A dynamic intelligent early warning method for ground subsidence induced by tunnel boring machine (TBM)

Abstract

The present invention discloses an intelligent dynamic early warning method for ground subsidence induced by shield tunneling, comprising the following steps: data acquisition, feature extraction, model establishment, and prediction and early warning. Data acquisition includes acquiring characteristic factors such as engineering geological conditions, hydrogeological conditions, tunnel geometric characteristics, shield construction parameters, and ground subsidence data; feature extraction extracts the main features through a random forest model; model establishment uses a multi-factor Taylor series extended Kalman filter to fuse characteristic factors, and optimizes the weight coefficients of each characteristic factor through a differential evolution algorithm; prediction and early warning uses a model to predict the ground subsidence at the next time step, identifies catastrophic singularities based on a gradient ratio criterion, and performs ground subsidence early warning. The present invention combines multi-source data fusion technology with catastrophic theory to solve the problem of difficulty in early warning of ground subsidence disasters when a shield machine passes through a composite stratum, assists the shield control center in parameter setting and decision-making, and improves construction safety and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster monitoring and early warning, and mainly to an intelligent dynamic early warning method for ground collapse induced by shield tunneling. Background Art

[0002] With the acceleration of urbanization, the importance of developing and utilizing underground space has become increasingly prominent. As a relatively mature underground engineering construction technology, the shield method has been widely used due to its advantages such as high safety, fast excavation speed, and low environmental pollution. However, when the shield machine crosses the composite strata over a long linear distance, it is easy to encounter problems such as tool wear, mud cake in the center of the cutterhead, difficulty in controlling the shield attitude, excessive surface settlement, misalignment of the segments, and floating. It may even cause the shield machine to advance slowly, the shield machine to stop and jam, and the over-excavation of the slag to eventually develop into ground collapse, posing a serious threat to the safety of surrounding buildings (structures) or deep and shallow buried pipelines and the people.

[0003] Traditional shield construction is controlled by the project itself. Technical personnel rely on experience to set parameters and drive the shield machine. The labor intensity is high and the level of understanding and cooperation between man, machine and environment is limited. It is difficult to respond quickly and effectively when encountering extreme working conditions such as ground collapse caused by instability of the excavation interface. For this reason, large-diameter shield construction will show a trend of intelligence, informatization and unmanned operation in the future. Among them, the key support for the intelligent construction technology of shield machines is to obtain, identify and analyze the massive data generated by shield machine construction and excavation, and then predict and warn of ground subsidence and ground collapse.

[0004] The existing methods for predicting ground subsidence mainly include empirical methods, analytical solutions, numerical simulations, model testing, and machine learning. Due to the complexity of geological conditions and the spatial variation of geotechnical parameters, empirical methods and analytical solutions cannot accurately estimate ground subsidence. As for numerical simulation and model testing, the former has difficulties in determining model parameters, and the latter usually ignores some variables, which poses a challenge to the applicability of the two methods. Due to the ability to accurately predict multi-dimensional ground subsidence, machine learning methods such as artificial neural networks, support vector regression, and multivariate linear regression have also been applied, but these methods have some inherent limitations, such as poor robustness, over-reliance on data sets, difficulty in adjusting hyperparameters, and overfitting. Therefore, the above methods have certain defects in timely warning of ground collapse. Summary of the invention

[0005] Purpose of the invention: In view of the above-mentioned deficiencies in the prior art, an intelligent dynamic early warning method for ground collapse induced by shield tunneling is proposed to improve the real-time and accuracy of prediction and early warning.

