JP7576491B2 - Analysis device and analysis program - Google Patents

Description

The present invention relates to an analysis device and an analysis program.

Groundwater problems associated with tunnel construction include the collapse of tunnel faces due to large amounts of water seeping out and the impact on the surrounding environment. To address these issues, there is a need for a method to quickly and accurately predict the amount of water seeping out.

Therefore, the SDA-SWING (groundwater information construction) method, which incorporates sequential data assimilation (SDA) to predict spring water volume, is known as an information construction method that quickly predicts groundwater behavior by incorporating observed values during construction into a model and reflecting the results in the selection of excavation methods and the consideration of groundwater countermeasures (drainage methods, water stoppage methods) (for example, non-patent documents 1 and 2).

Kenji Takahashi, Yuzo Onishi, Toru Yasuda, and Shun Kumato, "Establishment of a Simple System for Information-Based Construction of Groundwater in Mountain Tunnels (SWING)," Proceedings of the Symposium on Underground Space, Japan Society of Civil Engineers, Vol. 13, pp. 147-150, 2008. Junya Ogiso, Tomofumi Koyama, Kenji Takahashi, Toru Yasuda, Yuzo Onishi: Development of a simplified groundwater information construction system (SWING) using sequential data assimilation, Journal of the Japan Society of Civil Engineers, Special Issue on Tunnel Engineering, Vol. 66, No. 1, pp. 9-15, 2012.

A prediction method has been developed that uses measurements of wellhead spring water obtained on-site to predict the amount of spring water and incorporates this into the construction stage, using one-dimensional hydraulic formulas; however, there have been reported cases where the predicted results differ from the observed values. One of the reasons for this is thought to be the difficulty of grasping the physical quantities of the ground required for prediction in advance in unexcavated areas.

Taking the above facts into consideration, the present invention aims to accurately predict the observed values of slices corresponding to unexcavated parts of the ground.

The analysis device according to the present invention includes a setting unit that sets explanatory variables, which are physical quantities including the stratum division of the slice obtained in advance, for each slice obtained by dividing a rock mass model that models the rock mass to be excavated in the excavation direction; an objective variable prediction unit that predicts the objective variable of the slice corresponding to the portion to be excavated based on an objective variable, which is a physical quantity correlated with the explanatory variable, of the slice corresponding to the excavated portion that is on the front side of the portion to be excavated in the excavation direction; a learning unit that uses the explanatory variables and the objective variable of the slice corresponding to the excavated portion as learning data to learn an estimation model for estimating the objective variable from the explanatory variables; and a learning unit that learns an estimation model for estimating the objective variable from the explanatory variables after excavating the portion to be excavated, based on an objective variable that is a physical quantity correlated with the explanatory variable of the slice corresponding to the portion to be excavated. The system includes an observation unit that acquires observed values, which are physical quantities corresponding to the excavation target portion, an identification unit that identifies the objective variable of the slice corresponding to the excavation target portion from the observed values of the slice corresponding to the excavation target portion using an observation equation that represents the relationship between the observed values and the objective variable, and an observation value prediction unit that predicts the objective variable of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the identified excavation target portion, the explanatory variable of the slice corresponding to the unexcavated portion that is further in the excavation direction than the excavation target portion, and the estimation model, and predicts the observed value of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the unexcavated portion and the observation equation.

According to the analysis device of the present invention, the setting unit sets explanatory variables, which are physical quantities including the stratum classification of the slices obtained in advance, for each slice obtained by dividing a natural ground model that models the natural ground to be excavated in the excavation direction. The objective variable prediction unit predicts the objective variable of the slice corresponding to the portion to be excavated based on the objective variable, which is a physical quantity correlated with the explanatory variable of the slice corresponding to the excavated portion that is on the near side of the portion to be excavated in the excavation direction.

Then, a learning unit learns an estimation model for estimating the objective variable from the explanatory variables, using the explanatory variables and the objective variable of the slice corresponding to the excavated portion as learning data. An observation unit obtains an observation value, which is a physical quantity related to the objective variable of the slice corresponding to the excavated portion, after the portion to be excavated is excavated.

