A method for predicting water inflow of tunnel excavation by horizontal drilling

Abstract

The present application provides a method for predicting the water inflow of tunnel excavation by horizontal drilling, relating to the technical field of water inflow prediction in tunnel construction, comprising: obtaining water inflow time series data and geological parameter data; using the Goodman formula based on rock mass integrity correction to calculate the permeability coefficient of each monitoring point according to the water inflow time series data; decomposing the water inflow time series data to obtain a plurality of intrinsic mode functions and a residual term; forming a training data set; constructing a prediction model containing a GRU neural network structure and a physical constraint mechanism, training the prediction model using the training data set until the loss function converges, and obtaining the trained prediction model; obtaining the current monitoring data and inputting it into the trained prediction model to predict the water inflow and water inflow risk level of the to-be-excavated section. The present application can realize fine and high-precision prediction of water inflow during tunnel excavation, avoiding safety hazards caused by coarse-grained prediction of traditional methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel construction water inflow prediction, and particularly relates to a method for predicting tunnel excavation water inflow by horizontal drilling. BACKGROUND

[0002] Tunnel water inrush disaster is one of the main geological disasters threatening the safety of tunnel engineering, which has a significant impact on engineering construction safety. Sudden water inrush may not only cause casualties and damage to mechanical equipment, but also cause a series of chain problems such as delay of construction period and cost overrun. The existing tunnel water inflow prediction methods mainly include empirical formula method, numerical simulation method and statistical analysis method. These methods usually rely on vertical drilling and geophysical prospecting means to obtain geological parameters. However, the empirical formula method relies too much on empirical parameters and is difficult to accurately reflect the hydrological dynamic change law under complex geological conditions; the numerical simulation method has a solid theoretical basis, but the calculation amount is huge and the key parameters are difficult to obtain, and the practicality is limited; and the statistical analysis method is difficult to achieve effective early warning of sudden water inrush events due to low prediction accuracy, which has obvious technical limitations.

[0003] Chinese patent 202111449519.4 discloses a tunnel excavation water inflow prediction method based on horizontal directional drilling exploration hole, which attempts to solve some defects of traditional prediction methods. However, this method still has technical bottlenecks: it uses formula method to segmentally predict tunnel water inflow, and uses the distance between vertical drill holes as the segmentation basis, but the vertical drill hole spacing is usually 200-500 meters or even larger. Considering that tunnel construction generally adopts the combination of drill and blast method and TBM construction, the single excavation cycle footage is about 2-3 meters, so long prediction segmentation will inevitably lead to coarse prediction result granularity and insufficient accuracy, which cannot timely warn local water inrush risk and thus brings potential hidden dangers to construction safety.

[0004] In summary, the existing technology mainly has three technical defects: first, the traditional vertical drilling monitoring method is difficult to continuously and accurately reflect the change characteristics of the tunnel axial hydrogeological conditions; second, most prediction models use empirical formula method, which lacks real-time dynamic monitoring data support and cannot adapt to the complex changes of geological conditions; third, the prediction results are often too macro and strong in integrity, which cannot realize high-precision local segmentation prediction and cannot meet the actual needs of tunnel construction safety. Therefore, it is urgent to develop a new type of tunnel water inflow prediction method that can realize fine prediction. SUMMARY

[0005] Therefore, the present application provides a method for predicting water inflow in tunnel excavation by horizontal drilling, which realizes fine and high-precision prediction of water inflow in the tunnel excavation process, avoids the safety hazards caused by the coarse-grained prediction of the traditional method, provides a forward-looking early warning for water inflow risks in the construction process, and effectively improves the safety and efficiency of tunnel construction.

[0006] The technical scheme of the present application is implemented as follows:

[0007] The present application provides a method for predicting water inflow in tunnel excavation by horizontal drilling, which comprises:

[0008] S1, segmentally arranging horizontal directional drilling along the tunnel axis and installing a monitoring system to obtain time series data of water inflow and geological parameter data of each monitoring point;

[0009] S2, using the Goodman formula based on rock mass integrity correction to calculate the permeability coefficient of the location of each monitoring point according to the time series data of water inflow;

[0010] S3, using the complete ensemble empirical mode decomposition method to decompose the time series data of water inflow to obtain a plurality of intrinsic mode functions and a residual term;

[0011] S4, forming a training data set based on the intrinsic mode function, the residual term, the permeability coefficient and the geological parameter data;

[0012] S5, constructing a prediction model containing a GRU neural network structure and a physical constraint mechanism, training the prediction model using the training data set until the loss function converges, and obtaining the trained prediction model;

[0013] S6, obtaining current monitoring data and inputting the trained prediction model to predict the water inflow and water inflow risk level of the to-be-excavated section.

[0014] Preferably, in step S1, segmentally arranging horizontal directional drilling and installing a monitoring system comprises:

[0015] Using a geological radar to conduct advanced detection of the geology in front of the tunnel to identify potential high-risk water inflow areas;

[0016] Setting a monitoring point every 15-20 meters of drilling in ordinary sections and a monitoring point every 5-10 meters of drilling in high-risk sections;

[0017] Deploying a high-precision flow sensor and a data transmission device at each monitoring point.