[0006] Technical solution: An intelligent dynamic early warning method for ground collapse induced by shield tunneling, comprising the following steps: data acquisition, feature extraction, model establishment, prediction and early warning; the data acquisition includes: acquiring engineering geological conditions and hydrogeological conditions in the geological survey stage, acquiring tunnel geometric features in the tunnel design stage, acquiring shield construction parameters and ground settlement data in the shield construction stage, and constructing a feature set; the feature extraction uses a random forest algorithm to calculate the importance of each characteristic factor to surface settlement at different times, and takes the most important factors and time factors as the main features; the model establishment is used to establish a dynamic prediction model for ground settlement; the prediction and early warning iteratively predicts the settlement amount of the next time step based on the established dynamic prediction model for ground settlement, and identifies the catastrophic singularity according to the gradient ratio criterion, and issues an early warning signal for ground collapse.

[0007] Furthermore, the engineering geological conditions are mechanical parameters of composite stratum rock and soil, including density, Young's modulus, friction angle, cohesion, and rock stratum proportional coefficient; the hydrogeological conditions are hydrological parameters, including groundwater level; the tunnel geometric characteristics include tunnel burial depth, tunnel diameter, and tunnel spacing; the shield construction parameters include shield thrust, torque, penetration speed, soil bin pressure, grouting pressure, soil loss, and shield machine working status obtained via a sensor network.

[0008] Furthermore, in the random forest algorithm, the difference between the out-of-bag data error rate and the interference out-of-bag error rate of each feature is calculated, and the one with the largest difference and the time parameter are used as the main features input into the settlement dynamic prediction model.

[0009] Furthermore, a multi-factor Taylor series extended Kalman filter model optimized by differential evolution is established. The specific steps include: converting the classical Kalman filter into a multi-factor form and using Taylor series expansion to calculate the ground settlement deformation F induced by shield tunneling. s The formula is as follows:

[0010]

[0011] Among them, d, p, v are the characteristic factors extracted by random forest, t is the time parameter, α, β, γ and δ are the weight coefficients corresponding to these four parameters; g k-1 is the measurement noise at time (k-1);

[0012] The weight coefficients of the characteristic factors were optimized using the differential evolution algorithm and then input into the multi-factor Taylor series extended Kalman filter model to complete the model establishment.

[0013] Furthermore, in the prediction and early warning step, the plane manifold of time and ground subsidence is divided into three stages: stable segment 1, unstable segment, and stable segment 2, wherein stable segment 1 and stable segment 2 represent the stable continuous subsidence process before and after the ground collapse, and the unstable segment corresponds to the ground collapse event. The state vector of the next time step is continuously predicted based on the current measurement vector in the time series, and the ratio of the settlement difference of the next time step to the settlement difference of the current time step is calculated according to the gradient ratio criterion. When the ratio exceeds the designed tolerance value, a ground collapse disaster early warning is issued. The specific steps include: assuming S i-1 , S i and S i+1 are continuous behavior points in the ground subsidence plane manifold and have the same time interval, and their coordinates are (t i-1 ,y i-1 )、(t i ,y i ) and (t i+1 ,y i+1 ), then the straight line and The gradient k i and k i+1 They are:

[0014]

[0015]

[0016] Define the gradient ratio R i , and set the tolerance value [R]:

[0017]

[0018] When this equation holds true, it is considered that the behavior point S i+1 Entering the unstable stage, behavior point S i+1 is the catastrophic singularity, that is, the current t i The next time step at time t i+1 Over time, ground collapse will occur, triggering a ground warning.

[0019] Beneficial effects: The present invention discloses an intelligent dynamic early warning method for ground subsidence induced by shield tunneling, establishes a dynamic prediction model for ground subsidence, identifies ground subsidence warning signals based on a gradient ratio criterion that has theoretical support and is convenient for engineering application, and realizes early warning of ground collapse. It can solve the overfitting problem caused by massive high-dimensional data driven by existing ground collapse warning methods, and avoid the shortcomings of poor robustness and difficulty in transplantation of complex intelligent models, so that it can accurately predict and warn ground subsidence and ground collapse induced by shield tunneling in real time. It is particularly suitable for shield construction carried out in composite strata with complex geological conditions, and can effectively assist shield tunneling control to achieve disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the present invention;

[0021] Figure 2 Schematic diagram of the random forest structure of the present invention;

[0022] Figure 3 It is the Kalman filtering principle of the present invention;

[0023] Figure 4 It is a flow chart of the differential evolution algorithm of the present invention;

[0024] Figure 5 It is a schematic diagram of the mutation singularity of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described below in conjunction with the accompanying drawings.

