1. A method for predicting surrounding rock deformation during construction of a road tunnel based on a Bayes-LSTM model is characterized by comprising the following steps:

the method comprises the following steps:

a. and (3) acquiring vault settlement and peripheral convergence data: the data is continuous settlement data actually measured by an observation point in a period of time;

b. and (3) mechanism knowledge analysis: drawing a curve graph of vault settlement and peripheral convergence change, carrying out preliminary analysis and judgment on a curve change rule and a tunnel surrounding rock deformation rule, and eliminating the condition that actually measured data in a construction period are obviously abnormal due to special unfavorable geological reasons;

c. preprocessing data: normalizing the data by using a Python built-in function MinMaxScale, and converting the time sequence data into supervised data by using a series _ to _ super function;

d. dividing the data set: using a train _ test _ split function to take the first 67% of data as a training set and the last 33% of data as a test sample, and taking 20% of the training set as a verification set to verify the generalization capability of the model;

e. constructing an LSTM vault settlement and periphery convergence prediction model by using a Keras frame with a built-in Python;

f. building a parameter optimization model by using a Hyperopt Bayes parameter adjusting module built in Python, and setting Bayes optimization parameters and a search space;

g. optimizing the LSTM model: setting initial parameters, training a vault settlement and peripheral convergence prediction model of the LSTM, repeatedly training the LSTM by using an MSE (mean square error) as a loss function and adopting a parameter optimization model in f, judging the fitting effect of the model according to the loss of a training set verification set, and selecting a group of super-parameter combinations with minimum loss to construct an LSTM model;

h. loading a trained LSTM vault settlement and periphery convergence prediction model, predicting test set data, and outputting a predicted value y;

i. the resulting prediction data in h is denormalised using the inverse _ transform function built into Python.

2. The Bayes-LSTM model-based surrounding rock deformation prediction method for the construction period of the road tunnel according to claim 1, wherein:

in the step f, Bayes optimization parameters comprise the size of a hidden layer in an LSTM unit, the selection of an optimizer, the learning rate and the iteration times; the search space of the size of the hidden layer in the LSTM unit is 2-64, the optimizers are selected from adam ', rmspoop and adamax', the search space of the learning rate is 0.001-0.01, and the search space of the iteration times is 100-300.

3. The Bayes-LSTM model-based surrounding rock deformation prediction method for the construction period of the road tunnel according to claim 2, wherein:

in step g, training the prediction model of vault settlement and peripheral convergence of the LSTM comprises:

construction of forgetting Gate (forget gate) f_t：f_t＝σ(W_fh_t-1+W_fx_t+b_f)；

Building input gate (input gate) i_t：i_t＝σ(W_ih_t-1+W_ix_t+b_i)；

New cell information C at the present time_t：C_t＝f_t×C_t-1+i_t×tanh(W_ch_t-1+W_cx_t+b_c)；

Compute output gate (output gate) o_t：o_t＝σ(W_oh_t-1+W_ox_t+b_o)；

Calculate the final output h_t：h_t＝o_t×tanh(C_t)；

Wherein W and b represent weight matrix and bias parameter, respectively, σ is sigmoid function, i_tDetermining the information required for renewal into the cellular state, C_tFor new unit information at time t, o_tOutput part for determining cell state, h_tDenotes the time x at t_tThe output of the corresponding cell.

4. The Bayes-LSTM model-based surrounding rock deformation prediction method for the construction period of the road tunnel according to claim 3, wherein:

in step g, the mean square error MSE is selected as a loss function,wherein, y_iIn order to be the actual value of the measurement,is a predicted value.

5. The Bayes-LSTM model-based surrounding rock deformation prediction method for the construction period of the road tunnel according to claim 4, wherein:

step g, when the output of the gate is 0, forbidding all information to pass through; when the gate output is 1, this indicates that all information is allowed to pass.

6. The Bayes-LSTM model-based surrounding rock deformation prediction method for the construction period of the road tunnel according to claim 5, wherein:

and g, optimizing the LSTM surrounding rock deformation prediction model by a Bayes parameter optimization method to find out the optimal parameters, wherein the LSTM surrounding rock deformation prediction model is a double-layer LSTM model.