Then, an identification unit identifies the objective variable of the slice corresponding to the portion to be excavated from the observed values of the slice corresponding to the portion to be excavated, using an observation equation that represents the relationship between the observed values and the objective variable. An observation value prediction unit predicts the objective variable of the slice corresponding to the unexcavated portion, using the objective variable of the slice corresponding to the identified portion to be excavated, the explanatory variables of the slice corresponding to the unexcavated portion that is further in the excavation direction than the portion to be excavated, and the estimation model, and predicts the observed value of the slice corresponding to the unexcavated portion, using the objective variable of the slice corresponding to the unexcavated portion and the observation equation.

In this way, the explanatory variables and the objective variables of the slice corresponding to the excavated portion are used as learning data, and an estimation model for estimating the objective variables from the explanatory variables is learned. The objective variables of the slice corresponding to the identified portion to be excavated and the explanatory variables of the slice corresponding to the unexcavated portion are predicted using the estimation model. The observed values of the slice corresponding to the unexcavated portion are predicted using the observation equation, thereby making it possible to accurately predict the observed values of the slice corresponding to the unexcavated portion of the natural ground.

The estimation model according to the present invention is a neural network model that receives the explanatory variables as input and outputs the objective variables.

The learning unit according to the present invention uses the learned estimation model to determine the ratio of the objective variables between the stratum divisions, and the observed value prediction unit predicts the objective variable of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the identified portion to be excavated, the explanatory variables of the slice corresponding to the unexcavated portion that is further in the excavation direction than the portion to be excavated, and the ratio of the objective variables between the stratum divisions.

The objective variable prediction unit according to the present invention predicts the objective variable of the slice corresponding to the portion to be excavated based on the objective variable of the slice corresponding to the excavated portion by time updating the Kalman filter, with the objective variable of each slice being the state quantity at each time of the Kalman filter, and the identification unit identifies the objective variable of the slice corresponding to the portion to be excavated based on the objective variable estimated by the objective variable prediction unit by observation updating of the Kalman filter, the observed value of the slice corresponding to the portion to be excavated, and an observation equation expressing the relationship between the observed value and the objective variable.

The analysis device according to the present invention further includes an explanatory variable update unit that updates the explanatory variables of the slice corresponding to the unexcavated portion after the excavation target portion is excavated.

The analysis program of the present invention includes a computer, a setting unit that sets explanatory variables, which are physical quantities including the stratum division of the slice, obtained in advance for each slice obtained by dividing a rock mass model that models the rock mass to be excavated in the excavation direction, a response variable prediction unit that predicts the response variable of the slice corresponding to the portion to be excavated based on a response variable, which is a physical quantity correlated with the explanatory variable, of the slice corresponding to the excavated portion that is on the front side of the portion to be excavated in the excavation direction, a learning unit that uses the explanatory variables and the response variable of the slice corresponding to the excavated portion as learning data to learn an estimation model for estimating the response variable from the explanatory variables, and after excavating the portion to be excavated, a response variable prediction unit that predicts the response variable of the slice corresponding to the portion to be excavated based on a response variable, which is a physical quantity correlated with the explanatory variable, of the slice corresponding to the portion to be excavated The analysis program functions as an observation unit that acquires observed values, which are physical quantities corresponding to the excavation target portion, an identification unit that identifies the objective variable of the slice corresponding to the excavation target portion from the observed values of the slice corresponding to the excavation target portion using an observation equation that represents the relationship between the observed values and the objective variable, and an observation value prediction unit that predicts the objective variable of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the identified excavation target portion, the explanatory variable of the slice corresponding to the unexcavated portion that is further in the excavation direction than the excavation target portion, and the estimation model, and predicts the observed value of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the unexcavated portion and the observation equation.

As described above, the analysis device and analysis program of the present invention have the effect of being able to accurately predict the observed values of slices that correspond to unexcavated parts of the ground.

1 is a block diagram showing an analysis device according to an embodiment of the present invention; 1 is a functional block diagram illustrating an analysis device according to an embodiment of the present invention. FIG. 13 is a diagram showing an example of slices obtained by dividing a natural ground model in the excavation direction. FIG. 1 is a diagram illustrating an example of a neural network model. FIG. 13 is a diagram showing an example of a well formula that is an observation equation in which the amount of spring water is an observation value. 4 is a flowchart showing the contents of an analysis processing routine of the analysis device according to the embodiment of the present invention. 1 is a graph showing examples of explanatory variables and response variables in an embodiment. FIG. 2 is a diagram showing a configuration of a neural network model in an embodiment. 1 is a graph showing prediction results of a response variable in an example. FIG. 1 illustrates operations in inverse calculation in the prior art. FIG. 1 illustrates operations in data assimilation in the prior art. FIG. 1 is a diagram illustrating filtering taking into account the observation space and the state space in the prior art. FIG. 1 is a diagram showing an example of the data assimilation range in spring water volume prediction. FIG. 13 is a diagram showing an example of adding observed values in spring water volume prediction.

