CN115081749A

CN115081749A - Bayesian optimization LSTM-based shield tunneling load advanced prediction method and system

Info

Publication number: CN115081749A
Application number: CN202210900027.0A
Authority: CN
Inventors: 吴贤国; 冯宗宝; 李昕懿; 刘俊; 陈虹宇; 王廷辉; 裴以军; 张金军; 戴小松; 徐文胜; 覃亚伟; 吴克宝
Original assignee: Huazhong University of Science and Technology
Current assignee: Huazhong University of Science and Technology
Priority date: 2022-07-28
Filing date: 2022-07-28
Publication date: 2022-09-20

Abstract

The invention discloses a shield tunneling load advanced prediction method and a shield tunneling load advanced prediction system based on Bayesian optimization LSTM, wherein the method comprises the following steps: acquiring monitoring data of shield construction parameters, and performing data preprocessing; performing redundant parameter analysis and filtering by using Pearson correlation analysis, and performing feature selection by using a random forest algorithm to obtain an optimal parameter set; and predicting the shield tunneling load of the LSTM by Bayesian optimization based on the optimal parameter set. The method comprises the steps of preprocessing acquired initial data, filtering irrelevant data such as shutdown data, abnormal values and missing values, filtering high coupling parameters through correlation analysis, performing importance sorting and feature selection by using RF on the basis to obtain an optimal parameter set, determining optimal hyper-parameters of an LSTM prediction model by using Bayesian optimization, constructing a cutter torque and total thrust prediction model based on the selected hyper-parameters, realizing accurate advance prediction of tunneling load, and providing reference for the operation of the shield tunneling machine.

Description

Bayesian optimization LSTM-based shield tunneling load advanced prediction method and system

Technical Field

The invention belongs to the technical field of shield tunnel construction load prediction, and particularly relates to a Bayesian optimization LSTM-based shield tunneling load advanced prediction method and system.

Background

In the tunneling construction process of the tunnel, the tunneling load can change along with the changes of geological environment, the tunneling speed of the shield tunneling machine and the operation state, the interaction between rock machines is reflected to a certain degree, and the method has important significance for avoiding the blocking of a cutter head, reducing the abrasion of the cutter head, ensuring the tunneling efficiency, adjusting the operation parameters of the shield tunneling machine and the like. In addition, improper control of the tunneling load can cause damage to the shield tunneling machine and damage to the upper soil body of the tunnel, which causes serious economic loss and safety accidents. Therefore, the method can predict the tunneling load in advance, provide reference and guidance for adjusting the operation parameters of the shield tunneling machine, and is an important work for ensuring the tunneling safety and the tunneling efficiency of the shield tunnel.

However, in practical engineering, due to unpredictability and complex variability of geological environment and timing variability of load during tunneling, achieving accurate tunneling load advance prediction is still a difficult and challenging task.

In order to solve the problems, the patent CN 110895730 a discloses a TBM tunneling parameter prediction method based on an LSTM algorithm, which predicts the rotation speed and the propulsion speed of a cutterhead in a stable stage according to an ascending stage, provides a follow-up operation parameter suggestion for a TBM driver, and improves the construction informatization and intelligentization level of the shield industry. However, the method mainly aims at TBM construction, the processing of the acquired data is not accurate enough, and the evaluation test of the predicted model is not convincing, so that the prediction result is not reasonable as a reference basis.

Disclosure of Invention

Aiming at the defects or improvement requirements of the prior art, the invention provides a Bayesian optimization LSTM-based shield tunneling load advanced prediction method and a Bayesian optimization LSTM-based shield tunneling load advanced prediction system, which preprocess acquired initial data, filter independent data such as shutdown data, abnormal values and missing values, filter highly-coupled parameters through correlation analysis, perform importance ranking and feature selection by using RF on the basis to obtain an optimal parameter set, determine optimal super-tunneling parameters of an LSTM prediction model by Bayesian optimization, construct a cutter torque and total propulsion prediction model based on the selected super-parameters, realize accurate advanced prediction of load, and provide reference for the operation of a shield tunneling machine.

In order to achieve the above object, according to an aspect of the present invention, there is provided a shield tunneling load advanced prediction method based on bayesian optimization LSTM, including the following steps:

s100, acquiring monitoring data of shield construction parameters, and performing data preprocessing;

s200, performing redundant parameter analysis and filtering by using Pearson correlation analysis, and performing feature selection by using a random forest algorithm to obtain an optimal parameter set;

s300, based on the optimal parameter set, predicting the shield tunneling load of the Bayesian optimization LSTM.

Further, the S300 includes:

s310, data normalization processing is carried out, and the specific formula is as follows:

wherein the content of the first and second substances,

which represents the standard value after the normalization,

and

defaults to 1 and 0, respectively, representing normalized maximum and minimum values,

representing the value of a sample to be measured,

and

maximum and minimum values of the sample values, respectively;

s320, LSTM hyperparametric selection based on Bayes: adjusting parameters by a Bayesian optimization method to obtain optimal values of the hyper-parameters;

and S330, evaluating and comparing the model performance.