Background

In recent years, with the development of traffic construction, the number of highway tunnels is also rapidly increased, the balanced state of the stress strain of rock-soil bodies can be damaged in the tunnel excavation process, the problems of stress redistribution, surrounding rock deformation and the like can also follow, the large deformation of the surrounding rock is one of common construction disasters in tunnel construction, the large deformation of the surrounding rock can damage a supporting structure and invade a section boundary, and the large deformation has great influence on the normal construction of the tunnel. However, tunnel surrounding rock has the characteristics of discontinuity, heterogeneity and the like, so that theoretical calculation of surrounding rock deformation is difficult, vault settlement and surrounding convergence are accurately predicted in construction, and the method has very important significance for analysis of surrounding rock deformation and surrounding rock stability. The method is characterized in that a large number of research results of students at home and abroad on the prediction of the tunnel surrounding rock deformation are researched and researched, a common prediction method is statistical regression and numerical simulation, a new solution is provided for solving the problem of the tunnel surrounding rock deformation along with the development of an artificial intelligence technology, some students propose a model which uses a support vector machine as a theoretical basis, optimizes model parameters by using a particle swarm algorithm and a chaos theory, constructs a chaos optimization PSO-SVM model to predict the tunnel surrounding rock deformation, and the students adopt gray theory analysis, time sequence analysis and a BP neural network.

Although students at home and abroad make more researches on the tunnel surrounding rock deformation prediction method, the prediction models have certain limitations, for example, the BP neural network is greatly influenced by an initial value, and problems such as network gradient disappearance and gradient explosion easily occur.

Disclosure of Invention

The invention aims to provide a Bayes-LSTM model-based surrounding rock deformation prediction method during construction of a highway tunnel, which can improve the accuracy of surrounding rock deformation prediction and reduce the training time of samples, thereby ensuring the construction safety and the construction progress of the highway tunnel.

The technical scheme adopted by the invention is as follows:

a method for predicting surrounding rock deformation during construction of a road tunnel based on a Bayes-LSTM model is characterized by comprising the following steps:

the method comprises the following steps:

a. and (3) acquiring vault settlement and peripheral convergence data: the data is continuous settlement data actually measured by an observation point in a period of time;

b. and (3) mechanism knowledge analysis: drawing a curve graph of vault settlement and peripheral convergence change, carrying out preliminary analysis and judgment on a curve change rule and a tunnel surrounding rock deformation rule, and eliminating the condition that actually measured data in a construction period are obviously abnormal due to special unfavorable geological reasons;

c. preprocessing data: normalizing the data by using a Python built-in function MinMaxScale, and converting the time sequence data into supervised data by using a series _ to _ super function;

d. dividing the data set: using a train _ test _ split function to take the first 67% of data as a training set and the last 33% of data as a test sample, and taking 20% of the training set as a verification set to verify the generalization capability of the model;

e. constructing an LSTM vault settlement and periphery convergence prediction model by using a Keras frame with a built-in Python;

f. building a parameter optimization model by using a Hyperopt Bayes parameter adjusting module built in Python, and setting Bayes optimization parameters and a search space;

g. optimizing the LSTM model: setting initial parameters, training a vault settlement and peripheral convergence prediction model of the LSTM, repeatedly training the LSTM by using an MSE (mean square error) as a loss function and adopting a parameter optimization model in f, judging the fitting effect of the model according to the loss of a training set verification set, and selecting a group of super-parameter combinations with minimum loss to construct an LSTM model;

h. loading a trained LSTM vault settlement and periphery convergence prediction model, predicting test set data, and outputting a predicted value y;

i. the resulting prediction data in h is denormalised using the inverse _ transform function built into Python.

In the step f, Bayes optimization parameters comprise the size of a hidden layer in an LSTM unit, the selection of an optimizer, the learning rate and the iteration times; the search space of the size of the hidden layer in the LSTM unit is 2-64, the optimizers are selected from adam ', rmspoop and adamax', the search space of the learning rate is 0.001-0.01, and the search space of the iteration times is 100-300.