The following describes in detail an embodiment of the present invention with reference to the drawings.

<Principle of the embodiment of the present invention>
(Prior art and its problems)

In recent years, efforts have been made in various places to incorporate methods such as inverse calculation (data assimilation, etc.) and deep learning in tunnel excavation simulations to improve the prediction accuracy. Here, we will discuss the problems associated with each method.

(Inverse calculation/forward calculation (data assimilation))
In a typical simulation problem, the relationship between the observed value y and the state quantity x is replaced by the formula y = h(x). The state quantity x is the main parameter that constitutes the simulation model, but it is difficult to observe directly, so it is generally estimated from the observed value y. For example, the state quantity x is a soil constant, and the observed value y is the amount of ground deformation and water level drop caused by excavation.

In an inverse calculation problem, the state quantity x that matches the observed value y is calculated. The solution is determined deterministically, and there is only one state quantity x for one observed value y.

Now, if the state quantity before the observed value is obtained is x _t-1 , the state quantity after the observed value is obtained is x _t , the observed value is y _t , the change in time of x _t is f _t (x), and the formula is h _t (x), the operations in inverse analysis can be expressed as shown in FIG. 10.

In data assimilation, it is assumed that the state quantity _xt and the observation value _yt are affected by known Gaussian noise due to their respective characteristics, and the noise in the respective distributions is set as _υt and _ωt . In this case, in the inverse analysis, the observation value is displayed as a probability distribution as shown in Figure 11.

Now, consider that the formula y = h(x) is constant and only the state variable x is updated over time. Under the conditions that f(x) and h(x) are both linear and both υ and ω are Gaussian noise, the probability distribution of _xt reflecting the observed value can be obtained by a formula.

x _t =F·x _t−1 +υ υ∼N(0,Q): This state space model is represented by a normal distribution with variance value Q.

y _t =H·x _t +ω ω∼N(0,R): This observation model is represented by a normal distribution with variance R.

It is assumed that the distribution x _t-1 at a certain time t-1 follows a normal distribution with a mean value x _t-1 and a variance value V _t-1 as shown below.

x _t-1 ~N(x _t-1 , V _t-1 )

When an observation value _yt is given, the state quantity _xt can be calculated as follows: This operation is called filtering (see FIG. 12).

K _t =V _t-1・H'/(H・V _t-1・H'+R _t )
x _t =x _t-1 +K _t・(y _t -H・x _t-1 )
V _t =V _t-1・(1-K _t・H)

The relationship between _yt and _xt is a linear mapping of the coefficient H (observation equation), and this is used to calculate _Kt, which is the update coefficient for the posterior distribution _xt ·variance value _Vt and the prior distribution xt _-1 · _Vt-1 . _Kt is the ratio of Vt _-1 to _Rt . This simply calculates the mean value _xt and variance value _Vt of the posterior distribution using the ratio _Kt of the variance values Vt _{-1 and Rt in the prior distribution xt- 1 and} the observation distribution _yt . This _Kt is called the Kalman gain.

Once the posterior distribution _xt is calculated, state predictions for one period ahead and beyond from that point in time can be performed using the state equation and the observation equation. The equation for one period ahead prediction is as follows:

x _t =F・x _t−1
V _t =F・V _t−1・F'+Q _t

In the calculation of _Vt , it is a prerequisite that the state equation F is linear.

From then on, this process is repeated each time observational data is obtained to improve the accuracy of the posterior distribution.

The biggest problem with applying inverse calculation to tunnel excavation simulation is that this method only identifies the objective variables in the areas that affect the observed values. In particular, in tunnel excavation with a long construction length, it is not possible to identify the objective variables for unconstructed sections outside the affected area, and the results of the forward calculation (future prediction) are not updated (see Figure 13).