[0026] An intelligent dynamic early warning method for ground collapse induced by shield tunneling, such as Figure 1 As shown, the following steps are included: data acquisition, feature extraction, model establishment, prediction and early warning. The engineering geological conditions and hydrogeological conditions are obtained in the geological survey stage, the tunnel geometric characteristics are obtained in the tunnel design stage, and the shield construction parameters and ground subsidence data are obtained in the shield construction stage to construct a feature set. The importance of each characteristic factor to the surface subsidence is calculated in stages using the random forest algorithm, and the main features are selected accordingly. A dynamic prediction model for ground subsidence is established by combining the multi-factor Taylor series extended Kalman filter and the differential evolution algorithm. The settlement amount of the next time step is predicted based on the established dynamic prediction model for ground subsidence, and the catastrophic singularity is identified based on the gradient ratio criterion, and an early warning signal for ground collapse is issued.

[0027] In order to construct the feature set, mechanical parameters of composite strata rock and soil such as gravity, Young's modulus, friction angle, cohesion, rock formation proportional coefficient and hydrological parameters such as groundwater level are obtained in the geological survey stage, tunnel geometric characteristics such as tunnel burial depth, tunnel diameter, and tunnel spacing are obtained in the tunnel design stage, and shield construction parameters such as thrust, torque, penetration speed, soil bin pressure, grouting pressure, soil loss, shield machine working status and ground settlement data are obtained through the sensor network in the shield construction stage. The shield machine working status is represented by binary, the stop state is represented by 0, and the continuous excavation state is represented by 1.

[0028] Among them, gravity (γ), Young's modulus (E s ), friction angle and cohesion (C) need to be corrected by soil thickness, soil depth and tunnel depth. The formula for correcting the above parameters is as follows:

[0029]

[0030]

[0031]

[0032]

[0033] Among them, t i represents the thickness of the i-th soil layer, h i represents the depth of the i-th soil layer, and h represents the tunnel burial depth.

[0034] Rock layer ratio coefficient (F c ) is used to characterize the composite characteristics of the rock and soil layers on the tunnel excavation surface. Its value range is [0,1], where 0 means all layers are soil and 1 means all layers are rock. Its expression is:

[0035]

[0036] Among them, S hard represents the rock area of ​​the tunnel face, S total Represents the total area of ​​the tunnel face.

[0037] After the feature set is constructed, the random forest algorithm is used to extract its features. Figure 2As shown, random forest uses the out-of-bag (OOB) principle to extract features and adopts bootstrap sampling. To ensure the equal probability of sample selection, samples are randomly extracted from the data set of N samples and each sample is copied and put back into the data set. Those samples that are not extracted are called out-of-bag (OOB) data. According to statistical theory, this data theoretically accounts for 1 / 3 of the data set. When using out-of-bag data for prediction, an error rate can be calculated. Introduce random white noise interference for a feature, and then recalculate the error rate using out-of-bag data. The difference between the two error rates can be used to indicate the importance of the feature. The larger the error rate, the more important this feature is to the learner. The following are the specific steps for calculating feature importance:

[0038] Step 1: For each decision tree, select the corresponding out-of-bag data to determine the out-of-bag data error rate, denoted as errOOB1.

[0039] Step 2: Introduce random noise interference to the feature X of the out-of-bag data and recalculate the out-of-bag data error, denoted as errOOB2.