Further, the step S320 specifically includes the following steps:

s321, defining a model training error MSE to be minimized into a target function, randomly generating an initialization parameter combination in an optimization range, and training an LSTM prediction model according to the initialization parameter combination to obtain a model output result MSE value;

s322, introducing the initialization parameter combination into a Gaussian process, taking a fitted initial Gaussian model as a proxy model of a target function, and correcting the initial proxy model by using the LSTM prediction model training error MSE value obtained in the previous step to enable the distribution of the initial proxy model training error MSE to be closer to the real distribution of the model training error MSE;

s323, selecting a target function minimum value as a next parameter combination to be evaluated through an expected improved acquisition function, and repeating the LSTM prediction model training and Gaussian model correction of the previous two steps;

and S324, when the maximum iteration times are reached, stopping the optimization process, wherein the corresponding super-parameter combination is the optimal super-parameter when the MSE value of the LSTM prediction model is minimum.

Further, the S330 specifically includes:

s331, evaluating the performance of the prediction model by using three common indexes of a decision coefficient, a root mean square error and an average absolute error;

s332, the Bayesian optimization LSTM prediction model is compared with the prediction performance of other common models.

Further, the S200 includes the steps of:

s210 Pearson parameter correlation analysis: measuring linear correlation between the parameters by using a Pearson correlation coefficient;

s220 feature selection based on RF algorithm: based on the RF algorithm, the influence of the change of the characteristic variable on the Gini index is evaluated, so that the importance of the variable is measured, and the further screening of the parameters is carried out by combining a characteristic selection method.

Further, the S220 specifically includes:

s221, establishing a regression decision tree: predicting OOB by using a random forest model to obtain the mean square residual errors of b pieces of data outside the bags, wherein the mean square residual errors are

The calculation formula is as follows:

wherein the content of the first and second substances,

in order to be able to measure the amount of data,

representing the true value of the dependent variable in the out-of-bag data;

expressing the predicted value of the regression model, and randomly changing the data outside the bag

Characteristic parameter

And calculating a new out-of-bag error accuracy

A value;

s222 generates an error matrix: when generating the regression decision tree, randomly selecting the characteristic parameters to split the decision tree and dividing the parameters

In that

Randomly replacing the data samples outside the bags to form a new OOB test set, predicting the new test set by using the established random forest regression model again to obtain a new OOB residual mean square

Generating an error matrix

Comprises the following steps:

wherein

The number of influencing factor variables;

the number of training sample sets;

s223 performing importance scoring: by using

Subtracting the corresponding row of the error matrix, averaging the subtracted result, and dividing by the standard error to obtain the variable

Mean square residual average degradation of, i.e. importance score of, feature variable

It can be expressed as:

wherein, the first and the second end of the pipe are connected with each other,

is as follows

A mean square residual of the samples;

is the standard error;

s224, feature selection: and sequentially removing the features with the minimum importance through feature selection, intercepting to obtain different numbers of feature subsets, respectively modeling and evaluating the feature subsets, and determining the optimal subset according to an evaluation result.

Further, the S100 includes:

s110, shutdown data filtering: selecting four main construction parameters of propulsion speed, cutter head rotating speed, total propulsion force and cutter head torque as discrimination indexes, wherein the discrimination functions are as follows:

in the above formula, the first and second carbon atoms are,

、

respectively corresponding to the total propelling force, the cutter head torque, the propelling speed and the cutter head rotating speed,

representing the product of four parameter values. When the product of the four parameters is not 0, the shield machine is in a working state, and corresponding data are reserved; when the product of the four parameters is 0, namely the value of any one parameter is 0, the shield machine is considered to be in a non-working state, and data in the non-working state are directly eliminated;

s120, abnormal value identification and processing: identifying and removing abnormal values by using a box type graph method, and filling the removed abnormal values by using data of the second before or the second after the abnormal values;

s130, constant parameter filtering: identifying and rejecting constant parameters

S140 missing value processing: the direct deletion method is combined with the constant filling method, and different situations of data missing and data dividing are processed, so that the validity of the data is better ensured.

According to a second aspect of the invention, a shield tunneling load advanced prediction method based on Bayesian optimization LSTM is provided, which comprises the following steps:

the first main module is used for acquiring monitoring data of shield construction parameters and performing data preprocessing;

the second main module is used for analyzing and filtering redundant parameters by using Pearson correlation analysis and selecting characteristics by using a random forest algorithm to obtain an optimal parameter set;

and the third main module is used for predicting the shield tunneling load of the Bayesian optimization LSTM based on the optimal parameter set.

According to a third aspect of the present invention, there is provided an electronic apparatus comprising:

at least one processor, at least one memory, and a communication interface; wherein the content of the first and second substances,

the processor, the memory and the communication interface are communicated with each other;

the memory stores program instructions executable by the processor, which invokes the program instructions to perform the method.

According to a fourth aspect of the invention, there is provided a non-transitory computer readable storage medium storing computer instructions which cause the computer to perform the method.

