Displacement amount learning device, displacement amount learning method, displacement amount prediction device, and displacement amount prediction method

Abstract

To provide a displacement amount learning device, a displacement amount learning method, a displacement amount prediction device, and a displacement amount prediction method capable of predicting a displacement amount with accuracy as reducing labor load necessary for analysis further than previous ones.SOLUTION: A displacement amount learning device 10 is provided with a leaning data acquisition section 11 and a learning processing section 12 to learn a displacement amount generated on an inner peripheral surface of a tunnel. The learning data acquisition section 11 acquires a group of a whole or a part of observed evaluation, timbering rigidity, earth covering and excavation diameter on a tunnel work face, and a displacement amount of the tunnel work face. The learning process section 12 comprises an AI model to perform machine-learning on the AI model using the learning data so as to output the displacement amount by inputting a whole or a part of the observed evaluation, the timbering rigidity, the earth covering and the excavation diameter.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a displacement amount learning device, a displacement amount learning method, a displacement amount prediction device, and a displacement amount prediction method. [Background technology]

[0002] In mountain tunnel construction, two-dimensional and three-dimensional FEM (Finite Element Method) analysis is often used to predict the amount of displacement inside a tunnel. Regarding the prediction of displacement inside a tunnel, there are technologies such as those described in Patent Documents 1 and 2. The tunnel soundness determination device described in Patent Document 1 comprises a convergence displacement prediction means and an index value calculation means. The convergence displacement prediction means uses measured values ​​obtained by actually measuring the displacement of the inner surface of the tunnel at measurement locations, performs multiple regression analysis on a relational expression showing the relationship between the measured values ​​and the convergence value of the displacement of the inner surface of the tunnel, and calculates the convergence value of the displacement of the inner surface of the tunnel. The index value calculation means calculates an index value related to the soundness of the tunnel using the convergence value of the displacement of the inner surface of the tunnel, geological condition values, and structural condition values. The effect prediction method described in Patent Document 2 is a method for predicting the effect of countermeasures planned to be implemented in a tunnel, and predicts the deformation suppression effect of the countermeasures planned to be implemented in the deformed tunnel based on the internal displacement velocity of the tunnel. In this effect prediction method, the deformation suppression effect of the countermeasures planned to be implemented in a tunnel is predicted based on prior numerical analysis data obtained by numerically analyzing the deformation suppression effect of the countermeasures in advance, with conditions that affect the deformation suppression effect of the countermeasures as parameters, and the internal displacement velocity of the tunnel. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2011-163017 [Patent Document 2] Japanese Patent Publication No. 2020-200648 [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] However, conventional methods for predicting displacement had drawbacks, such as the time and effort required for analysis, and the variability in prediction accuracy due to uncertainties in the ground conditions. From this perspective, the present invention provides a displacement amount learning device, a displacement amount learning method, a displacement amount prediction device, and a displacement amount prediction method that can predict displacement amounts with high accuracy while reducing the effort required for analysis compared to conventional methods. [Means for solving the problem]

[0005] The displacement amount learning device according to the present invention is a displacement amount learning device that learns the amount of displacement occurring on the inner surface of a tunnel, comprising a learning data acquisition unit and a learning processing unit. The learning data acquisition unit is used for observation and evaluation of the tunnel face. , support rigidity, soil cover and excavation diameter The pair of the displacement amount of the tunnel face is acquired as training data. The training processing unit has an AI model and the observation evaluation , the support rigidity, the soil cover and the excavation diameter The AI ​​model is trained using the training data so that it outputs the displacement amount when input. The observational evaluation includes at least an evaluation score related to compressive strength. In the displacement learning device according to the present invention, observation and evaluation of the tunnel face By inputting all of the following: support stiffness, soil cover, and excavation diameter, the displacement amount is output using the training data. The AI ​​model is trained using machine learning. By using this trained AI model, it's possible to predict the trend of displacement occurring on the inner surface of the tunnel. Therefore, it's possible to predict displacement that reflects the ground conditions without complex analysis, and to perform construction management based on this prediction in real time. The aforementioned observation and evaluation may further include evaluation points for at least one of the following observation items: stability of the tunnel face, condition of the uncut surface, weathering and alteration, proportion of the fractured area in the face, crack spacing, condition of the cracks, morphology of the cracks, seepage, and deterioration due to water.