In addition, if there are multiple stratum divisions within the affected area (see Figure 3), there will also be multiple objective variables, and so the inverse calculation requires the same number of observations as the objective variables. However, in most cases, only one observation is obtained at the tunnel face position during tunnel excavation, and a separate observation must be obtained ahead of the face (see Figure 14). For example, to solve the equation y = ax1 + bx2, two values of y are required.

Furthermore, in normal inverse/forward calculation processing, the value of the objective variable x is identified using only the observed value y, and the explanatory variable z is not currently mentioned. For example, there is a clear correlation between the explanatory variable z and the objective variable x, such as the relationship between crack density and permeability coefficient, but it is difficult to prove this relationship with an equation, and when the dimensions of the explanatory variables increase, such as elastic wave velocity and uniaxial compressive strength in addition to cracks, it becomes even more difficult to derive a relational equation.

Therefore, it is common to use an observation equation whose relationship is clearly shown, and then use only the observed values to back-calculate the dependent variable.

(Correlation between hydraulic conductivity and soil overburden in the SWING method (well formula))
In the SWING method, the tunnel spring water volume at the slice cross section is calculated by Bear's formula, which combines rainfall infiltration with Dupuit's quasi-uniform flow. This formula is a type of well formula, and assumes that the water level directly above the tunnel drops to the tunnel excavation depth immediately after excavation, and then gradually expands to the steady-state water level drop range over time. Since the target is a mountain tunnel, the groundwater level before excavation is assumed to be about GL or GL-5m.

In this formula, the spring water volume _qt and the hydraulic conductivity k are expressed by a proportional relationship related to the logarithmic hydraulic conductivity, _qt = α√k.

When considering the water flow in the infiltration theory, the water level directly above the tunnel does not change immediately after excavation, but gradually drops over time. Therefore, the flow rate _qt calculated by the SWING method is a value that includes an error due to this modeling. When the soil overburden is small, the steady-state water level is expected to drop to the tunnel position, so this error is small. At this time, the spring water volume (= permeability coefficient identified by the SWING method) is proportional to the square of the soil overburden. However, when the soil overburden is large, the groundwater level does not drop to the tunnel position even in the steady state, and the error becomes large. At this time, the effect of the soil overburden on the spring water volume is expected to be small.

Therefore, if future predictions are made assuming that the hydraulic conductivity of the natural ground is constant, the amount of spring water calculated by the SWING method increases in proportion to the square of the soil cover. However, in reality, once the soil cover exceeds a certain level, there is almost no change in the amount of spring water, so the correlation between the soil cover and the amount of spring water approaches logarithmic proportion. If assimilation is performed using the SWING method for such a spring water volume, the hydraulic conductivity will be a small value due to modeling errors. Therefore, to improve the prediction accuracy of the SWING method for natural ground with a uniform hydraulic conductivity, it is necessary to correct the set hydraulic conductivity according to the soil cover.

In this embodiment, this correlation is calculated internally by including soil cover as an explanatory variable in machine learning, and the hydraulic conductivity is automatically corrected. As a result, even if there is an error in the observation equation, highly accurate predictions are possible by incorporating machine learning.

(Machine Learning)
As described above, a clear correlation exists between explanatory variable z and objective variable x, but it is difficult to prove this relationship with an equation. However, if there are multiple combinations of objective variable x corresponding to explanatory variable z, it is possible to estimate an equation from them using a method such as regression analysis.

When the explanatory variable z is high-dimensional, it is difficult to carry out regression analysis, etc., but with the recent improvements in computer performance, it is now possible to use machine learning to calculate this correlation relatively easily without performing complex formulation work.

In this embodiment, by incorporating this machine learning process into the flow of reverse and forward calculations, when the explanatory variable z is updated in addition to the observed value y, it is possible to predict the objective variable x of the unconstructed section from the results.

In addition, the explanatory variable z is unrelated to the range of influence of the observation equation, making it possible to update the objective variable in unconstructed sections outside that range. Furthermore, since it is performed independently of the inverse calculation, objective variables are calculated for multiple rock types, and by assuming that their ratio is constant, the objective variable x has a single dimension, making it possible to identify a single observed value y.