[0018] Preferably, the geological parameter data includes: groundwater level, drilling diameter, water-bearing body length, rock mass quality index RQD, fault fracture zone width, karst development degree and environmental impact factors.

[0019] Preferably, step S3 comprises:

[0020] S31, adding white noise with different amplitudes to the water inflow time series data to generate multiple groups of data;

[0021] S32, performing empirical mode decomposition on each group of data after adding noise;

[0022] S33, averaging the decomposition results of all groups to obtain the intrinsic mode function and residual term.

[0023] Preferably, the amplitude of white noise in step S31 ranges from 0.1 to 0.3 times the standard deviation of the original data, and the number of noise groups is 50-100.

[0024] Preferably, the Goodmann formula based on rock mass integrity correction in step S2 is:

[0025]

[0026] In the formula, K is the permeability coefficient; Q is the water inflow time series data; L is the length of the horizontal borehole through the water-bearing body; H is the water head difference, representing the vertical height difference between the groundwater level and the tunnel drainage surface; d is the cross-sectional diameter of the horizontal borehole; and a(RQD) is the correction coefficient based on rock mass integrity, which is calculated according to the rock mass quality index RQD, and the formula is as follows:

[0027]

[0028] In the formula, C1 is the basic correction value, which ranges from 0.15 to 0.25; C2 is the gain coefficient, which ranges from 0.75 to 0.85; k is the curve steepness coefficient, which controls the sensitivity of the correction coefficient to RQD, and ranges from 0.04 to 0.06; and R0 is the inflection point parameter, which represents the critical value of the change of the correction coefficient, and ranges from 45 to 55.

[0029] Preferably, the prediction model comprises:

[0030] The feature preprocessing layer combines the intrinsic mode function and the residual term to form time series features, and combines the permeability coefficient and the geological parameters to form static features;

[0031] The feature encoding layer extracts features from the time series features and the static features, and concatenates them in the feature dimension to form a fusion feature vector;

[0032] The time series modeling layer includes multiple stacked bidirectional GRU units for capturing time series dependencies, and the output of each GRU unit is weighted by an attention mechanism to obtain a context feature representation;

[0033] The prediction layer maps the context feature representation to the water inflow prediction value through a fully connected network.

[0034] The risk assessment layer calculates the ratio of the water inrush prediction value to the safety threshold value, and divides the water inrush risk level according to the ratio size.

[0035] Preferably, the loss function is:

[0036] L total =L MSE +λ1(N)·L physical +λ2(N)·L temporal

[0037]

[0038] In the formula, L total is the total loss function; L MSE is the mean square error loss; L physical is the physical constraint loss; L temporal is the time continuity loss; m is the sample number; y i is the actual water inrush of the i-th sample; is the predicted water inrush of the i-th sample; A tunnel is the cross-sectional area of the tunnel; K i is the permeability coefficient of the i-th sample corresponding to the monitoring point; is the hydraulic gradient of the i-th sample corresponding to the monitoring point, indicating the change of water head per unit distance; Δt i represents the time interval between two samples i and i+1; λ1(N) and λ2(N) are weight coefficients that change with the cumulative data amount N.

[0039] Preferably, step S4 comprises:

[0040] S41, the intrinsic mode function and the residual term are taken as time series features, and the permeability coefficient and the geological parameter data are taken as static features;

[0041] S42, for each monitoring point, the time series features of continuous T time points and the corresponding static features are selected to form an input feature vector; the water inrush amount of T+1 time points is taken as an output prediction target;

[0042] S43, a sliding window method is used, the window length is M, the sliding step is 1, and a plurality of input-output sample pairs are sequentially constructed;

[0043] S44, the sample pairs are randomly divided into a training set, a validation set and a test set to form a training data set.

[0044] Preferably, the determination of the water inrush risk level in step S6 comprises:

[0045] The ratio of the predicted water inrush amount to the safety threshold value is calculated to obtain a risk index P;

[0046] Risk levels are divided according to the risk index P: P<0.3 is low risk, 0.3≤P<0.6 is medium risk, 0.6≤P<0.8 is high risk, and P≥0.8 is extremely high risk.

[0047] The present application has the following beneficial effects over the prior art:

[0048] (1) The present application realizes high-precision prediction of water inflow in tunnel excavation by segmenting horizontal directional drilling along the tunnel axis and installing a monitoring system, processing time series data of water inflow by a complete ensemble empirical mode decomposition method, and using a prediction model that fuses GRU neural network and physical constraint mechanism. Compared with traditional vertical drilling and empirical formula prediction methods, the present application can capture the continuous changes of axial hydrogeological conditions of the tunnel and provide more refined prediction results, providing prospective early warning of water inflow risk in the tunnel construction process;

[0049] (2) The segmented horizontal directional drilling strategy adopted by the present application sets a monitoring point every 15-20 meters in ordinary sections and every 5-10 meters in high-risk sections, breaking through the 200-500 meter spacing limit of traditional vertical drilling. This intensive layout makes the distribution of monitoring points more reasonable, accurately reflects the spatial variation characteristics of the hydrogeological conditions along the tunnel, and matches the monitoring granularity with the tunnel construction single-cycle footage, effectively improving the identification ability of the prediction model for local small-scale water inflow risk;

