The intelligent smart geotechnical belt is buried horizontally inside the roadbed, and its end is exposed from the slope of the roadbed side slope. It is connected with a cable to the data collection station. The collection station is uploaded to the cloud server through the network, and the automation of monitoring data is realized with the help of the cloud server. Collection monitoring.

Further, the data set preprocessing is to use a normalization method to process the training data set, which can map the training data set to the range of [0,1], thereby enhancing the generalization ability of the lightweight neural network.

Further, the training process of the lightweight neural network model is to divide the training data set into two parts, a verification set and a training set, randomly select limited data as a verification set, and use the rest as a training set.

Compared with the prior art, the beneficial effects of the invention are:

1. The internal and external joint monitoring method can realize high-precision real-time intelligent monitoring of roadbed deformation data.

Specifically, in this patent, the short-baseline double-difference Beidou monitoring scheme is used to observe the deformation of the subgrade. For the collected data, the combination of the combined double-difference geometric distance and the combination of the ionospheric observation value minus the satellite-to-earth distance are used to observe value is processed. Weaken or eliminate the influence of station-to-satellite geometric distance, atmospheric delay and various noises on positioning, thereby reducing measurement errors and improving the positioning accuracy of the Beidou global satellite system.

Specifically, the smart conductive polymer used in this patent has good durability, high strength, and low manufacturing cost. The reinforcement itself has self-detection technology and does not need to embed sensors inside the material. Moreover, the detection is accurate and timely, and the performance is stable. By extracting the deformation and force characteristic information of the reinforcement, it can realize the location of the soil rupture surface in the entire section of the subgrade cross section, diagnose the damage state of the reinforced soil, and provide early warning for soil safety. Provide more reasonable technical means.

It can be seen that the integrated intelligent monitoring can be realized remotely by using the internal and external joint detection method, without the need for surveyors to visit the site to manually read the monitoring data, and the monitoring system can not only achieve high-precision real-time monitoring, but also can accompany the service life of the subgrade Enable persistent monitoring.

_2. Use the loss function and L1 norm penalty to realize the lightweight construction of the neural network.

Specifically, the least square error function is used as the loss function to train the network parameters, and a relatively sparse structure of its parameters is obtained in advance by using the L ₁ norm penalty on the loss function to complete the lightweight construction of the neural network. This method avoids various problems such as underfitting, slow convergence or overfitting due to the transformation of a series of influencing factors such as the expansion of the input data set scale, the increase in the number of layers, and the complexity of the structure or connection form.

3. A new quasi-Newton method based on subgradient and active set methods is used to drive the lightweight neural network.

来代替牛顿法中的海塞尔矩阵

避免涉及计算海塞尔矩阵的逆矩阵。该计算过程十分繁琐且无法保证迭代过程中海塞尔矩阵保持正定性，因此容易使神经网络的学习陷入局部最优，而无法获得准确的预测结果。在拟牛顿法基础上再采用主动集法对其中的超参数进行二次规划，彻底获得轻质神经网络的稀疏结构，从而有效降低轻质神经网络的过拟合风险，于此同时也能显著提升轻质神经网络的学习效率。Specifically, using an approximate matrix

to replace the Hessel matrix in Newton's method

Avoid involving computing the inverse of a Hessel matrix. The calculation process is very cumbersome and cannot guarantee the positive definiteness of the Hessel matrix during the iterative process, so it is easy to make the learning of the neural network fall into a local optimum, and it is impossible to obtain accurate prediction results. On the basis of the quasi-Newton method, the active set method is used to perform quadratic programming on the hyperparameters, and the sparse structure of the lightweight neural network is thoroughly obtained, thereby effectively reducing the risk of over-fitting of the lightweight neural network. At the same time, it can also significantly Improve the learning efficiency of lightweight neural networks.

4. Using lightweight neural network to realize high-precision and high-efficiency prediction of subgrade settlement.

Based on the first-hand data measured by the internal and external joint monitoring method, most of the settlement values obtained by using the lightweight neural network to predict are consistent with the corresponding observed settlement values, and the prediction accuracy of the lightweight neural network for subgrade settlement is better And the prediction ability is stable, and the lightweight neural network can greatly improve the training efficiency.