In general, compared with the prior art, the above technical solution contemplated by the present invention can achieve the following beneficial effects:

1. the invention discloses an advanced prediction method for shield tunneling load, provides a real-time prediction frame for shield tunneling load, and constructs an intelligent prediction system for shield tunneling load. The intelligent tunneling load prediction method and the intelligent tunneling load prediction system firstly move relevant data in a database for preprocessing, abnormal values, missing values and irrelevant data are cleaned, then parameter correlation analysis and feature selection are carried out based on the preprocessed data, redundant parameters are filtered, finally a Bayesian optimization LSTM tunneling load prediction model is constructed based on the selected optimal parameter set, advance prediction is respectively carried out on cutter head torque and total thrust, and timely and useful information and guidance are provided for tunneling construction of a tunnel;

2. the shield tunneling load advanced prediction method provides a new idea and a solution for the advanced prediction of the tunneling load by applying the LSTM network for modeling the time series data in the engineering field. In order to realize accurate advanced prediction of the tunneling load of the shield tunnel and provide guidance and reference for the operation of the shield machine;

3. the shield tunneling load advanced prediction method provided by the invention has the advantages that the acquired initial data is preprocessed, irrelevant data such as shutdown data, abnormal values and missing values are filtered, high coupling parameters are filtered through correlation analysis, on the basis, RF is utilized to carry out importance sorting and feature selection, the optimal parameter sets of cutter head torque and total propulsion force are respectively obtained, and conditions are provided for high-precision prediction of tunneling load based on a Bayesian optimization LSTM model.

4. According to the shield tunneling load advanced prediction method, the optimum hyperparameter of the LSTM prediction model is determined by adopting Bayesian optimization, and a cutterhead torque and total propulsive force prediction model is constructed on the basis of the selected hyperparameter, so that accurate advanced prediction of the tunneling load is realized, and reference is provided for the operation of a shield tunneling machine.

Drawings

Fig. 1 is a flowchart of a shield tunneling load advanced prediction method and system based on bayesian optimization LSTM according to an embodiment of the present invention;

fig. 2 is a schematic structural diagram of a shield tunneling load advanced prediction method and system based on bayesian optimization LSTM according to an embodiment of the present invention;

fig. 3 is a schematic physical structure diagram of an electronic device according to an embodiment of the present invention;

FIG. 4 is an example of an outlier identification result provided by an embodiment of the present invention;

fig. 5 is a diagram of the result of the evaluation of the importance of the shield parameters provided by the embodiment of the present invention (fig. 5 (a) is a cutterhead torque result, and fig. 5 (b) is a total thrust result);

fig. 6 is an MSE trend graph of different feature subsets according to an embodiment of the present invention (fig. 6 (a) shows a cutter torque trend, and fig. 6 (b) shows a total thrust trend);

fig. 7 is a bayesian optimization iterative process diagram provided by an embodiment of the present invention (fig. 7 (a) is a cutterhead torque iterative process, and fig. 7 (b) is a total thrust iterative process);

fig. 8 is a cutter head torque prediction result graph based on bayesian optimization LSTM according to an embodiment of the present invention (fig. 8 (a) is a cutter head torque actual value, fig. 8 (b) is a cutter head torque predicted value, and fig. 8 (c) is a comparison graph of the cutter head torque actual value and the predicted value);

fig. 9 is a total propulsion prediction result graph (fig. 9 (a) is a total propulsion actual value, fig. 9 (b) is a total propulsion predicted value, and fig. 9 (c) is a comparison graph of the total propulsion actual value and the predicted value) based on the bayesian optimization LSTM according to the embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

As shown in fig. 1, the invention provides a shield tunneling load advanced prediction method based on bayesian optimization LSTM, which comprises the following steps:

s200, performing redundant parameter analysis and filtration by using Pearson correlation analysis, and performing feature selection by using a Random Forest (RF) algorithm to obtain an optimal parameter set;

s300, based on the optimal parameter set, conducting Bayesian optimization LSTM shield tunneling load prediction.

Specifically, the S100 performs data preprocessing including filtering shutdown data and constant parameters in the original data, identifying and processing abnormal values and missing values in the data, and primarily cleaning irrelevant information and parameters in the data, including the following steps:

s110, shutdown data filtering:

the data in the shutdown state are mainly collected under the conditions of cutter replacement, shutdown rest, maintenance and the like of the shield machine, the judgment of the non-working state of the shield machine can be identified according to the construction parameters of the shield machine, and when the collected construction parameter data is 0, the shield machine can be considered to be in the non-working state. Selecting four main construction parameters of propulsion speed, cutter head rotating speed, total propulsion force and cutter head torque as discrimination indexes, wherein the discrimination functions are as follows:

（1）

（2）

in the above formula, the first and second carbon atoms are,

、

representing the product of four parameter values. When the product of the four parameters is not 0, the shield machine is in a working state, and corresponding data are reserved; when the product of the four parameters is 0, namely the value of any one parameter is 0, the shield machine is considered to be in a non-working state, and data in the non-working state are directly removedAnd (4) dropping.

the abnormal values are identified and removed by adopting a box type graph method, the box type graph method has no restrictive requirement on data, the distribution form of the data does not need to be assumed in advance, and the box type graph method is used for judging the abnormal values based on the quartile. The box graph method is used for identifying abnormal values, the upper limit and the lower limit of normal values are determined according to the quartile, and the calculation method comprises the following steps:

（3）

（4）

（5）

wherein the content of the first and second substances,

、

respectively representing the lower quartile and the upper quartile,

the data are arranged in the order from small to large for the number of data,

is the difference between the four-part numbers,

、

lower and upper limits of the data, respectively. Upon identification of an abnormal value, when the data value is at

In between, normal data, when the data value is less than

Or greater than

If so, the abnormal value can be determined and should be directly removed. For removed outliers, the data from the second before or after the value is used for padding.