[0050] (3) The present application introduces the Goodmann formula based on rock mass integrity correction to calculate the permeability coefficient, and modifies the permeability characteristics under different rock mass quality grades through the alpha(RQD) correction coefficient. This correction coefficient is constructed using a Sigmoid function, which can accurately describe the nonlinear characteristics of the permeability coefficient with the change of rock mass integrity, effectively solving the problem of poor adaptability of traditional empirical formulas to different integrity rock masses;

[0051] (4) The present application uses the complete ensemble empirical mode decomposition method to process time series data of water inflow, adds white noise with an amplitude of 0.1-0.3 times the standard deviation to the original data and generates 50-100 groups of data, and takes the average after empirical mode decomposition of each group of data, effectively overcoming the mode mixing and endpoint effect problems existing in the traditional EMD method;

[0052] (5) The prediction model constructed by the application adopts a hierarchical network structure, extracts time sequence features and static features through a feature preprocessing layer respectively, and realizes effective fusion of multi-source heterogeneous data by combining a feature encoding layer and a time sequence modeling layer. In particular, the time sequence modeling layer adopts a bidirectional GRU structure with an attention mechanism, which can simultaneously capture the forward and backward dependency relationships of the water inflow time sequence data, and adaptively give different weights to the features of different time points, effectively suppressing noise interference while retaining long-term memory. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0054] Figure 1 The method flowchart of the present application is shown in the figure.

[0055] Figure 2 The technical implementation diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described in detail below. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] As shown in the figure, the present application provides a method for predicting water inflow of tunnel excavation by horizontal drilling, comprising: Figure 1

[0058] S1, segmentally arranging horizontal directional drilling along the tunnel axis and installing a monitoring system to obtain water inflow time sequence data and geological parameter data of each monitoring point;

[0059] S2, using the Goodman formula based on rock mass integrity correction to calculate the permeability coefficient of the position of each monitoring point according to the water inflow time sequence data;

[0060] S3, decomposing the water inflow time sequence data by using the complete ensemble empirical mode decomposition method to obtain a plurality of intrinsic mode functions and a residual term;

[0061] S4, forming a training data set based on the intrinsic mode function, the residual term, the permeability coefficient and the geological parameter data;

[0062] ​S5, a prediction model containing a GRU neural network structure and a physical constraint mechanism is constructed, and the prediction model is trained using a training data set until the loss function converges, to obtain a trained prediction model;

[0063] S6, current monitoring data is obtained and input into the trained prediction model to predict the water inflow and water inflow risk level of the to-be-excavated section.

[0064] As shown in Figure 2 The technical implementation process of the present application is as follows: first, segmentally arrange horizontal directional drilling along the tunnel axis and install a monitoring system, set a monitoring point every 15-20 meters in ordinary sections and every 5-10 meters in high-risk sections to obtain water inflow time series data and geological parameter data; second, calculate the permeability coefficient using the Goodman formula based on rock mass integrity correction, and introduce the correction coefficient α(RQD) constructed by the Sigmoid function to finely adjust different integrity rock masses; then, process the water inflow time series data using the complete set empirical mode decomposition method, generate multiple groups of data by adding white noise and take the average to overcome the mode mixing problem of traditional EMD, and obtain stable intrinsic mode functions and residual terms; next, the intrinsic mode functions and residual terms are taken as time series features, and the permeability coefficient and geological parameters are taken as static features to construct a training data set; subsequently, a prediction model combining a GRU neural network structure and a physical constraint mechanism is designed, and a composite loss function containing mean square error loss, physical constraint loss based on Darcy's law and time continuity loss is used for training; finally, the current monitoring data is input into the trained model to predict the water inflow of the to-be-excavated section, and the risk level is divided according to the ratio of the prediction result to the safety threshold, providing accurate early warning for tunnel construction.

[0065] Specifically, in an embodiment of the present application, in step S1, segmentally arranging horizontal directional drilling and installing a monitoring system includes:

[0066] Using a geological radar to conduct advanced detection of the geology in front of the tunnel to identify potential high-risk water inflow areas;

[0067] In ordinary sections, a monitoring point is set every 15-20 meters drilled, and in high-risk sections, a monitoring point is set every 5-10 meters drilled;

[0068] A high-precision flow sensor and a data transmission device are deployed at each monitoring point.

[0069] In this embodiment, first, geological radar is used for advanced detection of the geology in front of the tunnel, identifying potential high-risk water gushing areas such as fault fracture zones, karst development areas, etc. Geological radar can effectively detect geological structure anomalies in front of the tunnel by emitting electromagnetic waves and receiving reflected signals, providing guidance for subsequent borehole layout. After obtaining the geological radar detection results, horizontal directional drilling is carried out along the tunnel axis. The borehole has a certain slope to allow groundwater to flow out naturally, facilitating water gushing monitoring. The borehole layout adopts a differentiated strategy: in ordinary sections, a monitoring point is set every 15-20 meters drilled, and in high-risk areas (such as fault fracture zones, karst development areas, etc. determined by geological radar detection), a monitoring point is set every 5-10 meters drilled, forming a reasonable monitoring network layout. For hard rock strata areas, drilling can be implemented synchronously with TBM tunneling machinery to improve construction efficiency. At each monitoring point, high-precision flow sensors and data transmission devices are deployed. The flow sensor is used to accurately measure the water gushing volume of the borehole, and the data transmission device transmits monitoring data in real time to the data processing center. The monitoring system uses wireless transmission, and each monitoring point is equipped with an RFID tag for positioning and identification.