[0006] The displacement amount learning method according to the present invention is a displacement amount learning method for learning the amount of displacement occurring on the inner surface of a tunnel, comprising a learning data acquisition step and a learning processing step. In the learning data acquisition step, observation and evaluation of the tunnel face are performed. , support rigidity, soil cover and excavation diameter The pair of the observed and evaluated tunnel face is acquired as training data. , the support rigidity, the soil cover and the excavation diameter The AI ​​model is trained using the training data so that it outputs the displacement amount when input. The observational evaluation includes at least an evaluation score related to compressive strength. In the displacement learning method according to the present invention, observation and evaluation of the tunnel face By inputting all of the following: support stiffness, soil cover, and excavation diameter, the displacement amount is output using the training data. The AI ​​model is trained using machine learning. By using this trained AI model, it's possible to predict the trend of displacement occurring on the inner surface of the tunnel. Therefore, it's possible to predict displacement that reflects the ground conditions without complex analysis, and to perform construction management based on this prediction in real time.

[0007] The displacement prediction device according to the present invention comprises a prediction processing unit having the trained AI model learned by the above-described displacement learning method. The prediction processing unit observes and evaluates the tunnel face at the predicted location. , support rigidity, soil cover and excavation diameter The amount of displacement of the tunnel face at the predicted position is predicted by inputting the data into the trained AI model. Furthermore, the displacement amount prediction method according to the present invention includes a prediction processing step in which the displacement amount is predicted using the trained AI model trained by the above-described displacement amount learning method. In the prediction processing step, the tunnel face at the predicted position is observed and evaluated. , support rigidity, soil cover and excavation diameter The amount of displacement of the tunnel face at the predicted position is predicted by inputting the data into the trained AI model. In the displacement prediction device and displacement prediction method according to the present invention, observation and evaluation of the tunnel face The displacement amount is output by inputting all of the following: support stiffness, soil cover, and excavation diameter. By using a machine learning-based AI model, it is possible to predict the trend of displacement occurring on the inner surface of a tunnel. Therefore, it is possible to predict displacement amounts that reflect the condition of the ground without performing complex analyses, and construction management based on these predictions can be carried out in real time.

[0008] The learning processing step may comprise an initial learning step and an additional learning step. In said initial learning step, the AI model is subjected to machine learning using information collected when constructing tunnels other than the tunnel to be predicted. In said additional learning step, additional learning is performed on said initially trained AI model using information collected when constructing a part of said tunnel to be predicted. In said prediction processing step, said displacement amount is predicted by inputting information obtained when constructing the remaining portion of said tunnel to be predicted into said additionally trained AI model. According to this configuration, it is possible to predict the displacement amount reflecting the condition of the natural ground to be predicted while reducing the information collected from the tunnel to be predicted. Effects of the Invention

[0009] According to the present invention, it is possible to accurately predict a displacement amount while reducing the labor required for analysis compared to conventional techniques. Brief Description of the Drawings

[0010] [Figure 1] It is a configuration diagram of a tunnel construction support system according to an embodiment of the present invention. [Figure 2] It is an example of contents recorded in a face observation record. [Figure 3] It is a diagram for explaining accuracy verification of the present invention, showing the relationship between explanatory variables and objective variables given to an AI (Artificial Intelligence) model. [Figure 4] It is a diagram for explaining accuracy verification of the present invention, showing the number of measurement cross-sections at the site used in the accuracy verification. [Figure 5] It is a diagram for explaining accuracy verification of the present invention, showing the frequency distribution of displacement amounts measured at all sites. [Figure 6] It is a diagram for explaining the learning method used in accuracy verification of the present invention, where (a) is a schematic diagram of learning method I, (b) is a schematic diagram of learning method II, and (c) is a schematic diagram of learning method III. [Figure 7] This figure shows the results of the verification using learning method I. [Figure 8] This figure shows the results of the verification using learning method II. [Figure 9] This figure shows the results of the verification using learning method III. [Figure 10] This is the result of the importance analysis in the accuracy verification of the present invention. [Modes for carrying out the invention]

[0011] Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings as appropriate. Each figure is only a schematic representation to the extent that the present invention can be fully understood. Therefore, the present invention is not limited to the illustrated examples. In each figure, common or similar components are denoted by the same reference numerals, and their redundant descriptions are omitted.