Description of drawings

Fig. 1 is a technical flow chart of a roadbed subsidence monitoring and prediction method;

Fig. 2 is the topological diagram of the basic feedforward neural network;

Fig. 3 is the contrast chart of measured value and predicted value of subgrade settlement of widening subgrade of horizontal filling of embodiment 1;

Fig. 4 is the comparison chart of computing efficiency between common BP neural network and lightweight neural network in embodiment 1;

Fig. 5 is the contrast chart of measured value and predicted value of settlement of subgrade filling widening subgrade of embodiment 2;

Fig. 6 is the comparison chart of computing efficiency between common BP neural network and lightweight neural network in embodiment 2;

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

A method for monitoring and predicting highway subgrade settlement involved in the present invention, as shown in Figure 1, includes the following processes:

Optimize the topology of the basic neural network S1, add penalty items to the learning parameters S2, use the new quasi-Newton method to train the neural network S3, use the internal and external joint monitoring method to monitor the subgrade displacement S4, preprocess the measured data set S5, and train the lightweight neural network S6.

S1 optimizes the topology of the basic neural network:

表示输入层中第n个量以权重w向隐藏层中第j个神经元输入；以符号

表示第i个隐藏层中第j个神经元的偏置项。然后，被连接的第j个神经元可将输入量通过一个激活函数进行转换(激活神经元)，激活函数为某个特定的非线性函数，被转换后的输入量称为净输入量，以符号

表示第i个隐藏层中第j个激活神经元的净输入量。最后，激活的神经元会将净输入量传递到输出层输出。In the basic neural network shown in Figure 2, each input quantity is connected to each neuron by the signal strength (weight) represented by the unknown coefficient w and the strength adjustment item (bias) represented by b, and the symbol

Indicates that the nth quantity in the input layer is input to the jth neuron in the hidden layer with weight w; the symbol

Denotes the bias term for the jth neuron in the ith hidden layer. Then, the connected jth neuron can convert the input through an activation function (activation neuron). The activation function is a specific nonlinear function, and the converted input is called the net input. symbol

Indicates the net input to the j-th activated neuron in the i-th hidden layer. Finally, activated neurons pass the net input to the output layer output.

Based on the basic feed-forward neural network shown in Figure 2, a complex neural network with more hidden layers is constructed, and the net input in the above learning process will be used as a "new input" with new weights and biases. neurons in the next hidden layer. Similarly, the neurons in this layer will also convert these "new inputs" again through the activation function, and then input the converted results to the next hidden layer, and so on until they are passed to the output layer for output. The described basic neural network topology optimization process can be expressed as follows:

是预先输入到神经网络中的第k个预置参数，由权重和偏置组成，通过预先对神经网络进行训练得到；a_k是激活函数，研究中采用双曲正切函数，即a_k＝(e^2x-1)/(e^2x+1)。Among them, I _ij is a matrix composed of input quantities x, and the subscript indicates that the i-th quantity is input to the j-th neuron of the first hidden layer; i is the number of input quantities; j is the number of neurons in each hidden layer ; e _i is the residual; w _k is the weight of the kth neuron of the first hidden layer; b _k is the bias of the kth neuron of the first hidden layer;

is the kth preset parameter input into the neural network in advance, which is composed of weight and bias, and is obtained by training the neural network in advance; a _k is the activation function, and the hyperbolic tangent function is used in the research, that is, a _k =( e ^2x -1)/(e ^2x +1).

The input is passed to the next hidden layer, and so on, until it is passed to the output layer for output. The described basic neural network topology optimization process can be expressed as follows:

S2 adds a penalty term to the learning parameters:

Use the least squares error function as the loss function to train the network parameters, and the loss function can be expressed as follows:

The neural network parameter training process is the process of minimizing the input value x _ij through the loss function. By using the L ₁ norm penalty on the loss function to obtain a relatively sparse structure of its parameters in advance, the lightweight construction of the neural network is completed. The process described can be expressed as follows:

w_k、b_k稀疏程度的拉格朗日乘数因子，β_j ^[k]为超参数，是训练过程中得到的一系列调节值。Among them, λ _{k, j} and λ _k are determined

The Lagrangian multiplier factor of the degree of sparsity of w _k and b _k , β _j ^[k] is a hyperparameter, which is a series of adjustment values obtained during the training process.