[0070] Specifically, the depth and flow information of each monitoring point are recorded synchronously, and geological parameter data are collected, including:

[0071] Groundwater level, obtained by a water level gauge in the borehole, represents the vertical height difference between the groundwater level and the tunnel drainage surface.

[0072] Borehole diameter, the cross-sectional diameter of the horizontal directional drilling, determined according to the drill bit specifications used during drilling.

[0073] Aquifer length, the length of the horizontal borehole passing through the aquifer, obtained through geological records and electrical logging techniques during drilling.

[0074] Rock Mass Quality Index (RQD), obtained through borehole core sampling analysis, representing the degree of rock mass integrity, calculated as the percentage of the cumulative length of core segments longer than 10 cm to the total drilling length.

[0075] Fault fracture zone width, determined by geological radar detection and the presence of fault fracture zones during drilling, recording the width of the fault fracture zone.

[0076] Karst development degree, evaluated through core analysis during drilling, geological radar detection, and hydrogeological investigation, etc., and can be classified into weak, medium, and strong levels.

[0077] Environmental impact factors, including rainfall, atmospheric pressure changes, surrounding construction activities, and other external factors that may affect water gushing volume, obtained through weather stations, environmental monitoring equipment, etc.

[0078] At the same time, the water inflow time series data, i.e., the water inflow value sequence measured at different time points of each monitoring point, is collected.

[0079] All data is marked with a timestamp to establish a unified space-time database.

[0080] Specifically, in an embodiment of the present application, step S2 calculates the permeability coefficient of each section according to the water inflow of each section of the borehole, and uses a modified Goodman formula for derivation:

[0081]

[0082] In the formula, K is the permeability coefficient; Q is the water inflow time series data; L is the length of the horizontal borehole passing through the water-bearing body; H is the water head difference, representing the vertical height difference between the groundwater level and the tunnel drainage surface; d is the diameter of the horizontal borehole cross section; and a(RQD) is a correction coefficient based on rock mass integrity, which is calculated according to the rock mass quality index RQD, and the formula is as follows:

[0083]

[0084] In the formula, C1 is a basic correction value, with a value range of 0.15-0.25; C2 is a gain coefficient, with a value range of 0.75-0.85; k is a curve steepness coefficient, controlling the sensitivity of the correction coefficient to RQD, with a value range of 0.04-0.06; and R0 is a turning point parameter, representing the critical value of the correction coefficient change, with a value range of 45-55.

[0085] The correction coefficient a(RQD) constructed by introducing the Sigmoid function in the present embodiment can accurately describe the nonlinear characteristics of the permeability coefficient changing with the rock mass integrity. When the RQD value is low, it indicates that the rock mass is highly broken, the correction coefficient is close to C1+C2, and the permeability coefficient is large; when the RQD value is high, it indicates that the rock mass is highly complete, the correction coefficient is close to C1, and the permeability coefficient is small. This correction method effectively solves the adaptability problem of the traditional Goodman formula in different integrity rock masses.

[0086] In the actual calculation process, first, the water inflow time series data Q, the water-bearing body length L, the water head difference H, the borehole diameter d, and the rock mass quality index RQD are extracted from the monitoring data obtained from S1. Then, the correction coefficient a(RQD) is calculated according to the rock mass quality index RQD, and finally, the permeability coefficient K of the position where each monitoring point is located is calculated by substituting into the corrected Goodman formula.

[0087] Specifically, in an embodiment of the present application, step S3 includes:

[0088] S31, adding white noise of different amplitudes to the water inflow time series data to generate multiple groups of data.

[0089] Specifically, white noise with different amplitudes is added to the original water inflow time series data X(t) to generate multiple sets of noise superimposed data. The amplitude of the white noise ranges from 0.1 to 0.3 times the standard deviation of the original data, and the number of noise sets is 50-100 sets. The formula is as follows:

[0090] X j (t)=X(t)+ε j ·σ·N j (t),j=1,2,...,M

[0091] In the formula, X j (t) is the data after adding noise to the jth set; X(t) is the original water inflow time series data; ε j is the amplitude coefficient, with a value range of 0.1-0.3; σ is the standard deviation of the original data; N j (t) is the jth set of standard white noise sequence (mean is 0, standard deviation is 1); M is the number of noise sets, with a value range of 50-100 sets.

[0092] S32, the data after adding noise to each set is empirically mode decomposed.

[0093] Specifically, the data X j (t) after adding noise to each set is empirically mode decomposed to a series of intrinsic mode functions (IMF) and a residual term.