[0012] <Configuration of the tunnel construction support system according to the embodiment> Referring to Figure 1, the configuration of the tunnel construction support system 1 according to the embodiment will be described. Figure 1 is a configuration diagram of the tunnel construction support system 1 according to the embodiment. The tunnel construction support system 1 is a system that supports the smooth progress of tunnel construction by predicting the amount of displacement that occurs on the inner surface of the tunnel. The tunnel construction support system 1 can be used for various types of tunnels without being limited to any particular type. In this embodiment, a mountain tunnel will be used as an example to explain the system. The tunnel construction support system 1 mainly comprises a displacement learning device 10 and a displacement prediction device 20. If the displacement learning device 10 itself predicts the tunnel displacement, the displacement learning device 10 and the displacement prediction device 20 can be configured as a single device.

[0013] The displacement learning device 10 is a device that learns the relationship between information about the tunnel face and the amount of displacement occurring on the inner surface of the tunnel. Information about the tunnel face includes (1) observation and evaluation of the tunnel face, (2) support stiffness, (3) overburden, and (4) excavation diameter. The displacement learning device 10 learns using all or part of (1) observation and evaluation of the tunnel face, (2) support stiffness, (3) overburden, and (4) excavation diameter, along with the amount of displacement on the inner surface of the tunnel. Information about the tunnel face may include pre-construction survey data and construction data obtained during construction, and may be information collected during pre-construction surveys or construction. The displacement prediction device 20 is a device that predicts the amount of displacement occurring on the inner surface of the tunnel based on information about the tunnel face. The items of information about the tunnel face used by the displacement prediction device 20 correspond to the items used in the learning of the displacement learning device 10. For example, if "observation and evaluation of the tunnel face" and "overburden" were used in the learning of the displacement learning device 10, the information used by the displacement prediction device 20 to make predictions will be "observation and evaluation of the tunnel face" and "overburden". Furthermore, the amount of displacement predicted by the displacement prediction device 20 is a value corresponding to the amount of displacement used in the learning of the displacement learning device 10. For example, if the final displacement of the inner surface of the tunnel was used in the learning of the displacement learning device 10, the value predicted by the displacement prediction device 20 will be the final displacement.

[0014] (Configuration of the displacement learning device) As shown in Figure 1, the displacement learning device 10 comprises a learning data acquisition unit 11 and a learning processing unit 12. The learning data acquisition unit 11 and the learning processing unit 12 are implemented, for example, by program execution processing. When these functions are implemented by program execution processing, the program for implementing these functions is stored in a storage unit (not shown). The learning processing unit 12 is equipped with an AI (Artificial Intelligence) model. The AI ​​model can derive better results by comparing the results of classification, prediction, and judgment based on the given data with the actual correct answer and adjusting various parameters. The AI ​​model is also called a "learning model" or "learner." In this embodiment, we will explain using "MST (Memory Saving Tree)" as the AI ​​model.

[0015] MST is a tree-structure-based AI algorithm that makes predictions by automatically learning branching conditions based on the learned data. MST is lightweight, highly accurate, and allows for additional learning. For example, AI algorithms such as "Neural Network" and "Random Forest" require implementation on CPUs (Central Processing Units) or GPUs (Graphics Processing Units) with abundant computing resources, but MST can be implemented on microcontrollers, making it lightweight and fast. Compared to multiple regression analysis, MST's main advantages are its ability to perform accurate inference with less effort even for nonlinear correlations, and its ability to adapt to data with different characteristics through additional learning. Additional learning is a method of adapting an AI model that has completed initial learning by sequentially providing additional data. Many AI models have the problem of "forgetting" the initial learning during additional learning, but MST has a mechanism that allows for additional learning while retaining the initial learning. It should be noted that "MST" is just one example of an AI model, and the algorithm of the AI ​​model is not particularly limited. Other algorithms (including the multiple regression analysis, neural networks, and random forests described here) may also be used as AI models.