S130 constant parameter filtering: identifying and rejecting constant parameters

In the whole shield tunneling system, the set values of some shield parameters are constant, and the parameters belong to constant parameters. Because they do not change along with the advancing of the shield machine in the construction process, the data of the shield machine has no great reference significance for the prediction and the control of the construction parameters. In order to reduce the interference of unnecessary data and the workload of subsequent analysis, the constant parameters can be directly filtered out, identified by observation and eliminated.

The S140 specifically includes the following steps:

s141 deletes directly: when the data missing value proportion of one characteristic parameter exceeds 30%, the available effective information in the parameter data is considered to be too little and should be directly deleted;

s142 constant padding: and when the proportion of the missing value in the parameter data is lower than 30%, the parameter data is indicated to have more available effective information, and the missing value is filled by adopting the median of the parameter data.

Specifically, the S200 specifically includes:

s210 Pearson parameter correlation analysis: measuring a linear correlation between the parameters using a Pearson Correlation Coefficient (PCC);

to further analyze the relationship between different parameters, a linear correlation between the parameters was measured using Pearson Correlation Coefficient (PCC). The calculation formula of the Pearson correlation coefficient is expressed as:

（6）

in the above formula, the first and second carbon atoms are,

is that

Measurement of

And

measurement of

The covariance between the two is such that,

and

respectively corresponding to the average values of the two variables,

and

the deviation of the two variables is varied separately. Correlation coefficient

Has a value in the interval [ -1,1 [)]According to the size and the positive and negative of the value, the correlation between the parameters can be analyzed and judged. When in use

When, a positive correlation between the two variables is illustrated; when the temperature is higher than the set temperature

When, a negative correlation between the two variables is illustrated. When in use

Has a value of [0,0.09]In between, it indicates no correlation between the two parameters; when in use

Has a value of [0.09,0.3 ]]In between, it indicates that there is low correlation between the two parameters; when in use

Has a value of [0.3,0.5 ]]In between, a moderate correlation between the two parameters is indicated; when in use

Has a value of [0.5, 1]]In between, a significant correlation between the two parameters is illustrated.

And (3) evaluating the influence of the change of the characteristic variable on the Gini index based on the RF algorithm so as to measure the importance of the variable. The method is mainly based on the evaluation of the importance degree of the OBB error to the parameters, after the parameters are randomly replaced, the model mean square residual error reduction (% Inc MSE) or the model precision reduction (Inc Node Purity) is measured, so that the importance of the characteristic parameters can be judged, and the calculation steps are as follows:

s221, a regression decision tree is established. Predicting OOB (out-of-bag data) by using a random forest model to obtain the mean square residual errors of the b out-of-bag data, wherein the mean square residual errors are

The calculation formula is as follows:

（7）

wherein the content of the first and second substances,

in order to be the amount of data,

representing the true value of the dependent variable in the out-of-bag data;

Characteristic parameter

And calculating a new out-of-bag error accuracy

A value;

s222 generates an error matrix. When generating the regression decision tree, randomly selecting the characteristic parameters to split the decision tree and dividing the parameters

In that

Random substitution is performed in each data sample outside the bag, thereby forming aPredicting the new OOB test set again by using the established random forest regression model to obtain a new OOB residual mean square

Generating an error matrix

Comprises the following steps:

（8）

wherein

The number of influencing factor variables;

the number of training sample sets;

s223 performs importance scoring. By using

It can be expressed as:

（9）

is as follows

A mean square residual of the samples;

is the standard error. The higher the importance score of a characteristic variable, the higher the importance of that variable to the model outcome.

The importance scoring is carried out on the characteristic variables through the random forest, although the sorting result of the variables can be obtained, it cannot be determined which variables should be filtered, and the further screening of the parameters needs to be carried out by combining a characteristic selection method. And sequentially removing the features with the minimum importance through feature selection, intercepting to obtain different numbers of feature subsets, respectively modeling and evaluating the feature subsets, and determining the optimal subset according to an evaluation result.

The advanced prediction of the shield tunneling load in the step S300 comprises prediction of shield cutterhead torque and total propelling force, and the step S300 specifically comprises the following steps:

（10）

wherein the content of the first and second substances,

which represents the standard value after the normalization,

and

defaults to 1 and 0, respectively, indicatingThe normalized maximum and minimum values are then compared to each other,

represents the value of the sample,

and

maximum and minimum values of sample values, respectively;

in order to eliminate the influence of different parameter scales on the model solving efficiency and precision, the data needs to be normalized and scaled to the same order of magnitude. And a min-max normalization method is selected, the calculation is simple, and the relation between original data can be reserved. The present invention chooses to normalize the data to the interval [0,1 ]. The calculation formula of the normalization process is as follows:

（10）

wherein the content of the first and second substances,

which represents the standard value after the normalization,

and

represents the value of the sample,

and

are respectively a sample valueMaximum and minimum values of.

the Bayesian optimization method is adopted for parameter adjustment, the complex parameter search problem can be solved only by a small number of iterations, the optimal value of the hyper-parameter is obtained, and the method is suitable for the hyper-parameter adjustment of the LSTM prediction model.