It is possible to directly associate the observed value y with the explanatory variable z using machine learning alone without using inverse calculation, but it is generally possible to associate multiple observed items (e.g. deformation amount, component stress, amount of well water, range of influence, etc.) with the objective variable through one observation equation, allowing for efficient calculations.

<Configuration of the analysis device according to the embodiment of the present invention>
As shown in FIG. 1, an analysis device 100 according to an embodiment of the present invention includes a CPU 12, a graphics card 13, a GPU 14, a RAM 16, a HDD 18, a communication interface 21, and a bus 23 for interconnecting these components.

The CPU 12 and GPU 14 execute various programs. The RAM 16 is used as a work area when the CPU 12 executes the various programs. The HDD 18, which serves as a recording medium, stores various programs and data, including a program for executing the analysis processing routine described below.

When the analysis device 100 in this embodiment is represented by functional blocks along with a program for executing an analysis processing routine, it becomes as shown in FIG. 2. The analysis device 100 has an input unit 10, a calculation unit 20, and an output unit 50.

The input unit 10 receives as input data representing each slice obtained by dividing a natural ground model, which is a model of the natural ground to be excavated, in the excavation direction, and explanatory variables, which are physical quantities including the stratum divisions determined in advance for each slice (see Figure 3).

The explanatory variable z can be obtained before the tunnel excavation reaches the relevant slice based on design conditions, preliminary surveys (forward exploration, etc.), past performance, etc., and is a value that is expected to be related to the objective variable but is difficult to formulate.

Examples of explanatory variables include rock type, which is a geological layer classification, as well as soil cover, tunnel cross-sectional dimensions, crack density, natural ground elastic wave velocity, uniaxial compressive strength, and rainfall intensity.

The input unit 10 also accepts as input data representing the correlation between explanatory variables and objective variables obtained from other excavation work results.

The objective variables are values that can be converted into various observed values through the observation equation. These values are determined by determining the explanatory variable z that corresponds to each stratum classification. For example, objective variables include the rock mass deformation coefficient, hydraulic conductivity, or rainfall infiltration rate.

In addition, after the excavation target portion is excavated, the input unit 10 receives as input the observed value of the slice corresponding to the excavation target portion, which is a physical quantity related to the objective variable.

Here, the various observed values y are values obtained from various behaviors that occur in the tunnel when the tunnel progress reaches each slice, and cannot be measured in advance. For example, the observed values include support displacement, face extrusion amount, appropriate charge amount, face image (feature value), or well water amount.

In addition, after the portion to be excavated is excavated, if there are explanatory variables to be updated for a slice corresponding to an unexcavated portion, the input unit 10 accepts as input the explanatory variables of the slice corresponding to the unexcavated portion.

The calculation unit 20 includes a setting unit 22, a target variable prediction unit 24, a learning unit 26, an explanatory variable update unit 28, an observation unit 30, an identification unit 32, and an observed value prediction unit 34.

The setting unit 22 sets explanatory variables for each slice of the ground model that models the ground to be excavated.

The setting unit 22 also sets an initial objective variable x for each slice that corresponds to the stratum classification included in the explanatory variables, based on the correlation between the objective variable and the stratum classification obtained from the results of other excavation works.

The objective variable prediction unit 24 predicts the objective variables of the slice corresponding to the portion to be excavated based on the objective variables of the slice corresponding to the excavated portion that is on the near side of the portion to be excavated in the excavation direction.

Specifically, the objective variable prediction unit 24 predicts the objective variable of the slice corresponding to the portion to be excavated based on the objective variable of the slice corresponding to the excavated portion by updating the Kalman filter over time, with the objective variable of each slice being the state quantity of the Kalman filter at each time. Here, the Kalman filter is updated over time with each slice in the excavation direction of the natural ground model corresponding to each time step of the Kalman filter. In other words, the time change of the objective variable is not predicted.

The learning unit 26 generates explanatory variables and objective variables for each slice corresponding to the excavated portion as learning data, and learns a neural network model for estimating the objective variable from the explanatory variables based on the learning data (see Figure 4).

In this way, as tunnel excavation progresses, the identified objective variable x and the corresponding explanatory variable z are obtained for each slice in the already constructed section, and the explanatory variables and identified objective variables of each slice are generated as learning data, and this learning data is subjected to machine learning to train the neural network model.