[0016] The learning data acquisition unit 11 shown in Figure 1 acquires (including creation) learning data for training the AI ​​model. The learning data is a set of explanatory variables and an objective variable. Here, the explanatory variables are information about the tunnel face, and the objective variable is the displacement of the inner surface of the tunnel. Therefore, the learning data is composed of, for example, a set of explanatory variables such as observation and evaluation of the tunnel face, support stiffness, overburden, and excavation diameter, and the objective variable, the displacement of the inner surface of the tunnel. It is desirable that the correspondence between the items constituting the learning data be strict, but slight discrepancies in the position of the tunnel face between each item are acceptable. The learning data is preferably created based on construction results from similar tunnel construction. Alternatively, the observation and evaluation of the tunnel face, support stiffness, overburden, excavation diameter, and displacement of the inner surface of the tunnel may be pre-registered in the memory unit, and the learning data acquisition unit 11 may acquire the registered information as learning data. The memory unit may be provided by a device connected via a communication line (for example, a cloud system).

[0017] <(1) Observation and evaluation of the tunnel face> Referring to Figure 2, the observation and evaluation of the tunnel face will be explained. Figure 2 shows an example of the contents recorded in the tunnel face observation log. In this embodiment, the information recorded in the tunnel face observation log is assumed to be the observation and evaluation of the tunnel face. As shown in Figure 2, the face observation log records an evaluation category for each observation item. There are 10 observation items, for example: "A: Face stability", "B: Condition of the uncut surface", "C: Compressive strength", "D: Weathering and alteration", "E: Percentage of fractured area in the face", "F: Crack spacing", "G: Condition of cracks", "H: Morphology of cracks", "I: Water seepage (visually observed amount)", and "J: Water deterioration". Note that the face observation log shown in Figure 2 is an example, and the observation items are not limited to those shown in Figure 2. The evaluation categories range from "1" as the minimum value to "4-6" as the maximum value (which varies depending on the observation item).

[0018] For example, regarding the observation items for face stability, "1" is registered as the evaluation category when the face is stable; "2" is registered as the evaluation category when rock blocks fall off from the face; "3" is registered as the evaluation category when extrusion of the face occurs; and "4" is registered as the evaluation category when the face cannot stand on its own and collapses or flows out. Further, regarding the fracture morphology, "1" is registered as the evaluation category in the case of random blocky morphology; "2" is registered as the evaluation category in the case of columnar morphology; "3" is registered as the evaluation category in the case of layered, flaky or tabular morphology; and "4" is registered as the evaluation category in the case of earthy-sandy, fragmented or unconsolidated morphology. Note that although descriptions of evaluation categories for "B: State of bare excavation surface", "C: Compressive strength", "D: Weathering alteration", "E: Proportion of fractured parts in the face", "F: Fracture spacing", "G: State of fractures", "I: Spring water (visually estimated volume)", and "J: Deterioration by water" are omitted herein, they are as shown in Figure 2.

[0019] <Support stiffness> The support stiffness is calculated based on the support pattern, for example, by the following formula (1). K=K C +K S ··Formula (1) K: support stiffness, K C : shotcrete stiffness, K S : steel support stiffness Further, K C is calculated by the following formula (2), and K S is calculated by the following formula (3). K C ={E C (a 2 -a i 2 )} / [(1+ν){(1-2ν)a 2 +a i 2}]··Formula (2) E C : modulus of elasticity of shotcrete, ν: Poisson's ratio of shotcrete, a: excavation radius, a i : radius to the inner side of the shotcrete K S =ES A S / S S a...Equation (3) E S : Elastic modulus of steel support structure, A S : Cross-sectional area of ​​steel support structure, S S :Spacing of steel support structures, a:Excavation radius

[0020] <Soil cover> Overburden is the height from the top of a tunnel (underground structure) constructed underground to the ground surface directly above it. The overburden value is obtained, for example, from the results of a prior geological survey (such as a longitudinal section). <Drilling diameter> The excavation diameter may be the excavation width at the tunnel SL position, twice the upper half excavation radius, or the diameter of a circle equivalent to the excavation cross-sectional area. However, it is desirable to use a value calculated using a unified method from the above as training data. <Displacement amount> Displacement amounts refer to the amount of settlement and internal displacement within the tunnel caused by excavation. These displacement amounts are generally obtained through settlement measurements and internal displacement measurements, which are carried out as part of tunnel measurement and management. For example, the displacement amount used as training data is the value at the point when the measurement cross-section has moved sufficiently away from the tunnel face, the effects of excavation have ceased, and the displacement has converged (hereinafter, this displacement amount will be referred to as the final displacement amount).