[0094] The basic idea of empirical mode decomposition is to decompose a complex time series signal into a series of intrinsic mode functions with different characteristic scales, each of which represents an oscillation mode in the original signal. The decomposition process is realized through an iterative "sifting" process, the steps of which are as follows:

[0095] 1. Identify all local maximum points and minimum points in X j (t);

[0096] 2. Fit the upper envelope line through the maximum points and the lower envelope line through the minimum points;

[0097] 3. Calculate the average of the upper and lower envelope lines, denoted as m(t);

[0098] 4. Calculate the difference between X j (t) and m(t) to obtain h(t) = X j (t)-m(t);

[0099] 5. Determine whether h(t) satisfies the two conditions of IMF: (a) the number of maximum points and minimum points differs by no more than 1; (b) the local mean is zero;

[0100] 6. If h(t) does not satisfy the IMF condition, take h(t) as a new signal and repeat steps 1-5 until the IMF condition is satisfied;

[0101] 7. Define h(t) satisfying the condition as the first IMF component, denoted as IMF1;

[0102] 8. Calculate the residual r(t) = X j (t) - IMF1;

[0103] 9. Take r(t) as a new input signal and repeat the above steps to obtain IMF2, IMF3,..., until the residual function becomes a monotonic function or its amplitude is less than a preset threshold.

[0104] For each set of data X j (t), after the above decomposition, the following is obtained:

[0105]

[0106] wherein, is the ith intrinsic mode function obtained by decomposing the jth set of data; r j (t) is the residual term of the jth set of data; N j is the number of intrinsic mode functions obtained by decomposing the jth set of data.

[0107] S33, average the decomposition results of all groups to obtain the intrinsic mode functions and the residual term.

[0108] Specifically:

[0109]

[0110] wherein, IMF i (t) is the final ith intrinsic mode function; r(t) is the final residual term; and N is the number of intrinsic mode functions, and takes a value of 5-7.

[0111] Specifically, in an embodiment of the present application, step S4 comprises:

[0112] S41, take the intrinsic mode functions and the residual term as time sequence features, and take the permeability coefficient and the geological parameter data as static features; the time sequence features usually include 5-7 intrinsic mode functions (IMF1, IMF2,..., IMF n) and 1 residual term (r), each of which is a time-varying sequence. These features reflect the variation of water inflow at different time scales, where high-frequency IMFs reflect short-term fluctuations, low-frequency IMFs reflect medium-term changes, and the residual term reflects long-term trends. Static features include the permeability coefficient K and geological parameter data (groundwater level height, borehole diameter, water-bearing body length, rock mass quality index RQD, fault fracture zone width, karst development degree, and environmental impact factors), which are relatively stable at each monitoring point location and reflect the spatial hydrogeological characteristics.

[0113] S42, for each monitoring point, select the time series features and corresponding static features of the consecutive T time points to form an input feature vector; take the water inflow at the T+1 time point as the output prediction target; specifically represented as:

[0114] Input feature vector: X i = [IMF1(t i-T+1 :t i ), IMF2(t i-T+1 :t i ), …, IMF n (t i-T+1 :t i ), r(t i-T+1 :t i ), K i , G i ];

[0115] Output prediction target: y i = Q(t i+1 ).

[0116] Where, IMF j (t i-T+1 :t i ) represents the value of the jth intrinsic mode function at t i-T+1 to t i ; r(t i-T+1 :t i ) represents the value of the residual term at t i-T+1 to t i ; K i represents the permeability coefficient of the ith monitoring point; G i represents the geological parameter data of the ith monitoring point; Q(t i+1 ) represents the actual water inflow at t i+1 .

[0117] S43, use the sliding window method, the window length is M, the sliding step is 1, and a plurality of input-output sample pairs are sequentially constructed; the process of the sliding window is as follows:

[0118] Firstly, the window is placed at the start position of the time series to construct the first sample pair; then the window is slid forward by one unit to construct the second sample pair; this is repeated until the window reaches the end of the time series. This method can make full use of time series data to generate a large number of training samples from limited monitoring data. For example, assuming that a monitoring point has 1500 time point data, T = 24, M = 1440, then 1500-24-1+1 = 1476 sample pairs can be constructed. Each sample pair contains 24 time point input features and 1 time point output target.

[0119] S44, randomly divide the sample pairs into a training set, a validation set and a test set to form a training data set.

[0120] The division ratio is: training set 70%, validation set 15%, and test set 15%.

[0121] Specifically, in an embodiment of the present application, the prediction model comprises:

[0122] The feature preprocessing layer combines the intrinsic mode function and the residual term to form the time series feature, and combines the permeability coefficient and the geological parameters to form the static feature; specifically, the obtained multiple intrinsic mode functions (IMF1, IMF2, …, IMF n ) and a residual term (r) are combined to form a time series feature matrix with a dimension of T x (n+1), where T is the length of the time window, and n+1 is the number of features; the calculated permeability coefficient K and the obtained geological parameter data (groundwater level, borehole diameter, water-bearing body length, rock mass quality index RQD, fault fracture zone width, karst development degree and environmental impact factors) are combined to form a static feature vector; all features are standardized to eliminate the dimension effect.

[0123] The feature encoding layer extracts features from the time series features and the static features, and concatenates them in the feature dimension to form a fusion feature vector.

[0124] The feature encoding layer independently encodes different types of features, and then fuses them to form a unified representation:

[0125] For the time series features, a one-dimensional convolution network (Conv1D) is used for feature extraction, with a convolution kernel size of 3 and a step size of 1 to extract local time series patterns; for the static features, a fully connected network is used for feature mapping to capture the nonlinear relationship between the features; the processed time series features and static features are concatenated in the feature dimension to form a fusion feature vector, which is used as the input of the time series modeling layer.

