Tunnel construction water inflow calculation method based on machine learning

Abstract

The invention discloses a tunnel construction water inflow calculation method based on machine learning, and belongs to the field of high-precision prediction of tunnel water inflow disasters. Precise prediction of the tunnel water inflow is one of the focus problems concerned by underground engineering and is directly related to the engineering construction progress, the cavern stability and the safety of constructors. Therefore, a method for calculating the water inflow of the tunnel by constructing the prediction model by integrating multiple machine learning algorithms is provided. The method comprises the steps that a geological condition database is established for a tunnel section with the water inflow needing to be calculated according to drilling data, field investigation data and long-term observation data in the early stage, then screening analysis is conducted on the data in the database, influence factors are selected and preprocessed, iteration updating of a prediction model is conducted through the selected influence factors, and the prediction result is obtained. Finally, a high-precision tunnel water inflow prediction result can be obtained through calculation by inputting on-site investigation data of a tunnel construction area.

Description

Technical Field

[0001] The present invention relates to the field of high-precision prediction of tunnel water inrush disasters, and particularly to a calculation method for tunnel construction water inflow based on machine learning. Background Technique

[0002] In various projects such as water conservancy, transportation, and underground mineral mining, tunnel excavation is involved and sudden water inrush accidents may occur. The occurrence probability of sudden water inrush is high, the water inrush volume is large, and the water pressure is high. Moreover, because its occurrence process is sudden and the location cannot be easily judged accurately, it is also very difficult to predict its scale and dynamic characteristics. The problem of sudden water inrush often affects the project construction, causes the surrounding rock to become unstable, falls off and blocks the tunnel (cavity), and at the same time causes ground collapse, ground settlement and pollution of the surrounding water environment, affecting the long-term operation and the ecological environment on the ground, and even causing more serious life safety and property losses.

[0003] At present, there are many calculation methods for water inflow, which can be mainly summarized into the following several types: approximate methods, theoretical methods, stochastic mathematics methods, non-linear theory methods, and numerical methods. Among them, approximate methods mostly come from the summary of engineering practice and have good prediction accuracy for calculating water inrush problems under simple geological conditions. The calculation results of theoretical methods often need to be corrected by combining with measured data, and the theoretical accuracy may be insufficient in practical applications. Stochastic mathematics methods are sensitive to abnormal data, and the results may be significantly affected by data noise or modeling errors. Non-linear theory methods require a large amount of fine observational data, especially in terms of time series and spatial distribution. Numerical methods need to be verified by field data or theoretical models, and it may be difficult to judge the credibility of the results when lacking experience. Machine learning methods are emerging means for quantitative analysis and calculation of groundwater and are also one of the important means, especially applicable under complex geological conditions.

[0004] At present, in the aspect of accurate prediction of tunnel water inrush, the methods for obtaining various hydrogeological parameters are relatively perfect, but how to use each parameter to complete simple, fast, convenient and effective water inrush prediction is still a difficult point. Therefore, based on algorithms such as Random Forest (RF) and BP (Back Propagation) neural network, the present invention constructs an accurate prediction model for tunnel water inrush to make up for the blank in the field of high-precision prediction of tunnel water inrush disasters. Summary of the Invention

[0005] Object of the Invention. The object of the present invention is to provide a calculation method for tunnel construction water inflow based on machine learning for the problem of solving tunnel water inflow by integrating machine learning models. This method has low time cost, high accuracy and is relatively convenient.

[0006] Technical Solution. To achieve the above object, the present invention provides a calculation method for tunnel construction water inflow based on machine learning, and the method includes the following specific steps:

[0007] 1) Collect geological environment data in the surrounding areas of tunnel projects, including drilling data, surface survey data, underground rock formation data, and hydrological information, and import them into the heterogeneous water gushing disaster big data repository;

[0008] 2) Using the Oshima formula and the modified Goodman formula, the water inflow per unit length and the maximum water inflow of the tunnel in the calculation area are preliminarily determined, and the actual level of water inflow disaster is determined by the water inflow per unit length, and the danger level of water inflow disaster is determined according to the maximum water inflow per unit length;

[0009] 3) Import the multi-source exploration information of the research section into a heterogeneous data repository built on the Hadoop system. According to the fitting results of the field data of the section and the field data of the previous water inrush accident area in the original database, fill the data gaps in the prediction section and eliminate abnormal data;

[0010] 4) Input the data in the database into the simulation software Groundwater Model System (GMS) for 3D geological modeling, and make supplementary corrections to the infill data based on the differences between the modeling and the actual conditions to complete further fitting and updating of the database;

[0011] 5) Using the negative terrain type, stratum lithology, adverse geological type, structural surface type and development degree, karst development degree, groundwater storage or water inflow type, rainfall, construction disturbance degree, rock formation inclination and rock formation inclination, permeability coefficient, groundwater level, horizontal and vertical distance of groundwater body from tunnel, tunnel length and tunnel depth, and groundwater head information in the database as input information of the prediction model, the initial model is iteratively updated, so that the predicted water inflow per unit length of the tunnel output by the model and the actual water inflow per unit length of previous water inflow disasters have a fitting degree of more than 85%, and the training of the neural network model is completed;

[0012] 6) Collect the above data of the study area and input them into the trained neural network model to output the water inflow per unit length of the tunnel, and use the water inflow per unit length to determine the actual level of water inflow disaster.