[0040] Step 3: Assuming there are N trees in the forest, the importance of feature X can be evaluated:

[0041]

[0042] The importance of each feature to ground subsidence in four stages (3 days before the shield machine passes, when the shield passes, 7 days after the shield passes, and 21 days after the shield passes) is calculated. In this embodiment, the three most important features and time parameters are taken as the main characteristic factors of the subsequent model. Without loss of generality, several most important characteristic factors can be selected according to actual conditions.

[0043] A multi-factor Taylor series extended Kalman filter model for differential evolution optimization is established. Figure 3 As shown in the figure, the operation of the Kalman filter requires two main processes: prediction and correction. The time update equation is mainly responsible for estimating the past, present or future, which can be used for filtering and smoothing; forward-calculating the current state variables and estimating the error covariance, constructing a priori estimates for the next time state, which can be used as a prediction; the state update equation is responsible for correcting the system. The time update equation, state update equation and error after state update of the Kalman filter are as follows:

[0044] Time update equation:

[0045]

[0046]

[0047] State update equation:

[0048]

[0049]

[0050] Error calculation:

[0051]

[0052] in, represents the posterior state estimate at time k, represents the prior state estimate at time k, Z k Represents the measured value (observation value); A, B and H are state transformation matrices, which are usually assumed to be constants and are mainly used to adjust the state transformation process; is the n×n prior estimation error covariance matrix; P k is the n×n order posterior estimation error covariance matrix; Q is the n×n order process noise covariance matrix; R is the m×m process noise covariance matrix; I is the n×n order identity matrix; K k is the Kalman gain, which is an n×m matrix.

[0053] The above classical Kalman filter is written in the form of multiple factors and expanded using Taylor series to linearize by taking partial derivatives of the process equation. Specifically, the surface settlement F s It can be expressed as:

[0054] F s =f(d k ,p k ,v k ,t k )

[0055] Further,

[0056]

[0057] Wherein, d, p, v are the characteristic factors extracted by random forest in this embodiment, t is the time parameter, subscripts k and (k-1) represent the k time and (k-1) time respectively; α, β, γ and δ are the weight coefficients of these four parameters, and their value range is [0, 1], and the initial value is set to 1; g k-1 is the measurement noise at time (k-1), which is assumed to be white noise with Gaussian distribution.

[0058] The more complex partial derivatives in the above expansion are expressed using simple symbols:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] f k =f(d k , p k , v k , t k )

[0068] Substituting into the Taylor expansion we get:

[0069]

[0070] g k-1 is very small and can be regarded as dynamic noise with an expected value of 0.

[0071] a k =a k-1 +b k-1 (d k-1 -d k-2 )+e k-1

[0072] b k =b k-1 +l k-1

[0073] c k =c k-1 +m k-1 (p k-1 -p k-2 )+h k-1

[0074] m k =m k-1 +o k-1

[0075] n k =n k-1 +i k-1 (v k-1 -v k-2 )+q k-1

[0076] ik =i k-1 +r k-1

[0077] j k =j k-1 +l k-1 (t k-1 -t k-2 )+s k-1

[0078] l k = l k-1 +u k-1

[0079] In the above equation, e k-1 , l k-1 ,h k-1 , o k-1 ,q k-1 , r k-1 ,s k-1 and u k-1 It can be regarded as a small perturbation, and the mathematical expectation is 0. From this we can get:

[0080]

[0081] To further simplify the expression,

[0082]

[0083]

[0084]

[0085] Therefore, the original Taylor expansion can be converted into the following matrix form:

[0086]

[0087] in, is a column vector containing all partial derivatives in the above expansion, is a column vector containing all perturbations, is a sparse matrix.

[0088] The differential evolution algorithm is used to optimize the weight coefficients of the above four characteristic factors, namely α, β, γ and δ. The basic idea of ​​the differential evolution algorithm is to find the optimal answer to a given problem by iteratively improving the population of candidate solutions. The process of the differential evolution algorithm is shown in Figure 4 , first define the objective function:

[0089]

[0090] Among them, z1,z2,…,zi …,z N are the parameters that need to be optimized, namely α, β, γ and δ, and N=4.