The learning unit 26 also uses the trained neural network model to determine the ratio of objective variables between stratum divisions. Specifically, the trained neural network model is used to determine objective variables x1', x2' corresponding to each stratum division, and a ratio α2 is determined such that x2' = α2 x x1' from objective variable x1' corresponding to the stratum division in the portion to be excavated and objective variable x2' corresponding to another stratum division. This ratio is determined for all stratum divisions.

After the excavation target portion is excavated, if there are explanatory variables to be updated for a slice corresponding to an unexcavated portion by performing forward exploration, the explanatory variable update unit 28 updates the explanatory variables of that slice.

After the observation unit 30 excavates the excavation target portion, it obtains the observation values of the slice corresponding to the input excavation target portion.

The identification unit 32 identifies the objective variable of the slice corresponding to the portion to be excavated by inverse calculation x = f'(y) using the observation equation f that expresses the relationship between the observation value y of the slice corresponding to the portion to be excavated and the objective variable x.

Specifically, the identification unit 32 identifies the objective variable of the slice corresponding to the portion to be excavated based on the objective variable estimated by the objective variable prediction unit 24 through observation updates of the Kalman filter, the observed values of the slice corresponding to the portion to be excavated, and the observation equation.

Here, the observation equation f is a simulation model equivalent to y = f(x) or y = f(x, z) that explains the relationship between the objective variable x, some of the explanatory variables z, and various observed values y, and contains errors associated with modeling.

For example, the observation equation f is a three-dimensional FEM model in which the deformation amount and the spring water amount are the observed values, a two-dimensional FEM model in which the deformation amount and the spring water amount are the observed values, an elastic support beam model in which the deformation amount is the observed value, or a well formula in which the spring water amount shown in Figure 5 is the observed value.

The observation value prediction unit 34 predicts the objective variables of the slices corresponding to the identified excavation target portion, the explanatory variables of the slices corresponding to the unexcavated portion that is further in the excavation direction than the excavation target portion, and the learning results of the neural network model, and predicts the observation value of the slice corresponding to the unexcavated portion using the objective variables of the slices corresponding to the unexcavated portion and the observation equation.

Specifically, the observation value prediction unit 34 predicts the objective variables of the slice corresponding to the undrilled portion using the objective variables of the slice corresponding to the identified portion to be excavated, the explanatory variables of the slice corresponding to the undrilled portion that is further in the excavation direction than the portion to be excavated, and the ratio of objective variables between the stratum divisions.

For example, the objective variable x1 for the stratum division 1 of the slice in the excavation target area and the corresponding rock type is set as the reference value, and the objective variables x2, x3, x4... for the other stratum divisions 2, 3, 4... are set as multiples of x1, as x2 = α2 x x1, x3 = α3 x x1, x4 = α4 x x1,... to find the objective variable x for each slice in the unconstructed section.

Then, the observation value prediction unit 34 predicts the observation value y of each slice in the unconstructed section by forward calculation using the observation equation y = f(x), and outputs it via the output unit 50.

<Operation of the analysis device>
Next, the operation of analysis device 100 according to the embodiment of the present invention will be described.

When performing excavation work on the ground to be excavated, the input unit 10 receives as input data representing each slice obtained by dividing the ground model, which is a model of the ground to be excavated, in the excavation direction, explanatory variables, which are physical quantities including the stratum divisions determined in advance for each slice, and data representing the correlation between the explanatory variables and the objective variables obtained from the results of other excavation work, and the analysis device 100 executes the analysis processing routine shown in FIG. 6.

First, in step S100, the setting unit 22 sets the input explanatory variables for each slice of the ground model that models the ground to be excavated.

Then, in step S102, the setting unit 22 sets an initial objective variable x corresponding to the stratum classification of the explanatory variable for each slice, based on the correlation between the stratum classification and the objective variable obtained from the results of other excavation work.

In step S104, the objective variable prediction unit 24 predicts the objective variable of the slice corresponding to the portion to be excavated based on the objective variable of the slice corresponding to the excavated portion that is on the near side of the portion to be excavated in the excavation direction.

In step S106, the learning unit 26 generates explanatory variables and objective variables for each slice corresponding to the excavated portion as learning data, and learns a neural network model for estimating the objective variable from the explanatory variables based on the learning data.

In step S108, the learning unit 26 uses the trained neural network model to determine the ratio of the objective variables between the strata sections.