[0013] Furthermore, in the step 1), when collecting engineering example data, construction tunnels with complete preliminary exploration data are selected, the missing rate of exploration information should be less than 30%, and the data of the surrounding areas should at least include information on tunnel water pressure, surrounding rock permeability coefficient, surrounding rock fragmentation degree, structural surface type and development degree, tunnel length and burial depth.

[0014] Furthermore, in step 2), the maximum water inflow per unit length of the tunnel is calculated as follows:

[0015]

[0016] In the formula, q 0 is the maximum water inflow per unit length of the tunnel (m 3 / d); K is the permeability coefficient of the surrounding rock (m / d); m is the empirical correction coefficient with a value of 0.86; h is the distance from the initial groundwater level line to the tunnel floor (m); r 0 is the equivalent circular radius of the tunnel cross-section (m); d is the equivalent circular diameter of the tunnel cross-section (m);

[0017] The corresponding relationship between the danger level of tunnel water inrush disaster and the predicted maximum water inflow is as follows: Grade I, q 0 > 1 (10 3 m 3 / d), extremely dangerous, with disastrous consequences: Grade II, 0.3 < q 0 < 1 (10 3 m 3 / d), highly dangerous, with serious hazards: Grade III, 0.05 < q 0 < 0.3 (10 3 m 3 / d), moderately dangerous, delaying construction: Grade IV, q 0 < 0.05 (10 3 m 3 / d), low danger;

[0018] The calculation of the water inflow per unit length of the tunnel is as follows:

[0019]

[0020] In the formula, Q is the water inflow per unit length of the tunnel (m / d); H is the height from the groundwater level to the tunnel center line (m); r is the tunnel radius (m); K is the permeability coefficient of the surrounding rock (m / d). This formula assumes that the initial groundwater level is horizontal and unchanged, the groundwater is in a stable state and satisfies radial flow, and the medium is homogeneous and isotropic;

[0021] The corresponding relationship between the occurrence level of tunnel water inrush disaster and the predicted water inflow is as follows: Grade I, Q > 5 (10 2 m 3 / d), extremely high possibility: Grade II, 1.5 < q < 5 (10 2 m 3 / d), high possibility: Grade III, 0.25 < q < 1.5 (10 2 m 3 / d), medium possibility: Grade IV, Q < 0.25 (10 2 m 3 / d), low possibility.

[0022] Further, in step 3), the established data repository should contain the material information in step 1), and screen and clean the existing information, and eliminate the data with a relative error of more than 50% in the fitting result with the actual situation. When using the data repository to complement the geological information of the study area, it should be further ensured that the proportion of missing data in the total data should be less than 25% to avoid the occurrence of data distortion.

[0023] Further, in step 4), the fitting coefficient between the corrected GMS three-dimensional data model and the actual situation should be kept above 0.9 to ensure the accuracy of the filled data.

[0024] Further, in step 5), the selected model is any one of SVM, AdaBoost, and BP.

[0025] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0026] A method for calculating the water inflow during tunnel construction based on machine learning proposed by the present invention is closer to the actual situation and has more accurate calculation ability compared with the traditional method that simply relies on raw data for calculation and modeling; compared with the traditional calculation method, this machine learning model pays more attention to the non-linear relationship between various calculation factors, and relies on in-depth learning of this non-linear relationship to obtain accurate water inflow calculation results, and is simple to operate and low in cost; when calculating the water inflow of the tunnel by integrating multiple machine learning algorithm methods, the method of the present invention can quickly and accurately obtain the predicted water inflow of different sections of the underground tunnel according to the on-site inspection data. In actual engineering applications, the average error between the water inflow calculated by this method and the actual water inflow is less than 7%, which has high economic and application value. The present invention is expected to provide an innovative technical idea for calculating the water inflow of tunnels by using machine learning algorithms to construct accurate prediction models. Description of the Drawings

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art.

[0028] Figure 1 It is a flow chart of solving the water inflow during tunnel construction based on machine learning involved in the present invention. Detailed Embodiments

[0029] The following further illustrates the present invention in conjunction with the drawings and specific embodiments.