The step S320 specifically includes the steps of:

And S330, evaluating and comparing the model performance.

S331 adopts a determination coefficient (

) Root mean square error (

) And average absolute error (

) Three general indicators are used to evaluate the performance of the prediction model.

Using a coefficient of determination (

) Root mean square error (

) And average absolute error (

The goodness of fit of the model predicted value and the actual observed value is measured, the value range is 0-1, the closer the value is to 1, the more perfect the fit of the model is, and the better the prediction performance is.

The standard deviation of the prediction error is calculated,

the difference between the predicted value of the model and the actual observed value is calculated, and the difference measures the deviation between the predicted value and the actual value. The smaller their values are, the smaller the deviation between the predicted value and the actual value of the model is, and the higher the prediction accuracy of the model is. The calculation formulas of the three evaluation indexes are as follows:

（11）

（12）

（13）

wherein, therein

The total number of data in the sample data set;

and

respectively representing a model predicted value and an actual observed value;

the average actual value is indicated.

In order to further verify the reliability of the Bayesian optimization LSTM prediction model established by the invention, the Bayesian optimization LSTM prediction model is compared with the prediction performance of other common models. RF, SVM and LSTM are the main methods for cutterhead torque and total thrust prediction, so three models, RF, SVM and LSTM, were chosen for comparison. In order to make the comparison result of the models more fair, the three prediction models of RF, SVM and LSTM and the Bayesian optimization LSTM prediction model adopt the same programming environment, performance evaluation indexes and data samples.

In order to make the objects, technical solutions and advantages of the present invention more clearly understood, in the following, with reference to the accompanying drawings, the present invention performs a related study on a fifth standard section interval tunnel of a first-stage project of 16-line rail transit in wuhan city, where the standard section is an interval from a guan village station to a guo bo central south station, and further describes a specific implementation process of the method of the present invention by taking the obtained project site data as an example:

(1) data pre-processing

The original data comprises 121 shield construction parameters of a shield machine cutter head driving system, a shield system, a propulsion system and a supporting system, wherein two parameters of cutter head torque and total propulsion are prediction objects, the time interval of each group of data is 3s, and the data amount is 3627 groups.

Due to the reasons of shield machine halt, sensor signal interference and the like, a large amount of abnormal data and invalid data exist in original data, and the data need to be preprocessed for subsequent analysis. According to the foregoing, the data cleaning and missing value processing are sequentially performed on the original data, and the method mainly includes the following four steps:

1) and (3) shutdown data filtering: and (3) judging the shutdown data in the original data set according to the formula (1) and the formula (2), identifying 27 groups of shutdown data, and directly removing the shutdown data.

2) Outlier identification and processing: abnormal value identification is performed by adopting a box graph method according to the formulas (3), (4) and (5), and the abnormal value identification result is shown in fig. 4 by taking the cutter head power, the propelling pressure, the hinging pressure, the shield tail sealing pressure and the output current of the frequency converter as examples. When the data value is less than Lmin or greater than Lmax, the abnormal value can be judged and directly removed. After abnormal value identification and data elimination, the number of parameters of the shield machine data set is reduced from 121 to 100, and the eliminated parameters are mainly parameters with all monitoring values of 0. For removed outliers, the data from the second before or after the value is used for padding.

3) Constant parameter filtering: 43 constant parameters are identified by observing the original data set, the constant parameters have no significance for the prediction of the tunneling load, the constant parameters are directly removed, and 57 parameters are also arranged in the removed data set.

4) Missing value processing: through analysis and calculation of all the data, the missing value proportion of all the parameter data is lower than 30%, which indicates that no parameter needs to be deleted, and the data of all the parameters have available information. And filling missing values of the remaining parameters by adopting a median filling method, and after the missing values are processed, obtaining 57 parameters in total.

(2) Parameter correlation analysis

And (3) carrying out correlation analysis on the remaining 57 parameters subjected to low variance value filtering, calculating correlation coefficients among the parameters according to an equation (6), removing redundant parameters according to the magnitude of the correlation coefficients, and reserving only one variable with high correlation. In order to retain effective information as much as possible, the discrimination interval of the redundant parameters is set to [0.8,1], when the absolute value of the correlation coefficient between two parameters is greater than 0.8, the correlation between the two parameters is extremely high, the similarity of the data information of the two parameters can be considered to be extremely high, and only one of the parameters needs to be retained.

According to the calculation result of the correlation coefficient, the absolute values of the correlation coefficients of 10 parameters such as the equipment bridge pressure, the hinge displacement, the total accumulated quantity of industrial water, the foam stock solution tank pressure and the like are all larger than 0.8, the absolute values have extremely high information similarity, and only the equipment bridge pressure parameter needs to be reserved. The correlation coefficients of the cutter head power, the output power of the frequency converter and the output torque of the frequency converter are all larger than 0.95, and only the cutter head power parameter needs to be reserved. After parameter screening according to the above analysis, 20 input parameters were finally retained.