S3 uses a new quasi-Newton method to train the neural network:

则施加L₁范数惩罚后的轻质神经网络的学习过程可表述为：Ruoling

Then the learning process _of the lightweight neural network after applying the L1 norm penalty can be expressed as:

的最小化过程，选择使用

的次梯度方向作为最陡下降方向，

对于ψ_r各分量的次梯度可表述为：In this patent, for

The minimization process, choose to use

The subgradient direction of is taken as the steepest descending direction,

The subgradient of each component of ψ _r can be expressed as:

in:

对于各参数的次梯度后，对

采用所述的基于次梯度的新型拟牛顿算法进行处理。先利用泰勒展开式得到

在ψ[τ]处的逼近形式：in getting

After the subgradient for each parameter, for

The described novel quasi-Newton algorithm based on the subgradient is used for processing. First use the Taylor expansion to get

Approximation form at ψ[τ]:

为海塞尔矩阵。in,

is a Hessel matrix.

In minimizing this approximation, the subgradient-based search direction can be expressed as:

The entire iterative form can be expressed as:

ψ[τ]+ηh[τ]→ψ[τ+1]

Where: η is the forward value obtained by the line search.

来代替海塞尔矩阵

从而避免复杂计算，即迭代形式可表述为：Then use the approximation matrix

to replace the Hessel matrix

In order to avoid complex calculations, the iterative form can be expressed as:

的逆矩阵

按如下表述进行构造：

the inverse matrix of

Construct as follows:

ζ[τ]＝ψ[τ+1]-ψ[τ]；

为单位矩阵。in,

ζ[τ]=ψ[τ+1]-ψ[τ];

is the identity matrix.

S4 adopts internal and external joint monitoring method to monitor subgrade displacement:

The Beidou monitoring system and the soil deformation monitoring system based on conductive polymers are used to realize the real-time monitoring of subgrade displacement by the internal and external joint monitoring method.

The Beidou monitoring system is composed of Beidou receiving antenna, fixed power supply or solar panel, receiver, wireless transmission module, etc. The receiver, wireless transmission module, etc. are all placed in an equipment protection box, which can be fixed near the Beidou receiving antenna by fixing brackets or expansion bolts, so as to achieve the purpose of protecting the equipment.

The Beidou receiving antenna is connected to the reference point of the specified position through the connecting support. The reference point is set at the toe of the new subgrade slope, the shoulder of the road and the top of the junction of the old and new subgrade in the test section.

Real-time monitoring uses three satellites of the second generation of Beidou for precise positioning, and uses the relative positioning method to obtain the relative position relationship between the point to be measured and the reference point through the baseline solution of the global satellite system.

Among them, the relative positioning method is to use more than two receivers to measure at the same time, and to eliminate the common error of the signals of the two receivers through differential calculations, so as to obtain relatively high-precision relative coordinates between the two receivers.

The soil deformation monitoring system and method based on conductive polymer (CN201310312664.7) is used to embed a smart conductive polymer into the interior of the roadbed when widening the roadbed to observe the horizontal displacement of the whole section. The soil deformation monitoring system based on conductive polymers uses smart conductive polymers as soil internal monitoring components. Connect cables at different measuring points on the strip-shaped conductive polymer, and the cables of many measuring points protrude from the end of the tape, and then pack the tape with an insulating protective sleeve to achieve waterproof, wear-resistant and anti-static, and then prepare an intelligent smart geotextile.

Among them, the intelligent smart geotechnical belt is buried horizontally inside the roadbed, and its end is exposed from the slope of the roadbed side slope. It is connected with a cable to the data collection station, and the collection station is uploaded to the cloud server through the network.

Real-time monitoring is to realize the internal deformation detection of the soil during the whole life cycle through the distributed self-inspection technology of the reinforcement deformation, and to locate the potential cracks and slip surfaces in the soil for a long time, diagnose the internal deformation state of the soil, and then realize the monitoring data with the help of the cloud server The automatic collection and monitoring provides reasonable technical means for the realization of soil safety early warning.

S5 measured data set preprocessing:

S_i、

机敏土工带的应变值分别记为

则有：The settlement observations of the subgrade are expressed as

S _i ,

The strain values of the smart geotechnical belt are recorded as

Then there are:

表示时间量，其它标记含义如上文所示。则训练轻质神经网络所使用的数据集如下所示：Among them: the superscript "1" indicates the first monitoring section;

Indicates the amount of time, and the meanings of other symbols are as shown above. Then the data set used to train the lightweight neural network is as follows:

in:

S _i =(S ₁ ,S ₂ ,...,S ₁₈₀ ) ^T

In order to improve the generalization ability of the lightweight neural network, the training data set needs to be normalized, and the normalization process is to process each column vector in the training data set as follows:

Among them: Z represents the normalized data; X represents the elements in each column of vectors; X _min and X _max represent the minimum and maximum elements in each column of vectors, respectively.