In step g, training the prediction model of vault settlement and peripheral convergence of the LSTM comprises:

construction of forgetting Gate (forget gate) f_t：f_t＝σ(W_fh_t-1+W_fx_t+b_f)；

Building input gate (input gate) i_t：i_t＝σ(W_ih_t-1+W_ix_t+b_i)；

New cell information C at the present time_t：C_t＝f_t×C_t-1+i_t×tanh(W_ch_t-1+W_cx_t+b_c)；

Compute output gate (output gate) o_t：o_t＝σ(W_oh_t-1+W_ox_t+b_o)；

Calculate the final output h_t：h_t＝o_t×tanh(C_t)；

Wherein W and b represent weight matrix and bias parameter, respectively, σ is sigmoid function, i_tDetermining the information required for renewal into the cellular state, C_tFor new unit information at time t, o_tOutput part for determining cell state, h_tDenotes the time x at t_tThe output of the corresponding cell.

In step g, the mean square error MSE is selected as a loss function,wherein, y_iIn order to be the actual value of the measurement,is a predicted value.

Step g, when the output of the gate is 0, forbidding all information to pass through; when the gate output is 1, this indicates that all information is allowed to pass.

And g, optimizing the LSTM surrounding rock deformation prediction model by a Bayes parameter optimization method to find out the optimal parameters, wherein the LSTM surrounding rock deformation prediction model is a double-layer LSTM model.

The invention has the following advantages:

1. the invention utilizes Bayes to optimize an LSTM model, and Bayes optimization parameters comprise: size of hidden layer in LSTM unit (units), selection of optimizer (optimizer), learning rate (learn _ rate), iteration number (epochs); the search space of the size of the hidden layer in the LSTM unit is 2-64, the optimizer selects adam ', ' rmspoop ', ' adamax ', the search space of the learning rate is 0.001-0.01, and the search space of the iteration times is 100-300. And the optimal parameters are trained according to different surrounding rock change curves, so that the accuracy of the surrounding rock deformation prediction is improved.

2. The LSTM network based on the invention is derived from RNN, and the LSTM network is an improved Recurrent Neural Network (RNN), which can solve the problem that RNN cannot handle long-distance dependence and can avoid the problems of gradient explosion and gradient disappearance in RNN. All RNNs have a chain form of repeating neural network modules. In a standard RNN, this duplicated module has only a very simple structure, such as a tanh layer. LSTM also has this chain structure, but its repeating unit has only one network layer unlike the unit in a standard RNN network, and it has four network layers inside it. Meanwhile, the LSTM has three "gates," which are a "forgetting gate", an "input gate", and an "output gate", respectively. LSTM relies on the structure of "gates" that allow information to selectively affect the state of the recurrent neural network at each time, allowing selective decisions as to which information to pass. Therefore, the LSTM reduces the time of model training and further improves the accuracy of model prediction, thereby ensuring the construction safety and the construction progress of the road tunnel.

Drawings

FIG. 1 is a flow chart of a method of the present invention;

FIG. 2 is a graph illustrating the variation of dome settlement for an example data set of the present invention;

FIG. 3 is a graph illustrating exemplary data peripheral convergence variation in accordance with the present invention.

Detailed Description

The present invention will be described in detail with reference to specific embodiments.

The invention relates to a Bayes-LSTM model-based surrounding rock deformation prediction method in a construction period of a road tunnel, which comprises the following steps:

a. and (3) acquiring vault settlement and peripheral convergence data: the data is continuous settlement data actually measured by an observation point in a period of time;

b. and (3) mechanism knowledge analysis: drawing a curve graph of vault settlement and peripheral convergence change, carrying out preliminary analysis and judgment on a curve change rule and a tunnel surrounding rock deformation rule, and eliminating the condition that actually measured data in a construction period are obviously abnormal due to special unfavorable geological reasons;

c. preprocessing data: normalizing the data by using a Python built-in function MinMaxScale, and converting the time sequence data into supervised data by using a series _ to _ super function;

d. dividing the data set: using a train _ test _ split function to take the first 67% of data as a training set and the last 33% of data as a test sample, and taking 20% of the training set as a verification set to verify the generalization capability of the model;

e. constructing an LSTM vault settlement and periphery convergence prediction model by using a Keras frame with a built-in Python;

f. building a parameter optimization model by using a Hyperopt Bayes parameter adjusting module built in Python, and setting Bayes optimization parameters and a search space;

g. optimizing the LSTM model: setting initial parameters, training a vault settlement and peripheral convergence prediction model of the LSTM, repeatedly training the LSTM by using an MSE (mean square error) as a loss function and adopting a parameter optimization model in f, judging the fitting effect of the model according to the loss of a training set verification set, and selecting a group of super-parameter combinations with minimum loss to construct an LSTM model;

h. loading a trained LSTM vault settlement and periphery convergence prediction model, predicting test set data, and outputting a predicted value y;

i. the resulting prediction data in h is denormalised using the inverse _ transform function built into Python.