The correlation results of the remaining 21 parameters and the cutter head torque and the total thrust force show that some of the remaining 21 input parameters still have significant correlation, such as the correlation between the equipment bridge pressure and the shield tail sealing pressure is-0.738, and the correlation is significant. There may still be coupling information between these parameters, requiring further analysis filtering. In addition, the output parameters of the cutter head torque have obvious correlation with cutter head power, output current of a frequency converter, propelling speed and the like, and the total propelling force has obvious correlation with propelling pressure, hinging pressure and the like, which shows that the parameters have great influence on the value of the tunneling load.

(3) Feature selection

In order to further filter redundant parameters, input parameters are subjected to importance analysis and feature selection by using a random forest, and an optimal parameter set is determined.

1) And establishing a regression decision tree. Predicting OOB (out-of-bag data) by using a random forest model to obtain the mean square residual errors of b out-of-bag data, wherein the mean square residual errors are respectively

The calculation formula is as follows:

（7）

wherein the content of the first and second substances,

in order to be able to measure the amount of data,

representing the true value of the dependent variable in the out-of-bag data;

Characteristic parameter

And calculating a new out-of-bag error accuracy

The value is obtained.

2) An error matrix is generated. When generating the regression decision tree, randomly selecting the characteristic parameters to split the decision tree and dividing the parameters

In that

Generating an error matrix

Comprises the following steps:

（8）

wherein

The number of influencing factor variables;

the number of training sample sets.

3) An importance score is made. By using

It can be expressed as:

（9）

wherein the content of the first and second substances,

is as follows

A mean square residual of the samples;

is the standard error. The higher the importance score of a characteristic variable, the higher the importance of the variable to the model result.

In Python, importance evaluation is performed on input parameters by using a random forest, the number of leaf nodes is set to 60, the depth of a tree is set to 12, cutter torque and total thrust are respectively used as output variables, importance scores of the remaining 21 parameters are calculated and ranked, and the result is shown in fig. 5.

As can be seen from fig. 5, the cutter head power, the output current of the frequency converter, and the screw machine rotation speed are the first three parameters of the importance ranking of the cutter head torque, and the propulsion pressure, the hinge pressure, and the earth bin pressure are the first three parameters of the importance ranking of the total propulsion thrust. From engineering practice experience, the torque of the cutter head is directly related to the torque of the cutter head, and the size of the torque of the cutter head is closely inseparable from the size of the power and the rotating speed of the cutter head, so that the power of the cutter head has obvious influence on the torque of the cutter head, and the size of the power of the cutter head has great correlation with the output current of a frequency converter, so that the power of the cutter head, the output current of the frequency converter and the rotating speed of a screw machine really have great influence on the torque of the cutter head. The total propelling force of the shield machine is the sum of the propelling pressure and other various pressures, the value of the total propelling force is directly related to the propelling pressure, and the total propelling force has certain relevance to other pressures in the shield system. In conclusion, the results of ranking the importance of the parameters of the cutter head torque and the total thrust force obtained based on the RF algorithm are reasonable.

And 5-fold cross validation is adopted to test the model accuracy, and feature selection is carried out on the basis of the importance sorting result to obtain the variation trend of Mean Square Error (MSE) of different feature subsets, as shown in figure 6. As can be seen from fig. 6, the model error MSE generally shows a trend of decreasing first and then increasing, which illustrates that the prediction accuracy of the model is effectively improved as some unimportant parameters are eliminated; when the number of the parameters of the feature subset reaches a certain value, the model prediction accuracy is reduced due to the fact that feature parameters are continuously removed, which means that important features are mistakenly deleted when feature parameters are further removed.

As can be seen from fig. 6, for the cutter head torque, when the number of parameters in the feature subset is 10, the mean square error value is minimum, and the model precision is highest; for the total thrust, when the number of parameters in the feature subset is 9, the mean square error value is minimum, and the model precision is highest, that is, the feature parameters at this time are respectively the optimal parameter sets of the cutter head torque and the total thrust. On the whole, the RF algorithm is combined with the feature selection for parameter screening, so that irrelevant feature parameters can be effectively eliminated, and the prediction precision of the model is further improved. The optimal feature subsets screened by combining the parameter importance ranking of fig. 5 are shown in table 1, the screened parameters are all parameters which are very important to cutter head torque and total thrust, and the parameters are used as input variables to construct a subsequent tunneling load prediction model.

(4) Bayesian optimization LSTM-based tunneling load prediction

1) Data normalization processing

And constructing a tunneling load Bayesian optimization LSTM prediction model based on the data filtered by the redundant parameters, wherein the data characteristics are shown in Table 2. In order to avoid the influence of inconsistent data levels on the training effect and precision of the model, the data are normalized according to the formula (10), so that all variables are in the same dimension. And after normalization processing, diversity is carried out on the data, the first 80% of the data is used as a training set for training an LSTM prediction model, and the second 20% of the data is used as a test set for verifying the performance of the prediction model. In order to ensure the reasonability of the model training process, the training set is subjected to diversity, wherein 80% of the training set is used as the training set, and 20% of the training set is used as the test set for model training.