[0126] The time series modeling layer includes multiple stacked bidirectional GRU units for capturing time series dependencies, and the output of each GRU unit is weighted by an attention mechanism to obtain a context feature representation.

[0127] Specifically, the temporal modeling layer contains 3-4 stacked bidirectional GRU units, each GRU unit containing 64-128 hidden neurons; the bidirectional GRU can simultaneously consider past and future information, enhancing the ability to capture temporal patterns; the computation process of each GRU unit is as follows:

[0128] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0129] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0130]

[0131] Among them, z t To update the gate, control the degree to which previous state information is retained; r t To reset the gate, control the influence of the previous state on the current candidate state; For candidate hidden state; h t Currently hidden; x t Input features include hydrological parameters, geological parameters, etc.; W z W r W represents the network weight matrix; b z b r b are bias terms.

[0132] The outputs of each GRU unit are weighted using an attention mechanism to highlight the impact of important time points:

[0133] Attention weight calculation: α t =softmax(v T ·tanh(W a ·h t +b a ));

[0134] Weighted context vector: c = ∑α t ·h t ;

[0135] The final result is a contextual feature representation containing rich temporal information.

[0136] The prediction layer maps contextual feature representations to predicted water flow rates through a fully connected network.

[0137] The prediction layer maps contextual feature representations to predicted inflow rates:

[0138] The model contains 2-3 fully connected layers, with the number of hidden layer neurons decreasing in turn (e.g. 128→64→32); the activation function uses ReLU to avoid the problem of gradient disappearance; the output layer of the latter is a single neuron, representing the predicted water inflow value; the output layer does not use an activation function, and directly outputs the linear combination result.

[0139] The risk assessment layer calculates the ratio of the predicted water inflow value to the safety threshold, and divides the water inflow risk level according to the ratio size.

[0140] The risk assessment layer calculates the ratio of the predicted water inflow value to the safety threshold, and divides the risk level according to the ratio size:

[0141] The safety threshold is determined according to the tunnel construction specification and the site conditions; the risk index P = predicted water inflow / safety threshold; the risk level is divided as follows: P < 0.3 is low risk, 0.3≤P < 0.6 is medium risk, 0.6≤P < 0.8 is high risk, and P≥0.8 is extremely high risk.

[0142] After the prediction model is constructed, it is trained using the training data set. The loss function during training introduces a physical constraint mechanism, so that the model prediction result not only conforms to the data law but also meets the physical law. The loss function is:

[0143] L total =L MSE +λ1(N)·L physical +λ2(N)·L temporal

[0144]

[0145] In the formula, L total is the total loss function; L MSE is the mean square error loss; L physical is the physical constraint loss; L temporal is the time continuity loss; m is the sample number; y i is the actual water inflow of the i-th sample; is the predicted water inflow of the i-th sample; A tunnel is the cross-sectional area of the tunnel; K i is the permeability coefficient of the monitoring point corresponding to the i-th sample; is the hydraulic gradient of the monitoring point corresponding to the i-th sample, representing the change of water head per unit distance; Δt i represents the time interval between two samples i and i+1; λ1(N) and λ2(N) are weight coefficients that change with the cumulative data amount N. The weight coefficient is calculated as follows:

[0146] λ i (N)=λ i0·e -γN

[0147] wherein, λ i0 is the initial weight; γ is the attenuation coefficient, and N is the cumulative data amount. When the data amount is small, the physical constraint weight is large, and as the data accumulation increases, the constraint weight gradually decreases, enhancing the fitting ability of the model to the measured data.

[0148] The specific training process of the model is as follows:

[0149] 1. Initialization: randomly initialize the model parameters, set the learning rate, batch size, training round number and other hyperparameters;

[0150] 2. Forward propagation: input data into the model to obtain the prediction result through each layer calculation;

[0151] 3. Loss calculation: calculate the loss value under the current parameters according to the above composite loss function;

[0152] 4. Back propagation: calculate the gradient of the loss function with respect to each parameter, and update the parameters using the optimizer; the optimizer uses the Adam algorithm, and the initial learning rate is set to 0.001; the learning rate decay strategy is adopted, and the learning rate is decayed to 0.9 times of the original every 10 training rounds;

[0153] 5. Early stopping strategy: monitor the model performance on the validation set, and stop training when the validation set loss does not decrease for 10 consecutive rounds;

[0154] 6. Model saving: save the model parameters with the best validation set performance.

[0155] During the training process, the change trend of the total loss function and its three components is monitored to ensure that the model learns both the statistical rules in the data and the hydrological physical laws. When the loss function converges to a stable value or reaches the preset maximum training round number, the training is completed, and the trained prediction model is obtained.

[0156] Specifically, in an embodiment of the present application, step S6 comprises:

[0157] First, the latest water inflow time series data and geological parameter data are collected from the horizontal directional drilling monitoring system in real time; then the water inflow time series data is subjected to complete ensemble empirical mode decomposition to obtain intrinsic mode functions and residual terms; then, combined with the calculated permeability coefficient, these features are arranged into an input vector according to the format during training; subsequently, the input vector is input into the trained GRU neural network prediction model to obtain the water inflow prediction value at the future time point; the risk index P is calculated by taking the ratio of the predicted water inflow to the safety threshold, and the water inflow risk level is divided according to the threshold (P<0.3 is low risk, 0.3≤P<0.6 is medium risk, 0.6≤P<0.8 is high risk, and P≥0.8 is extremely high risk).