In the step f, Bayes optimization parameters comprise the size of a hidden layer in an LSTM unit, the selection of an optimizer, the learning rate and the iteration times; the search space of the size of the hidden layer in the LSTM unit is 2-64, the optimizers are selected from adam ', rmspoop and adamax', the search space of the learning rate is 0.001-0.01, and the search space of the iteration times is 100-300.

In step g, training the prediction model of vault settlement and peripheral convergence of the LSTM comprises:

construction of forgetting Gate (forget gate) f_t：f_t＝σ(W_fh_t-1+W_fx_t+b_f)；

Building input gate (input gate) i_t：i_t＝σ(W_ih_t-1+W_ix_t+b_i)；

New cell information C at the present time_t：C_t＝f_t×C_t-1+i_t×tanh(W_ch_t-1+W_cx_t+b_c)；

Compute output gate (output gate) o_t：o_t＝σ(W_oh_t-1+W_ox_t+b_o)；

Calculate the final output h_t：h_t＝o_t×tanh(C_t)；

Wherein W and b represent weight matrix and bias parameter, respectively, σ is sigmoid function, i_tDetermining the information required for renewal into the cellular state, C_tFor new unit information at time t, o_tOutput part for determining cell state, h_tDenotes the time x at t_tThe output of the corresponding cell.

In step g, the mean square error MSE is selected as a loss function,wherein, y_iIn order to be the actual value of the measurement,is a predicted value.

Step g, when the output of the gate is 0, forbidding all information to pass through; when the gate output is 1, this indicates that all information is allowed to pass.

And g, optimizing the LSTM surrounding rock deformation prediction model by a Bayes parameter optimization method to find out the optimal parameters, wherein the LSTM surrounding rock deformation prediction model is a double-layer LSTM model.

Example (b):

as shown in fig. 1, the method comprises the following specific steps:

a. acquiring vault settlement and peripheral convergence data in the construction period of the highway tunnel: the data is continuous settlement data actually measured by an observation point in a period of time;

table 1 example of measured values of dome settlement

TABLE 2 examples of peripheral convergence measurements

b. And (3) mechanism knowledge analysis: drawing a curve graph of vault settlement and peripheral convergence change, carrying out preliminary analysis and judgment on a curve change rule and a tunnel surrounding rock deformation rule, and eliminating the condition that actually measured data in a construction period are obviously abnormal due to reasons such as special unfavorable geology and the like;

c. preprocessing data: normalizing the data by using a Python built-in function MinMaxScale, converting time series data into supervised data by using a series _ to _ super function, predicting data on the third day by using the accumulated settlement data of the first two days and the monitoring days as characteristic factors because the monitoring days are discontinuous, taking C1 as an example, and showing in Table 3;

TABLE 3C 1 example of supervised data

d. Dividing the data set: using a train _ test _ split function to take the first 67% of data as a training set and the last 33% of data as a test sample, and taking 20% of the training set as a verification set to verify the generalization capability of the model;

e. building an LSTM vault settlement and periphery convergence prediction model by using a Keras frame with a built-in Python;

f. building a parameter optimization model by using a Hyperopt Bayes parameter adjusting module built in Python, and setting Bayes optimization parameters and a search space;

g. optimizing the LSTM model: setting initial parameters, training a vault settlement and peripheral convergence prediction model of the LSTM, repeatedly training the LSTM by using an MSE (mean square error) as a loss function and adopting a parameter optimization model in f, judging the fitting effect of the model according to the loss of a training set verification set, and selecting a group of super-parameter combinations with minimum loss;

TABLE 4 parameter space for vault subsidence Bayes optimization and an example of the optimal parameters thereof

TABLE 5 examples of parameters for optimal peripheral convergence

h. Loading a trained LSTM vault settlement and periphery convergence prediction model, predicting test set data, and outputting a predicted value y;

i. the resulting prediction data in h is denormalised using the inverse _ transform function built into Python.

The invention is not limited to the examples, and any equivalent changes to the technical solution of the invention by a person skilled in the art after reading the description of the invention are covered by the claims of the invention.