2) Bayesian-based LSTM hyper-parameter selection

The Bayesian optimization method is adopted for parameter adjustment, the complex parameter search problem can be solved only by a small number of iterations, the optimal value of the hyper-parameter is obtained, and the method is suitable for the hyper-parameter adjustment of the LSTM prediction model. The method comprises the following steps:

2.1) defining a model training error MSE to be minimized into a target function, randomly generating an initialization parameter combination in an optimization range, training an LSTM prediction model according to the initialization parameter combination, and obtaining a model output result MSE value;

2.2) introducing the initialization parameter combination into a Gaussian process, taking a fitted initial Gaussian model as a proxy model of a target function, and correcting the initial proxy model by using the LSTM prediction model training error MSE value obtained in the last step to enable the distribution of the initial proxy model to be closer to the real distribution of the model training error MSE;

2.3) selecting the minimum value of the target function as the next parameter combination to be evaluated through the expected improved acquisition function, and repeating the LSTM prediction model training and Gaussian model correction of the previous two steps;

and 2.4) stopping the optimization process when the maximum iteration times are reached, wherein the corresponding super-parameter combination is the optimal super-parameter when the MSE value of the LSTM prediction model is minimum.

Optimizing the super-parameters of the LSTM prediction model by adopting Bayesian optimization, taking an MSE loss function of a training model as an objective function, wherein the optimized super-parameters comprise the number of network hidden layers (num layers), the number of hidden layer neurons (num units), the learning rate (learning rate) and the batch size (batch size), the number of input variables of the network input layer according to the cutter torque and the total propulsion is respectively 10 and 9, the number of network output layers is 1, the other parameters are kept in default settings, and the number of model iterations is controlled to be 50.

The model iteration number is controlled to be 50, and the Bayes optimization iteration process of the cutter head torque and the total propelling force is shown in FIG. 7. The minimum observed target in the graph is the minimum observed value before each iteration, namely the optimal value at the end of the iteration, and the estimated minimum target represents the minimum estimated value obtained according to the Bayesian optimization algorithm and determines which point is selected as the target. As can be seen from (a) and (b) in fig. 7, the estimated minimum target of the cutterhead torque and the total thrust is very close to the point line graph of the minimum observed target, which shows that the bayesian optimization process of the cutterhead torque and the total thrust is reasonable. In 50 iterations, the cutter head torque prediction model finds the minimum value of the objective function at the 22 th time, the total thrust prediction model finds the minimum value of the objective function at the 26 th time, the iteration times required for optimization are small, the Bayes optimization can be seen to quickly find the optimal hyper-parameter combination of the models, the search process is stable, and the search efficiency is high. The optimal hyperparametric results of LSTM obtained by bayesian optimization are shown in table 3.

3) Model performance evaluation and comparison

Performing hyper-parameter setting on the prediction model based on the result of Bayesian optimization, respectively establishing a Bayesian optimization LSTM prediction model for the cutter head torque and the total propulsion, and using the prediction model according to the formulas (11), (12) and (13)

、

And

three indicators were used to evaluate the predictive performance of the model. The results of the predictions of the cutterhead torque and the total thrust on the training set and the testing set obtained based on the Bayesian optimization LSTM prediction model are shown in FIG. 8.

As can be seen from (a), (b) and (c) in fig. 8, predicted on the cutter head torque training set

、

And

the model fitting degree is high, the training error is small, and the relation between the shield parameters and the cutter head torque is well learned through the Bayes optimization LSTM prediction model. Predicted cutter head torque on test set

、

And

the prediction accuracy of the model is high, which means that the accurate advance prediction of the cutterhead torque can be realized based on the Bayesian optimization LSTM prediction model, and the cutterhead torque can be predicted more accurately by training the first 80% of data, and the accurate advance prediction of the cutterhead torque can be realized based on the Bayesian optimization LSTM prediction model.

As can be seen from (a), (b) and (c) in FIG. 9, the prediction on the total propulsive force training set

、

And

0.919, 187.34, and 76.25, respectively, of predicted results on the test set

、

And

the prediction errors of the model test set are slightly higher than those of the training set, but the model training errors and the test errors are smaller, and the goodness of fit is about 0.9, which shows that the training is carried out through the first 80% of data, the total propulsive force of 20% after the prediction can be more accurate, a high-accuracy prediction result can be obtained based on the Bayesian optimization LSTM prediction model, and the accurate advanced prediction of the total propulsive force is realized.

In order to further verify the reliability of the Bayesian optimization LSTM prediction model established in the method, three models of LSTM, RF and SVM are adopted to predict the tunneling load of the shield tunnel, the prediction result is compared and analyzed with the Bayesian optimization LSTM prediction result, and the same input and output are selected for a plurality of prediction models to train and predict.