[0158] Specifically, in another embodiment of the present application, after the prediction model is trained, the trained GRU prediction model is compressed and accelerated using the TensorRT optimization technology to achieve efficient deployment on edge computing devices. The specific implementation is as follows:

[0159] Convert the FP32 (32-bit floating point) parameters of the model to INT8 (8-bit integer) or FP16 (16-bit floating point) format, achieve parameter precision reduction and performance improvement, and reduce the model size by 50%-75%. During the quantization process, the calibration dataset is used to ensure that the accuracy loss is controlled within 2%. Through vertical and horizontal layer fusion technology, multiple adjacent network layers (such as convolutional layers, batch normalization layers, activation layers) are merged into a single calculation unit, reducing intermediate result storage and memory access times, and reducing calculation delay. According to the deployment hardware characteristics (such as GPU model, memory bandwidth), the optimal CUDA kernel implementation is automatically selected to maximize hardware resource utilization.

[0160] The lightened model is deployed on the edge computing device. The model deployment adopts a microservice architecture, including the following components: data acquisition service, responsible for obtaining real-time data from the monitoring system; feature processing service, responsible for data preprocessing and feature extraction; inference service, loading the optimized model to perform prediction calculation; result distribution service, transmitting the prediction result to the construction control system.

[0161] This embodiment also designs an incremental learning strategy, so that the model can continuously learn from newly collected data and continuously improve the prediction accuracy. The specific implementation is as follows:

[0162] Incremental learning is triggered by the following two conditions:

[0163] Periodic trigger: incremental learning is triggered once automatically after each excavation cycle (usually 2-3 meters);

[0164] Error trigger: when the prediction error exceeds the preset threshold ε (ε = 15%) for N consecutive times (N = 5), trigger incremental learning.

[0165] A sliding time window mechanism is used to select data for incremental learning. The window size W is defined as the monitoring records of the last M time points (usually M = 1440, corresponding to 1 day of minute-level data). Each time incremental learning is triggered, the window slides forward, and the newly collected data is included in the training set, while the oldest part of the data is removed. The sliding window dataset is represented as:

[0166] D window = (X t , y t )|t∈[t current-M+1 , t current ]

[0167] In the formula, X t represents the input feature vector at time t; y t represents the actual water inflow at time t; t current represents the current time point.

[0168] The incremental learning updates the model parameters by the gradient descent method, and the mathematical expression is:

[0169]

[0170] In the formula, W t+1 is the updated network weight parameter set; W t is the current network weight parameter set; η t is a dynamic learning rate that decays over time; is the loss function gradient calculated based on the window data. The dynamic learning rate η t adopts a time decay strategy, and the calculation formula is:

[0171]

[0172] In the formula, η0 is the initial learning rate, which is set to 0.001; δ is the decay coefficient, which is set to 0.0005; t is the cumulative number of incremental learning.

[0173] Specifically, the loss function L window is consistent with the composite loss function used in the model training stage, and also contains three parts of mean square error loss, physical constraint loss and time continuity loss.

[0174] In order to prevent the model from having a catastrophic forgetting problem (i.e. the model forgetting the previously learned knowledge) in the incremental learning process, the following protection mechanism is adopted:

[0175] Experience replay strategy: maintain an experience pool with a capacity of K (usually K = 5000) to store historical important samples. Each time incremental learning, randomly extract P (usually P = 1000) samples from the experience pool and mix them with new data for training.

[0176] Parameter regularization: add a regularization term to the incremental learning loss function to limit the deviation degree of new parameters from original parameters:

[0177] L regularized = L window + λ reg · ||W t+1 -W t || 2

[0178] Wherein, λ reg is a regularization coefficient, usually set to 0.01.

[0179] Knowledge distillation: the original model prediction result is used as a "soft label" to assist training, and a distillation term is added to the loss function:

[0180] L distill = alpha * L window + (1-alpha) * KL(f(X; W t ), f(X; W t+1 ))

[0181] Wherein, alpha is a balance coefficient, and is set to 0.7;KL represents KL divergence, which measures the difference between two distributions;f(X; W) represents the predicted output of the model with parameters W for input X.

[0182] In conclusion, the present application collects the water inflow time series data and geological parameter data through horizontal drilling, then uses CEEMD decomposition to extract multi-scale time series features, combines the permeability coefficient calculated by physics, and inputs into the designed prediction model for water inflow prediction and risk level assessment. Through edge computing, millisecond-level response is realized, and the incremental learning mechanism is used to make the model continuously adapt to the change of geological conditions, effectively solving the problems of data and physical model fragmentation, low prediction accuracy and poor adaptability in traditional methods.