[0021] The learning processing unit 12 shown in Figure 1 trains an AI model using training data consisting of a set of information about the tunnel face and the displacement amount of the inner surface of the tunnel. Specifically, it trains the AI ​​model using information about the tunnel face as an explanatory variable and the displacement amount of the inner surface of the tunnel as the objective variable. In other words, the learning processing unit 12 trains the AI ​​model to output the displacement amount of the inner surface of the tunnel by inputting information about the tunnel face. The method of training the AI ​​model is not particularly limited, and various methods can be used. For example, (I) training using data other than the prediction point of the tunnel in question, (II) training using data from other sites other than the tunnel in question, and (III) training using both data other than the prediction point of the tunnel in question and data from other sites other than the tunnel in question. The number of training data to be used is not particularly limited, and training is terminated when, for example, the expected accuracy (correctness rate) is reached. The trained AI model is sent to the displacement amount prediction device 20 and stored.

[0022] (Configuration of the displacement prediction device) As shown in Figure 1, the displacement prediction device 20 comprises a prediction data acquisition unit 21 and a prediction processing unit 22. The prediction data acquisition unit 21 and the prediction processing unit 22 are implemented, for example, by program execution processing. When these functions are implemented by program execution processing, the program for implementing these functions is stored in a storage unit (not shown). The prediction data acquisition unit 21 acquires prediction data that forms the basis for predicting displacement. The prediction data is information about the tunnel face at the prediction location and consists of the same items as the training data. The prediction data is composed of, for example, a set of (1) observation and evaluation of the tunnel face, (2) support stiffness, (3) overburden, and (4) excavation diameter. It is desirable that the correspondence between the items that make up the prediction data be strict, but there may be some slight discrepancies in the position of the tunnel face between each item. The information of (1) observation and evaluation of the tunnel face, (2) support stiffness, (3) overburden, and (4) excavation diameter that makes up the prediction data may be measured or otherwise obtained using the same methods as the training data.

[0023] The prediction processing unit 22 uses information about the tunnel face at the predicted location (prediction data) to predict the amount of displacement occurring on the inner surface of the tunnel. The prediction processing unit 22 has a pre-trained AI model. The pre-trained AI model was trained using the displacement amount learning device 10. In other words, the pre-trained AI model is trained to output the amount of displacement occurring on the inner surface of the tunnel when information about the tunnel face is input. To put it another way, the prediction processing unit 22 takes information about the tunnel face as an explanatory variable and outputs the amount of displacement on the inner surface of the tunnel as the objective variable.

[0024] As described above, in the tunnel construction support system 1 according to this embodiment, the displacement amount learning device 10 uses learning data that includes all or part of the observation evaluation of the tunnel face, support stiffness, overburden, and excavation diameter to train an AI model. Then, the displacement amount prediction device 20 uses the trained AI model to predict the trend of displacement occurring on the inner surface of the tunnel. According to the tunnel construction support system 1, it is possible to predict displacement amounts that reflect the state of the ground without performing complex analysis, and construction management based on that prediction can be performed in real time.

[0025] The effects of the tunnel construction support system 1 according to the embodiment will be explained with reference to Figures 3 to 10. The inventors of the present invention trained an AI model using actual tunnel construction results and measurement data, and verified its accuracy by taking the RMSE (mean squared deviation) between the predicted values ​​and actual measured values ​​obtained by the trained AI model. For accuracy verification, MST was used as the AI ​​model. The explanatory and dependent variables given to the AI ​​model during training are shown in Figure 3. For accuracy verification, the following 10 items were used as observational evaluations of the tunnel face, as shown in Figure 2: "A: Face stability", "B: Condition of the uncut surface", "C: Compressive strength", "D: Weathering and alteration", "E: Proportion of fractured area in the face", "F: Crack spacing", "G: Condition of cracks", "H: Morphology of cracks", "I: Water seepage (amount observed visually)", and "J: Deterioration due to water".