The predicted performance of the four predictive models on the cutter head torque and total thrust in the training set and the test set are shown in table 4. As can be seen from Table 4, for the training set of cutter head torque and total thrust, four models are provided

The goodness of fit of the Bayes optimization LSTM prediction model is the highest, the root mean square error and the average relative error are the smallest, for a test set of cutter torque and total propulsion, the goodness of fit of the Bayes optimization LSTM prediction model in the four models is still the highest, and the error is the smallest, so that the Bayes optimization LSTM prediction model can better capture the actual value of the predicted target parameter and realize more accurate tunneling load prediction.

Based on the foregoing embodiment, as an optional embodiment, the shield tunneling load advanced prediction system based on bayesian optimization LSTM provided in the embodiment of the present invention includes:

It should be noted that, the apparatus in the apparatus embodiment provided by the present invention may be used for implementing methods in other method embodiments provided by the present invention, except that corresponding function modules are provided, and the principle of the apparatus embodiment provided by the present invention is basically the same as that of the apparatus embodiment provided by the present invention, so long as a person skilled in the art obtains corresponding technical means by combining technical features on the basis of the apparatus embodiment described above, and obtains a technical solution formed by these technical means, on the premise of ensuring that the technical solution has practicability, the apparatus in the apparatus embodiment described above may be modified, so as to obtain a corresponding apparatus class embodiment, which is used for implementing methods in other method class embodiments.

The method of the present invention is implemented by means of electronic devices, and therefore, a description of related electronic devices is necessary. To this end, an embodiment of the present invention provides an electronic apparatus, as shown in fig. 3, including: the system comprises at least one processor (processor), a communication Interface (communication Interface), at least one memory (memory) and a communication bus, wherein the at least one processor, the communication Interface and the at least one memory are communicated with each other through the communication bus. The at least one processor may invoke logic instructions in the at least one memory to perform all or a portion of the steps of the methods provided by the various method embodiments described above.

In addition, the logic instructions in the at least one memory may be implemented in software functional units and stored in a computer readable storage medium when sold or used as a stand-alone product. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the method embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment may be implemented by software plus a necessary general hardware platform, and may also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. Based on this recognition, each block in the flowchart or block diagrams may represent a module, a program segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

In this patent, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A shield tunneling load advanced prediction method based on Bayesian optimization LSTM is characterized by comprising the following steps:

s200, performing redundant parameter analysis and filtration by using Pearson correlation analysis, and performing feature selection by using a random forest algorithm to obtain an optimal parameter set;

2. The Bayesian optimization LSTM-based shield tunneling load advanced prediction method according to claim 1, wherein the S300 comprises:

wherein the content of the first and second substances,

which represents the standard value after the normalization,

and

represents the value of the sample,

and

maximum and minimum values of the sample values, respectively;

s320, LSTM hyperparameter selection based on Bayes: adjusting parameters by a Bayesian optimization method to obtain optimal values of the hyper-parameters;

and S330, evaluating and comparing the model performance.

3. The Bayesian optimization LSTM-based shield tunneling load advanced prediction method according to claim 2, wherein the S320 specifically comprises the following steps:

4. The Bayesian optimization LSTM-based shield tunneling load advanced prediction method according to claim 2, wherein the S330 specifically comprises:

s332 compares the prediction performance of the Bayesian optimization LSTM prediction model with other common models.

5. The Bayesian optimization LSTM-based shield tunneling load advanced prediction method according to claim 1, wherein the S200 comprises the following steps:

6. The Bayesian optimization LSTM-based shield tunneling load advanced prediction method according to claim 1, wherein S220 specifically comprises:

The calculation formula is as follows:

wherein the content of the first and second substances,

in order to be able to measure the amount of data,

representing the true value of the dependent variable in the out-of-bag data;

Characteristic parameter

And calculating a new out-of-bag error accuracy

A value;

s222 generates an error matrix: error matrix

Comprises the following steps:

wherein

The number of influencing factor variables;

the number of training sample sets;

the mean square residual error of the data outside the bag when the number of the influencing factor variables is i and the number of the training sample sets is j, i =1,2,3, … …, p, j =1,2,3, … …, b;

s223 performing importance scoring: obtaining variables

It can be expressed as:

wherein the content of the first and second substances,

is as follows

A mean square residual of the samples;

is the standard error;

the number of training sample sets;

for the out-of-bag data of the influence factor variable i and j sample setI =1,2,3, … …, p, j =1,2,3, … …, b;

7. The Bayesian optimization LSTM-based shield tunneling load advance prediction method according to claim 1, wherein the S100 comprises:

in the above formula, the first and second carbon atoms are,

、

representing the product of four parameter values; when the product of the four parameters is not 0, the shield is explainedThe mechanism is in a working state, and corresponding data are reserved; when the product of the four parameters is 0, namely the value of any one parameter is 0, the shield machine is considered to be in a non-working state, and data in the non-working state are directly eliminated;

s130, constant parameter filtering: identifying constant parameters and removing the constant parameters;

8. A shield tunneling load advanced prediction system based on Bayesian optimization LSTM is characterized by comprising the following steps:

the second main module is used for analyzing and filtering redundant parameters by utilizing Pearson correlation analysis and selecting characteristics by adopting a random forest algorithm to obtain an optimal parameter set;

9. An electronic device, comprising:

the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method of any one of claims 1 to 7.