[0030] The present invention provides a method for calculating the water inflow during tunnel construction based on machine learning, and this method includes the following specific steps:

[0031] 1) Collect geological environment data in the surrounding areas of tunnel projects, including drilling data, surface survey data, underground rock formation data, and hydrological information, and import them into the heterogeneous water gushing disaster big data repository;

[0032] 2) Using the Oshima formula and the modified Goodman formula, the water inflow per unit length and the maximum water inflow of the tunnel in the calculation area are preliminarily determined, and the actual level of water inflow disaster is determined by the water inflow per unit length, and the danger level of water inflow disaster is determined according to the maximum water inflow per unit length;

[0033] 3) Import the multi-source exploration information of the research section into a heterogeneous data repository built on the Hadoop system. According to the fitting results of the field data of the section and the field data of the previous water inrush accident area in the original database, fill the data gaps in the prediction section and eliminate abnormal data;

[0034] 4) Input the data in the database into the simulation software Groundwater Model System (GMS) for 3D geological modeling, and make supplementary corrections to the infill data based on the differences between the modeling and the actual conditions to complete further fitting and updating of the database;

[0035] 5) Using the negative terrain type, stratum lithology, adverse geological type, structural surface type and development degree, karst development degree, groundwater storage or water inflow type, rainfall, construction disturbance degree, rock formation inclination and rock formation inclination, permeability coefficient, groundwater level, horizontal and vertical distance of groundwater body from tunnel, tunnel length and tunnel depth, and groundwater head information in the database as input information of the prediction model, the initial model is iteratively updated, so that the predicted water inflow per unit length of the tunnel output by the model and the actual water inflow per unit length of previous water inflow disasters have a fitting degree of more than 85%, and the training of the neural network model is completed;

[0036] 6) Collect the above data of the study area and input them into the trained neural network model to output the water inflow per unit length of the tunnel, and use the water inflow per unit length to determine the actual level of water inflow disaster.

[0037] Furthermore, in the step 1), when collecting engineering example data, construction tunnels with complete preliminary exploration data are selected, the missing rate of exploration information should be less than 30%, and the data of the surrounding areas should at least include information on tunnel water pressure, surrounding rock permeability coefficient, surrounding rock fragmentation degree, structural surface type and development degree, tunnel length and burial depth.

[0038] Furthermore, in step 2), the maximum water inflow per unit length of the tunnel is calculated as follows:

[0039]

[0040] Where, q 0 is the maximum water inflow per unit length of the tunnel (m 3 / d); K is the permeability coefficient of the surrounding rock (m / d); m is the empirical correction coefficient with a value of 0.86; h is the distance from the initial groundwater level line to the tunnel floor (m); r 0 is the equivalent circular radius of the tunnel cross-section (m); d is the equivalent circular diameter of the tunnel cross-section (m);

[0041] The corresponding relationship between the risk level of tunnel water inrush disaster and the predicted maximum water inflow is as follows: Grade I, q 0 > 1 (10 3 m 3 / d), extremely high risk, disastrous consequences: Grade II, 0.3 < q 0 < 1 (10 3 m 3 / d), high risk, serious harm: Grade III, 0.05 < q 0 < 0.3 (10 3 m 3 / d), medium risk, delaying construction: Grade IV, q 0 < 0.05 (10 3 m 3 / d), low risk;

[0042] The calculation of the water inflow per unit length of the tunnel is as follows:

[0043]

[0044] Where, Q is the water inflow per unit length of the tunnel (m / d); H is the height from the groundwater level to the tunnel center line (m); r is the tunnel radius (m); K is the permeability coefficient of the surrounding rock (m / d). This formula assumes that the initial groundwater level is horizontal and unchanged, the groundwater is in a stable state and satisfies radial flow, and the medium is homogeneous and isotropic;

[0045] The corresponding relationship between the occurrence level of tunnel water inrush disaster and the predicted water inflow is as follows: Grade I, Q > 5 (10 2 m 3 / d), extremely high possibility: Grade II, 1.5 < Q < 5 (10 2 m 3 / d), high possibility: Grade III, 0.25 < Q < 1.5 (10 2 m 3 / d), medium possibility: Grade IV, Q < 0.25 (10 2 m 3 / d), low possibility.

[0046] Further, in step 3), the established data repository should contain the information in step 1), and screen and clean the existing information, and eliminate the data with a relative error of more than 50% in the fitting result with the actual situation. When using the data repository to complement the geological information of the study area, it should be further ensured that the proportion of missing data in the total data should be less than 25% to avoid the occurrence of data distortion.

[0047] Further, in step 4), the fitting coefficient between the corrected GMS three-dimensional data model and the actual situation should be kept above 0.9 to ensure the accuracy of the filled data.

[0048] Further, in step 5), the selected model is any one of SVM, AdaBoost, and BP.