[0091] Then define the fitness function, the expression of the fitness function is:

[0092]

[0093] Among them, C max Represents the maximum value of the objective function. The optimization goal is to minimize the value of the objective function while ensuring that the fitness is not negative.

[0094] The first step of the differential evolution algorithm is to randomly generate the initial population. The individuals in the population are called chromosomes, and each individual represents a potential solution. The formula for randomly generating the initial population is as follows:

[0095]

[0096] Where i = 1, 2, ..., P, P is the population size; z i (0) represents the i-th individual in the zero-generation population; r is a random number in the interval [0,1]; and Respectively represent Z i The upper and lower limits of the value.

[0097] Then, in each iteration, mutation, crossover, and selection operations are performed to combine existing solutions to generate new solutions. i (t), the mutation operation is performed through the following formula to obtain the corresponding mutant individual v i (t+1):

[0098] v i (t+1)=z r3 (t)+F×(z r1 (t)-z r2 (t))

[0099] Among them, r1, r2 and r3 are integers in {1, 2, …, i, …, P}, but they are not equal to each other; F is the scaling factor, and its value range is (0, 2].

[0100] For the mutant individual v in the current t generation population i (t+1) The crossover operation is performed by the following expression:

[0101]

[0102] where rand(j)∈[0,1] is the random number corresponding to the jth component; CR∈[0,1] is called the crossover probability; k is the coefficient corresponding to the i-th individual, which is randomly drawn from the integer sequence [1,2,…i,…,N]; z ij (t) represents the j-th dimension component of the i-th individual in the t-th generation population.

[0103] The selection operation uses a greedy algorithm to test individual u i (t+1) and the target individual z in the current t-generation population i (t) compete to enter the next generation of population z based on its fitness value i (t+1). The expression of the selection operation is as follows:

[0104]

[0105] Among them, f(u i (t+1) and f(z i (t) respectively represent u i (t+1) and z i (t)’s fitness.

[0106] This process is iterated until the termination criterion is met, and the individual with the best fitness in the group is returned as the best candidate solution found, that is, the four weight coefficients α, β, γ and δ. These optimized coefficient values ​​are input into the multi-factor Taylor series extended Kalman filter model to complete the model establishment.

[0107] Run the established model (see Figure 3 ), continuously iterating the time update equation and the measurement update equation, where the initial observation vector and the error covariance matrix are input into the two time update equations, and then the updated current time step posterior error covariance matrix is ​​used for the next iteration. The recursive loop is repeated, and the posterior estimate vector of the last time step is used to obtain the prior estimate of the current time step. The model continuously predicts the state vector of the next time step based on the current measurement vector, that is, the model realizes the prediction of the ground subsidence amount in the next time step.

[0108] Finally, the ground collapse is warned based on the catastrophe theory. The catastrophe theory believes that the singularity is the turning point of the sudden change of the state of the nonlinear system, which usually marks the catastrophic critical point of the system. Figure 5 As shown in the figure, the plane manifold of time and ground subsidence can be divided into three stages: stable segment 1, unstable segment, and stable segment 2. Stable segment 1 and stable segment 2 represent the stable and continuous subsidence process before and after the ground collapse, and the unstable segment corresponds to the ground collapse event. Therefore, the early warning of ground collapse can be transformed into searching for the starting point of the unstable segment in the plane manifold of ground subsidence in the space-time coordinates, that is, the catastrophic singularity.

[0109] S i-1 , S i and S i+1 are continuous behavior points in the ground subsidence plane manifold and have the same time interval, and their coordinates are (t i-1 ,y i-1 )、(t i ,y i ) and (t i+1 ,y i+1 ), then the straight line and The gradient k i and k i+1 They are:

[0110]

[0111]

[0112] Define the gradient ratio R i , and set the tolerance value [R]:

[0113]

[0114] When this equation holds true, it is considered that the behavior point S i+1 Entering the unstable stage, behavior point S i+1 is a catastrophic singularity, which means that the current t i The next time step (t i+1 time), ground collapse will occur, thus triggering a ground warning.