Using the normalization method to process the training data set can map the training data set to the range of [0,1], thereby enhancing the generalization ability of the lightweight neural network.

S6 trains a lightweight neural network:

The training data set is divided into two parts, the verification set and the training set, and 10% of the data is randomly selected as the verification set, and the rest is used as the training set. At the same time, in order to quantitatively measure the prediction ability of lightweight neural network and ordinary feedforward neural network, in addition to visually comparing the predicted value with the actual observation value, the root mean square error value RMSE is also used to analyze the prediction results. RMSE is calculated as follows:

表示预测值；S_i表示相应的实际观测值。in:

Indicates the predicted value; S _i indicates the corresponding actual observed value.

For the neural network, when the predicted value is completely consistent with the observed value, that is, when the prediction accuracy rate is 100%, the corresponding RMSE value is 0. In other words, the stronger the predictive ability of the neural network, the higher the accuracy and the lower the RMSE value, and vice versa.

Example 1: Monitoring and Prediction of Subgrade Subgrade Widening by Horizontal Filling

Relying on a highway reconstruction and expansion project in Shandong Province, the site is basically alluvial silt in the Yellow River Plain, with poor engineering properties, high groundwater level, and weak layers. After 20 years of service and operation, the overall deformation of the old subgrade is basically stable. The height of the old subgrade in this road section is 6m, the width of the right half of the top surface is 14m, the widening width is 7m, and the slope ratio is 1:1.75. The traditional horizontal layered filling is adopted, and the new subgrade is first filled with a width of 5m and a height of 2m. A roadbed subsidence monitoring and prediction method described in this patent is used to monitor and predict the subsidence of the horizontal filling and widening roadbed.

Figure 3 shows the comparison between the measured value and the predicted value of the subgrade settlement for horizontal filling and widening. It can be seen that the difference between the predicted value of the lightweight neural network and the actual value is 0.00721mm at the minimum and 0.42719mm at the maximum. From the perspective of prediction accuracy, the minimum accuracy rate of lightweight neural network in the working process is 88.67%, the highest accuracy rate is 99.87%, and the average accuracy rate is 96.16%. Prediction of settlement can maintain a high accuracy.

In the same computing environment, the ordinary BP neural network is used as a comparison to test the computational efficiency of the lightweight neural network and the ordinary feedforward neural network. Let each kind of neural network start from near the same larger RMSE, and after preliminary training to near a smaller RMSE value, record the time required for each as shown in Figure 4. Unified calculation of the time taken by the two neural networks to drop to 35. For the BP neural network, it takes longer for the RMSE value to drop to a lower level, which is 18.78 times that of the lightweight neural network. It can be seen that the lightweight neural network can greatly improve the training efficiency.

Example 2: Monitoring and Prediction of Subgrade Subgrade Settlement Widening

Relying on a highway reconstruction and expansion project in Shandong Province, the site is basically alluvial silt in the Yellow River Plain, with poor engineering properties, high groundwater level, and weak layers. After 20 years of service and operation, the overall deformation of the old subgrade is basically stable. The height of the old subgrade in this road section is 6m, the width of the right half of the top surface is 14m, the widening width is 7m, and the slope ratio is 1:1.75. The 20m-long test section of the road section is selected to widen the subgrade by partial filling, and the new subgrade is first filled with a width of 5m and a height of 2m. A roadbed subsidence monitoring and prediction method described in this patent is used to monitor and predict the subsidence of subgrade filling and widening roadbed.

Figure 5 shows the comparison between the measured value and the predicted value of the subgrade settlement for subgrade filling and widening. It can be seen that the difference between the predicted value of the lightweight neural network and the actual value is 0.00012mm at the minimum and 0.38246mm at the maximum. From the perspective of prediction accuracy, the minimum accuracy rate of the lightweight neural network in the working process is 89.00%, the highest accuracy rate is 99.99%, and the average accuracy rate is 96.51%, which shows that the lightweight neural network has a widening effect on partial filling. The subgrade settlement can be predicted with a high precision.