[0026] Furthermore, for accuracy verification, measurement and observation data from three sites (Site 1, Site 2, and Site 3) shown in Figure 4 were used. Sites 1 through 3 are construction sites located several kilometers apart but in the same geological age, and Sites 2 and 3 are in a continuous section. The geology of Site 1 is mainly slate, while the geology of Sites 2 and 3 is mainly slate and sandstone. The number of measurement cross-sections measured at each site is shown in Figure 4. The number of measurement cross-sections at Site 1 was "124", the number of measurement cross-sections at Site 2 was "43", and the number of measurement cross-sections at Site 3 was "44". The number of measurement cross-sections was highest at Site 1, approximately three times the number at Sites 2 and 3. Figure 5 shows the frequency distribution of displacement amounts measured at all sites. The frequency distribution shown in Figure 5 shows the frequency of final displacement amounts in "8 mm" units. As can be seen in Figure 5, there are many data points in the "8-24 mm" range, and relatively few data points of "40 mm" or more. Although not shown in the illustration, there is a difference in that there are no data points of "40 mm" or more at site 2, while there are a certain number of data points of "40 mm" or more at sites 1 and 3, and also data points of "80 mm" or more. Due to the large number of data points at site 1 compared to the other sites (sites 2 and 3), the distribution shown in Figure 5 shows a similar trend to the distribution at site 1.

[0027] For this accuracy verification, the AI ​​model was trained using the learning methods I to III shown in Figure 6. The details of each learning method are as follows: Learning Method I: Learn from all data except for the predicted locations and perform cross-validation. Learning Method II: Learn and cross-validate using only data from two other locations besides the predicted location. Learning Method III: In addition to Learning Method II, data from the prediction field is sequentially added for learning and cross-validation. Figures 7 to 9 show the results of verification using learning methods I to III for all data from three sites (sites 1 to 3). Figure 7 shows the results of verification using learning method I, Figure 8 shows the results of verification using learning method II, and Figure 9 shows the results of verification using learning method III. In Figures 7 to 9, the vertical axis represents the predicted value by the AI ​​model, and the horizontal axis represents the measured value. In Figures 7 to 9, the straight line extending from the bottom left to the top right represents 100% prediction accuracy, and since the vertical and horizontal scales in Figures 7 to 9 are "1:1", this straight line forms a 45-degree angle. In Figures 7 to 9, data below the 45-degree line indicate that the measured value was smaller than the predicted value, and data above the 45-degree line indicate that the measured value was larger than the predicted value.

[0028] The results of the verification of learning method I are shown in Figure 7. The RMSE when using learning method I is "9.75 mm," indicating that the predicted values ​​generally capture the trend of the measured values. On the other hand, it can be seen that for some individual data points, the predicted values ​​are about 0.5 or 3 times the measured values. The verification results for Learning Method II are shown in Figure 8. The RMSE when using Learning Method II is "17.25 mm," indicating that the predicted values ​​do not capture the trend of the measured values ​​in the range of large measured displacements (the predicted values ​​are smaller than the measured values). In particular, the prediction accuracy is low for data with large measured displacements (data of approximately 40 mm or more). As shown in Figure 5, this is likely because there is little data with large displacements in the training data, and therefore the learning of data in that range is not effectively reflected in the AI ​​model. The verification results for Learning Method III are shown in Figure 9. The RMSE when using Learning Method III is "11.01 mm," which is not as good as Learning Method I, but it is a significant improvement in accuracy compared to Learning Method II. From these results, it can be seen that the trend of displacement can be captured using a method that is applicable to actual construction sites, where data from other sites is used for initial training, and then construction results are used for additional training.

[0029] Furthermore, to analyze the characteristics of the data, an importance analysis of each explanatory variable was performed using an AI model trained on data from all construction sites. The results of the importance analysis are shown in Figure 10. In the data used for this accuracy verification, it was found that the importance of displacement prediction was highest in the following order: "compressive strength (observation item "C" in the face observation log)", "soil cover", "crack condition (observation item "G" in the face observation log)", "excavation diameter", and "support stiffness". This importance analysis result also shows that the tunnel construction support system 1 according to the embodiment is able to predict displacement amounts that capture the trends of measured values. In other words, generally, the deformation δ of a tunnel is proportional to the load P (=overburden) and the excavation diameter R, while it is inversely proportional to the stiffness Eg of the ground (a value affected by compressive strength and cracks) and the support stiffness Es (see equation (4)). →δ∝(P×R) / (Eg×Es)...Equation (4) The importance analysis results also show that these parameters are highly important, which is consistent with the mechanical considerations for the tunnel. Therefore, it can be inferred that the method described in the embodiment can be used to predict displacement amounts that capture the trends of measured values.