[0183] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting inflow of water in tunnel excavation by horizontal drilling, characterized by, The method comprises the following steps: S1, segmentally arranging horizontal directional drilling along the tunnel axis and installing a monitoring system to obtain time series data of water inflow and geological parameter data of each monitoring point; S2, using the Goodmann formula based on rock mass integrity correction to calculate the permeability coefficient of the location of each monitoring point according to the time series data of water inflow; The Goodmann formula based on rock mass integrity correction is: wherein, is the permeability coefficient; is the water inflow time series data; is the length of the horizontal borehole through the water-bearing body; is the water head difference, representing the vertical height difference between the groundwater level and the tunnel drainage surface; is the horizontal borehole cross-sectional diameter; is a correction coefficient based on rock mass integrity, which is calculated according to the rock mass quality index RQD, with the formula as follows: In the formula, is a basic correction value, and its value range is 0.15-0.25; is a gain coefficient, and its value range is 0.75-0.85; is a curve steepness coefficient, and controls the sensitivity of the correction coefficient to the change of RQD, and its value range is 0.04-0.06; is an inflection point parameter, and represents a critical value of the change of the correction coefficient, and its value range is 45-55; S3, decomposing the time series data of water inflow by using the complete set empirical mode decomposition method to obtain multiple intrinsic mode functions and a residual term; S4, forming a training data set based on the intrinsic mode functions, the residual term, the permeability coefficient and the geological parameter data; S5, constructing a prediction model containing a GRU neural network structure and a physical constraint mechanism, training the prediction model using the training data set until the loss function converges, and obtaining the trained prediction model; S6, obtaining the current monitoring data and inputting the trained prediction model to predict the water inflow and water inflow risk level of the to-be-excavated section.

2. A method for predicting inflow of water from a tunnel excavation by horizontal drilling according to claim 1, wherein In step S1, segmentally arranging horizontal directional drilling and installing a monitoring system comprises: Using a geological radar to conduct advanced detection on the geology in front of the tunnel to identify potential high-risk water inflow areas; Setting a monitoring point every 15-20 meters in ordinary sections and every 5-10 meters in high-risk sections; Deploying a high-precision flow sensor and a data transmission device at each monitoring point.

3. A method for predicting inflow of water from a tunnel excavation by horizontal drilling according to claim 1, wherein The geological parameter data includes: underground water level, borehole diameter, water-bearing body length, rock mass quality index RQD, fault fracture zone width, karst development degree and environmental impact factors.

4. The method for predicting inflow of a tunnel excavation by horizontal drilling according to claim 1, wherein Step S3 includes: S31, adding white noise of different amplitudes to the time series data of water inflow to generate multiple groups of data; S32, performing empirical mode decomposition on each group of data after adding noise; S33, averaging the decomposition results of all groups to obtain the intrinsic mode functions and the residual term.

5. A method for predicting inflow of water from a tunnel excavation by horizontal drilling according to claim 4, wherein The amplitude range of the white noise in step S31 is 0.1-0.3 times the standard deviation of the original data, and the number of noise groups is 50-100 groups.

6. The method for predicting inflow of a tunnel excavation by horizontal drilling according to claim 1, wherein The prediction model comprises: The feature preprocessing layer combines the intrinsic mode functions and the residual term to form time series features, and combines the permeability coefficient and the geological parameters to form static features; The feature encoding layer extracts features from the time series features and the static features, and concatenates them in the feature dimension to form a fusion feature vector; The time series modeling layer includes multiple stacked bidirectional GRU units for capturing time series dependencies, and the output of each GRU unit is weighted by an attention mechanism to obtain a context feature representation; The prediction layer maps the context feature representation to a water inflow prediction value through a fully connected network; The risk assessment layer calculates the ratio of the water inflow prediction value to the safety threshold, and divides the water inflow risk level according to the ratio size.

7. A method for predicting inflow of water from a tunnel excavation by horizontal drilling according to claim 6, wherein The loss function is: wherein, is the total loss function; is the mean square error loss; is the physical constraint loss; is the time continuity loss; is the number of samples; is the actual inrush water volume of the i-th sample; is the predicted inrush water volume of the i-th sample; is the tunnel cross-sectional area; is the permeability coefficient of the monitoring point corresponding to the i-th sample; is the hydraulic gradient of the monitoring point corresponding to the i-th sample, indicating the change in water head per unit distance; represents the time interval between two samples i and i+1; and is the weight coefficient varying with the cumulative data amount N.

8. The method for predicting inflow of a tunnel excavation by horizontal drilling according to claim 1, wherein, Step S4 includes: S41, taking the intrinsic mode functions and the residual term as time series features, and taking the permeability coefficient and the geological parameter data as static features; S42, for each monitoring point, selecting time series features and corresponding static features at T consecutive time points to form an input feature vector; taking the water inflow at T+1 time point as the output prediction target; S43, using a sliding window method, a window length of M, a sliding step of 1, and sequentially constructing a plurality of input-output sample pairs; S44, randomly dividing the sample pairs into a training set, a validation set, and a test set to form a training data set.

9. The method for predicting inflow of a tunnel excavation by horizontal drilling according to claim 1, wherein, The determination of the water inrush risk level in step S6 includes: Calculating the ratio of the predicted water inrush amount to the safety threshold to obtain a risk index P; According to the risk index P, the risk level is divided: P<0.3 is low risk, 0.3≤P<0.6 is medium risk, 0.6≤P<0.8 is high risk, and P≥0.8 is extremely high risk.

Images