In step S110, after the excavation target portion is excavated, if the explanatory variables to be updated for the slice corresponding to the undrilled portion are input by performing forward exploration, the explanatory variable update unit 28 updates the explanatory variables of that slice.

In step S112, the observation unit 30 obtains the observation values of the slice corresponding to the excavation target portion after excavating the excavation target portion.

In step S114, the identification unit 32 identifies the objective variable of the slice corresponding to the portion to be excavated by inverse calculation x = f'(y) using the observation equation f that expresses the relationship between the observation value y of the slice corresponding to the portion to be excavated and the objective variable x.

In step S116, the observation value prediction unit 34 predicts the objective variable of the slice corresponding to the identified excavation target portion, the explanatory variables of the slice corresponding to the unexcavated portion that is further in the excavation direction than the excavation target portion, and the learning results of the neural network model, and predicts the observation value of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the unexcavated portion and the observation equation. The prediction result is then output by the output unit 50.

In step S118, it is determined whether the excavation work has been completed. If the excavation work has not been completed, the slice of the portion to be excavated is changed to one further in the excavation direction, and the process returns to step S104. On the other hand, if the excavation work has been completed, the analysis processing routine is terminated.

<Example>
In this embodiment, in addition to rock type, which is a geological layer classification, the explanatory variables are earth covering, porosity, elastic wave velocity, and drilling energy when drilling holes for charging the tunnel face, as shown in Fig. 7, and the permeability coefficient is the objective variable. The explanatory variables are determined based on design conditions and various surveys, and the objective variable is determined by inverse calculation using an observation equation from observed values during excavation work.

The neural network model used for training uses a seven-layer perceptron structure with 100 nodes, as shown in Figure 8.

Figure 9 (A) shows a comparison between the measurement results using the SWING method and the predicted permeability coefficient of the unconstructed section when excavation work has been performed 300 m. Figure 9 (B) shows a comparison between the measurement results using the SWING method and the predicted permeability coefficient of the unconstructed section when excavation work has been performed 700 m. Figure 9 (C) shows a comparison between the measurement results using the SWING method and the predicted permeability coefficient of the unconstructed section when excavation work has been performed 1000 m.

In this way, it was confirmed that as the excavation work progressed, the neural network model's learning progressed and the prediction accuracy of the permeability coefficient in unconstructed sections improved.

As described above, the analysis device according to the embodiment of the present invention uses the explanatory variables and objective variables of the slice corresponding to the excavated portion as learning data to learn a neural network model for estimating the objective variables from the explanatory variables, predicts the objective variables of the slice corresponding to the unexcavated portion using the objective variables of the slice corresponding to the identified portion to be excavated, the explanatory variables of the slice corresponding to the unexcavated portion, and the ratio of the objective variables between the stratum divisions obtained from the neural network model, and predicts the observed values of the slice corresponding to the unexcavated portion using an observation equation, thereby making it possible to accurately predict the observed values of the slice corresponding to the unexcavated portion of the natural ground.

In addition, it is possible to identify the objective variable x outside the range of influence of the observation equation, which is not possible with normal inverse/forward calculations.

In addition, even if there are multiple geological layers within the range of influence of the observation equation, it is possible to identify the objective variable through inverse analysis using a single observed value.

In addition, even if the observation equation contains modeling errors, it is possible to identify a target variable that takes those errors into account, and to accurately predict the observed values of slices that correspond to unexcavated parts of the ground.

In addition, the accuracy of the prediction of the objective variable will improve as the excavation progresses, but it is also expected that the accuracy of the prediction of the objective variable will improve by using past performance data in advance predictions before excavation.

The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the spirit of the invention.

For example, in the above analysis processing routine, the setting of the initial objective variable x based on the correlation between the stratum classification and the objective variable obtained from the results of other excavation works in step S102 and the updating of the explanatory variables of the slice in step S110 may be omitted. In this case, the input of data representing the correlation between the explanatory variables and the objective variable obtained from the results of other excavation works by the input unit 10 and the input of the explanatory variables to be updated for the slice corresponding to the unexcavated portion may be omitted.

In addition, while the above embodiment assumes a horizontally excavated tunnel, the present invention may also be applied to vertical excavation work such as shaft and retaining wall excavation.