[0030] The inventor conducted importance analyses when training the model with only data from each of the three sites (Site 1 to Site 3). In Site 1, the importance of "crack condition (observation item "G" in the face observation log)" was higher than in Sites 2 and 3. While Sites 2 and 3 showed similar results in their importance analyses, Site 3 showed a significantly higher importance of "compressive strength (observation item "C" in the face observation log)." In Site 1, the importance of "crack spacing (observation item "F" in the face observation log)" was relatively high. In Site 2, the importance of "proportion of fractured area in the face (observation item "E" in the face observation log)" and "crack spacing (observation item "F" in the face observation log)" was relatively high. Furthermore, at the third site, the importance of "weathering and alteration (observation item "D" in the face observation log)" and "the proportion of fractured material in the face (observation item "E" in the face observation log)" was relatively high. From this, it can be seen that it is possible to predict the displacement of various tunnels by using observation items "A-J" in the face observation log, and that by appropriately selecting observation items "A-J" in the face observation log according to the characteristics of the tunnel, it is possible to make more accurate predictions. [Explanation of Symbols]

[0031] 1. Tunnel Construction Support System 10 Displacement learning device 11. Training Data Acquisition Unit 12 Learning Processing Unit 20 Displacement prediction device 21 Predictive Data Acquisition Unit 22 Prediction Processing Unit

Claims

1. A displacement learning device for learning the amount of displacement occurring on the inner surface of a tunnel, comprising a learning data acquisition unit and a learning processing unit, The aforementioned learning data acquisition unit acquires a set of data as learning data, including the observation and evaluation of the tunnel face, the support stiffness, the overburden and the excavation diameter, and the displacement of the tunnel face. The learning processing unit has an AI model, and uses the learning data to train the AI ​​model so that it outputs the displacement amount by inputting all of the observation evaluation, the support stiffness, the soil cover, and the excavation diameter. The aforementioned observational evaluation includes at least an evaluation score related to compressive strength, A displacement learning device characterized by the following features.

2. The aforementioned observation and evaluation further includes evaluation points for at least one of the following observation items: stability of the tunnel face, condition of the unexcavated surface, weathering and alteration, proportion of fractured areas in the face, crack spacing, crack condition, crack morphology, seepage, and deterioration due to water. Displacement amount learning device according to feature 1.

3. A displacement learning method for learning the amount of displacement occurring on the inner surface of a tunnel, comprising a learning data acquisition step and a learning processing step, In the aforementioned learning data acquisition process, a set of observational evaluations of the tunnel face, support stiffness, overburden, and excavation diameter, along with the displacement of the tunnel face, is acquired as learning data. In the learning process, the AI ​​model is trained using the learning data so that it outputs the displacement amount by inputting all of the observation evaluation, support stiffness, soil cover, and excavation diameter. The aforementioned observational evaluation includes at least an evaluation score related to compressive strength, A method for learning displacement, characterized by the following features.

4. The prediction processing unit comprises a trained AI model that has been trained by the displacement learning method described in claim 3, The prediction processing unit predicts the amount of displacement of the tunnel face at the predicted location by inputting all of the observation and evaluation of the tunnel face at the predicted location, the support stiffness, the overburden, and the excavation diameter into the trained AI model. A displacement prediction device characterized by the following features.

5. The method includes a prediction process step that predicts the displacement amount using the trained AI model trained by the displacement amount learning method described in claim 3, In the prediction processing step, the amount of displacement of the tunnel face at the predicted position is predicted by inputting all of the observation and evaluation of the tunnel face at the predicted position, the support stiffness, the overburden, and the excavation diameter into the trained AI model. A method for predicting displacement, characterized by the following features.

6. The learning process comprises an initial learning process and an additional learning process. In the initial learning process, the AI ​​model is trained using information collected when tunnels other than the target tunnel are constructed. In the aforementioned additional learning process, the AI ​​model that has already been initially trained is further trained using information collected when a portion of the tunnel to be predicted was constructed. In the prediction processing step, the amount of displacement is predicted by inputting information about the construction of the remaining part of the tunnel to be predicted into the AI ​​model that has been further trained. The displacement prediction method according to feature 5.

Images