In the same computing environment, the ordinary BP neural network is used as a comparison to test the computational efficiency of the lightweight neural network and the ordinary feedforward neural network. Let each kind of neural network start from near the same larger RMSE, and after initial training to near a smaller RMSE value, record the time required for each as shown in Figure 6. Unified calculation of the time taken by the two neural networks to drop to 35. For the BP neural network, it takes longer for the RMSE value to drop to a lower level, which is 20.16 times that of the lightweight neural network. It can be seen that the lightweight neural network can greatly improve the training efficiency.

Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these modifications and variations.

Claims (10)

1. A highway subgrade settlement monitoring and predicting method is characterized in that a light neural network model for subgrade settlement prediction is established by optimizing a topological structure of a basic feedforward neural network in form and adding punishment items to learning parameters of the basic neural network in algorithm, training of the neural network is completed by adopting a novel quasi-Newton method, subgrade displacement monitoring is performed by adopting an internal and external combined monitoring method after the light neural network is designed, and data set preprocessing is performed on obtained actual measurement data to complete a training process of the light neural network and use the training process for highway subgrade settlement monitoring and prediction.

2. The method for monitoring and predicting the subgrade settlement of the highway according to claim 1, wherein the basic feedforward neural network is a 3-layer feedforward neural network comprising 1 input layer, 1 hidden layer and 1 output layer, wherein the hidden layer comprises j neurons; for this basic feedforward neural network, input quantity I _n The data can enter a neural network from an input layer and then enter a hidden layer; in the middle hidden layer, there are j number of neurons, which are finally output through the output layer.

3. The method for monitoring and predicting the subgrade settlement of the highway according to claim 1, wherein the topological structure for formally optimizing the basic neural network is based on the basic feedforward neural network to construct a complex neural network with more hidden layers.

4. The method for monitoring and predicting the subgrade settlement of the highway according to claim 1, wherein the penalty term is algorithmically added to the learning parameters of the basic neural network by training network model parameters by using a least square error function as a loss function; by using L for the loss function ₁ And (4) punishing the norm to obtain a relatively sparse structure of the parameters in advance, so as to complete the lightweight construction of the neural network.

5. The method for monitoring and predicting the settlement of the roadbed of claim 1, wherein the novel quasi-Newton method is a novel quasi-Newton method for constructing the search direction and the search step size on the basis of a sub-gradient method and an active set method.

6. The method for monitoring and predicting the settlement of the road subgrade as claimed in claim 1, wherein the inside and outside combined monitoring method is to set reference points at specified positions on the side slope outside the monitored section, and use a Beidou monitoring system to perform periodic observation of the vertical displacement of the surface of the subgrade, and simultaneously use a soil deformation monitoring system and method based on the conductive polymer to embed the smart conductive polymer into the subgrade during widening the subgrade construction to perform horizontal displacement observation of the whole section.

7. The inside and outside combined monitoring method according to claim 6, wherein the big dipper monitoring system is positioned by three second generation big dipper satellites, the vertical displacement of the roadbed section surface is monitored by a global satellite system by adopting a relative positioning method, and the horizontal displacement in the roadbed is monitored by the soil deformation monitoring system based on the conductive polymer through an intelligent smart geotechnical belt.

8. The inside-outside combination monitoring method according to claim 6, wherein the soil deformation monitoring system based on conductive polymers adopts smart conductive polymers as soil interior monitoring elements; connecting cables at different measuring points on the banded smart conductive polymer, extending the cables at a plurality of measuring points out of the tail end of the band, and then packaging the band with an insulating protective sleeve to prepare an intelligent smart geoband; inside intelligent smart geotechnical zone level buried the road bed, its terminal domatic department from the road bed side slope exposes, is connected with the cable and is connected with the data acquisition station, and the high in the clouds server is passed to through the network in the acquisition station, realizes the automatic collection monitoring of monitoring data with the help of the high in the clouds server.

9. The method for monitoring and predicting the subgrade settlement of the highway according to claim 1, wherein the preprocessing of the data set is to process a training data set by a normalization method, and map the training data set into the range of [0,1] so as to enhance the generalization capability of the lightweight neural network.

10. The method for monitoring and predicting the settlement of the road subgrade as claimed in claim 1, wherein the training process of the light neural network is to divide a training data set into a validation set and a training set, randomly extract limited data as the validation set, and use the rest as the training set.

Images