Claims

1. A method for calculating water inflow in tunnel construction based on machine learning, characterized in that: The method comprises the following specific steps: 1) Collect geological environment data in the surrounding areas of tunnel projects, including drilling data, surface survey data, underground rock formation data, and hydrological information, and import them into the heterogeneous water gushing disaster big data repository; 2) Preliminarily determine the water inflow per unit length and the maximum water inflow of the tunnel in the calculation area, and use the water inflow per unit length to determine the actual level of the water inflow disaster, and determine the danger level of the water inflow disaster based on the maximum water inflow per unit length; 3) Import the multi-source exploration information of the research section into a heterogeneous data repository built on the Hadoop system. According to the fitting results of the field data of the section and the field data of the previous water inrush accident area in the original database, fill the data gaps in the prediction section and eliminate abnormal data; 4) Input the data in the database into the simulation software Groundwater Model System (GMS) for 3D geological modeling, and make supplementary corrections to the infill data based on the difference between the modeling and the actual situation to complete the fitting update of the database; 5) Using the negative terrain type, stratum lithology, adverse geological type, structural surface type and development degree, karst development degree, groundwater storage or water inflow type, rainfall, construction disturbance degree, rock formation inclination and rock formation inclination, permeability coefficient, groundwater level, horizontal and vertical distance of groundwater body from tunnel, tunnel length and tunnel depth, and groundwater head information in the database as input information of the prediction model, the initial model is iteratively updated, so that the predicted water inflow per unit length of the tunnel output by the model and the actual water inflow per unit length of previous water inflow disasters have a fitting degree of more than 85%, and the training of the neural network model is completed; 6) Collect the above data of the study area and input them into the trained neural network model to output the water inflow per unit length of the tunnel, and use the water inflow per unit length to determine the actual level of water inflow disaster.

2. The method for calculating water inflow in tunnel construction based on machine learning according to claim 1, characterized in that: In the step 1), when collecting engineering example data, a construction tunnel with complete preliminary exploration data is selected, the missing rate of exploration information is less than 30%, and the data of the surrounding area at least contains information on tunnel water pressure, surrounding rock permeability coefficient, surrounding rock fragmentation, structural surface type and development degree, tunnel length and burial depth.

3. The method for calculating water inflow in tunnel construction based on machine learning according to claim 1, characterized in that: In the step 2), the maximum water inflow per unit length of the tunnel is calculated as follows: Where q0 is the maximum water inflow per unit length of the tunnel (m 3 / d); K is the surrounding rock permeability coefficient (m / d); m is the correction coefficient, which is 0.86; h is the distance from the initial groundwater level to the tunnel floor (m); r0 is the radius of the equivalent circle of the tunnel cross section (m); d is the diameter of the equivalent circle of the tunnel cross section (m); The corresponding relationship between the tunnel water inrush hazard level and the predicted maximum water inrush is: Level I, q0>1(10 3 m 3 / d), extremely high risk, catastrophic consequences: Level II, 0.3<q0<1(10 3 m 3 / d), high risk, serious harm: Level III, 0.05<q0<0.3(10 3 m 3 / d), medium risk, construction delay: IV level, q0<0.05(10 3 m 3 / d), low risk; The water inflow per unit length of the tunnel is calculated as follows: Where Q is the water inflow per unit length of the tunnel (m / d); H is the height from the groundwater level to the centerline of the tunnel (m); r is the radius of the tunnel (m); K is the permeability coefficient of the surrounding rock (m / d). This formula assumes that the initial groundwater level is horizontal and constant, the groundwater is in a stable state and satisfies radial flow, and the medium is homogeneous and isotropic. The corresponding relationship between the level of tunnel water inrush disaster and the water inrush per unit length of the tunnel is: Level I, Q>5(10 2 m 3 / d), very high probability: Level II, 1.5<Q<5(10 2 m 3 / d), high probability: Grade III, 0.25<Q<1.5(10 2 m 3 / d), moderate possibility: IV level, Q<0.25(10 2 m 3 / d), low probability.

4. The method for calculating water inflow in tunnel construction based on machine learning according to claim 1, characterized in that: In the step 3), the constructed data repository contains the data information in step 1), and the existing information is screened and cleaned, and the data with a relative error of more than 50% with the actual fitting result is eliminated. When using the data repository to complete the geological information of the study area, ensure that the proportion of missing data in the total data is less than 25%.

5. The method for calculating water inflow in tunnel construction based on machine learning according to claim 1, characterized in that: In the step 4), the fitting coefficient between the corrected GMS three-dimensional data model and the actual situation is maintained above 0.

9.

6. The method for calculating water inflow in tunnel construction based on machine learning according to claim 1, characterized in that: In the step 5), the neural model selected is any one of SVM, AdaBoost, and BP.

Images