[0115] Without loss of generality, in addition to the extended Kalman filter, the extended particle filter or other machine learning methods can also be used to predict the ground subsidence in real time.

[0116] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An intelligent dynamic early warning method for ground collapse induced by shield tunneling, characterized in that: The following steps are involved: Data acquisition, feature extraction, model building, prediction and early warning; The data acquisition includes: acquiring engineering geological conditions and hydrogeological conditions in the geological survey stage, acquiring tunnel geometric features in the tunnel design stage, acquiring shield construction parameters and ground subsidence data in the shield construction stage, and constructing a feature set; the feature extraction uses a random forest algorithm to calculate the importance of each characteristic factor to the surface subsidence at different times, and takes the most important factors and time factors as the main features; the model establishment is used to establish a dynamic prediction model for ground subsidence; the prediction and early warning iteratively predicts the settlement amount of the next time step based on the established dynamic prediction model for ground subsidence, and identifies the catastrophic singularity according to the gradient ratio criterion, and issues an early warning signal for ground collapse; A multi-factor Taylor series extended Kalman filter model optimized by differential evolution is established. The specific steps include: converting the classical Kalman filter into a multi-factor form and using Taylor series expansion to calculate the ground settlement deformation F induced by shield tunneling. s The formula is as follows: Among them, d, p, v are the characteristic factors extracted by random forest, t is the time parameter, α, β, γ and δ are the weight coefficients corresponding to these four parameters; g k-1 is the measurement noise at time (k-1); The weight coefficients of the characteristic factors are optimized using the differential evolution algorithm and then input into the multi-factor Taylor series extended Kalman filter model to complete the model establishment; In the prediction and early warning step, the plane manifold of time and ground settlement is divided into three stages: stable segment 1, unstable segment, and stable segment 2, wherein stable segment 1 and stable segment 2 represent the stable continuous settlement process before and after the ground collapse, and the unstable segment corresponds to the ground collapse event. The state vector of the next time step is continuously predicted based on the current measurement vector in the time series, and the ratio of the settlement difference of the next time step to the settlement difference of the current time step is calculated according to the gradient ratio criterion. When the ratio exceeds the designed tolerance value, a ground collapse disaster early warning is issued. The specific steps include: assuming S i-1 , S i and S i+1 are continuous behavior points in the ground subsidence plane manifold and have the same time interval, and their coordinates are (t i-1 ,y i-1 )、(t i ,y i ) and (t i+1 ,y i+1 ), then the straight line and The gradient k i and k i+1 They are: Define the gradient ratio R i , and set the tolerance value [R]: When this equation holds true, it is considered that the behavior point S i+1 Entering the unstable stage, behavior point S i+1 is the catastrophic singularity, that is, the current t i The next time step at time t i+1 Over time, ground collapse will occur, triggering a ground warning.

2. The intelligent dynamic early warning method for ground collapse induced by shield tunneling according to claim 1 is characterized in that: The engineering geological conditions are the mechanical parameters of composite stratum rock and soil, including density, Young's modulus, friction angle, cohesion, and rock stratum proportional coefficient; the hydrogeological conditions are hydrological parameters, including groundwater level; the tunnel geometric characteristics include tunnel burial depth, tunnel diameter, and tunnel spacing; the shield construction parameters include shield thrust, torque, penetration speed, soil bin pressure, grouting pressure, soil loss, and shield machine working status obtained through a sensor network.

3. The intelligent dynamic early warning method for ground collapse induced by shield tunneling according to claim 1 is characterized in that: In the random forest algorithm, the difference between the out-of-bag data error rate and the interference out-of-bag error rate of each feature is calculated, and the one with the largest difference and the time parameter are used as the main features input into the settlement dynamic prediction model.

Images