The program of the present invention may also be provided by storing it on a storage medium.

REFERENCE SIGNS LIST 10 Input unit 20 Calculation unit 22 Setting unit 24 Objective variable prediction unit 26 Learning unit 28 Explanatory variable update unit 30 Observation unit 32 Identification unit 34 Observation value prediction unit 50 Output unit 100 Analysis device

Claims (6)

a setting unit that sets explanatory variables, which are physical quantities including a stratum classification of a slice obtained in advance, for each slice obtained by dividing a rock mass model that is a model of a rock mass to be excavated in the excavation direction;
a response variable prediction unit that predicts the response variable of the slice corresponding to the excavation target portion based on a response variable that is a physical quantity correlated with the explanatory variable of the slice corresponding to the excavated portion that is on the front side of the excavation target portion in the excavation direction;
a learning unit that learns an estimation model for estimating the objective variable from the explanatory variables by using the explanatory variables and the objective variable of the slice corresponding to the excavated portion as learning data;
an observation unit that acquires an observation value, which is a physical quantity associated with the objective variable, of the slice corresponding to the excavation target portion after the excavation target portion is excavated;
an identification unit that identifies the objective variable of the slice corresponding to the excavation target portion from the observation value of the slice corresponding to the excavation target portion by using an observation equation that represents a relationship between the observation value and the objective variable;
an observation value prediction unit that predicts the objective variable of the slice corresponding to the identified excavation target portion, the explanatory variables of the slice corresponding to an unexcavated portion that is further in the excavation direction than the excavation target portion, and the estimation model, and predicts the observation value of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the unexcavated portion and the observation equation;
An analysis device comprising: The analysis device according to claim 1, wherein the estimation model is a neural network model that receives the explanatory variables as input and outputs the objective variables. The learning unit calculates a ratio of the objective variable between the stratum divisions using the learned estimation model,
The analysis device described in claim 1 or 2, wherein the observation value prediction unit predicts the dependent variable of the slice corresponding to the identified portion to be excavated, the explanatory variables of the slice corresponding to the unexcavated portion that is further in the excavation direction than the portion to be excavated, and the ratio of the dependent variables between the stratum divisions. The objective variable prediction unit predicts the objective variable of the slice corresponding to the portion to be excavated based on the objective variable of the slice corresponding to the already excavated portion by updating the objective variable of each slice as a state quantity of the Kalman filter at each time point, by time updating the Kalman filter;
The analysis device according to any one of claims 1 to 3, wherein the identification unit identifies the dependent variable of the slice corresponding to the portion to be excavated based on the dependent variable estimated by the dependent variable prediction unit through observation updates of the Kalman filter, the observed values of the slice corresponding to the portion to be excavated, and an observation equation representing the relationship between the observed values and the dependent variable. The analysis device according to any one of claims 1 to 4, further comprising an explanatory variable update unit that updates the explanatory variables of the slice corresponding to the unexcavated portion after the excavation target portion is excavated. Computer,
a setting unit for setting explanatory variables, which are physical quantities including a stratum classification of a slice obtained in advance, for each slice obtained by dividing a rock mass model, which is a model of a rock mass to be excavated, in an excavation direction;
an objective variable prediction unit for predicting the objective variable of the slice corresponding to the excavation target portion based on an objective variable which is a physical quantity correlated with the explanatory variable of the slice corresponding to the excavated portion which is on the front side of the excavation target portion in the excavation direction;
a learning unit that learns an estimation model for estimating the objective variable from the explanatory variables by using the explanatory variables and the objective variable of the slice corresponding to the excavated portion as learning data;
an observation unit that acquires an observation value, which is a physical quantity associated with the objective variable, of the slice corresponding to the excavation target portion after the excavation target portion is excavated;
an identification unit that identifies the objective variable of the slice corresponding to the portion to be excavated from the observation values of the slice corresponding to the portion to be excavated, using an observation equation that represents the relationship between the observation values and the objective variable; and an observation value prediction unit that predicts the objective variable of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the identified portion to be excavated, the explanatory variables of the slice corresponding to an unexcavated portion that is further back in the excavation direction than the portion to be excavated, and the estimation model, and predicts the observation value of the slice corresponding to the unexcavated portion using the objective variable of the slice corresponding to the unexcavated portion and the observation equation.