JP7544634B2 - Weighted search method for tunneling prediction model - Google Patents

Description

The present invention relates to a weighting search method for an excavation prediction model, and in particular to a weighting search method for an excavation prediction model for determining the weighting of explanatory variables, which is used when creating an excavation prediction model for a shield tunneling machine through machine learning using artificial intelligence.

The shield tunneling method involves excavating the ground with a rotating cutter while pressing down and stabilizing the face at the tip of a known earth pressure or slurry type shield machine using mud soil pressure, muddy water pressure, etc., and forming a tunnel underground from the starting shaft to the arrival shaft while assembling a tunnel lining made of segments ring by ring behind the shield machine. This method is widely adopted for major tunnel construction in urban areas and plains.

Shield tunneling machines, such as earth pressure and slurry types, used in the shield tunneling method are equipped with a rotating cutter that cuts the tunnel face, a bulkhead, a cutter drive device, an earth removal mechanism, etc., at the front of a metal outer shell called a skin plate, and a shield jack and an erector device, etc., at the rear of the skin plate. The erector device is used to assemble the tunnel lining made of segments ring by ring, and while receiving a reaction force from the assembled tunnel lining, the shield jack pushes out the rotating cutter together with the skin plate, thereby excavating the shield tunnel while cutting the tunnel face.

In this type of shield tunneling method, it is necessary to precisely control the shield machine to excavate along the excavation plan extension line set underground, but because construction is carried out underground, various factors such as changes in the geological quality of the surrounding ground, changes in environmental conditions, and the characteristics of each shield machine make it difficult to control the operation so that the machine excavates precisely along the excavation plan extension line, and generally horizontal, vertical, and directional deviations from the excavation plan extension line occur at the tip of the shield machine. In particular, various factors resulting from the characteristics of each shield machine are called the "habits" of the shield machine at each construction site, and are difficult to grasp accurately. Conventionally, the shield machine was controlled by experienced workers using their experience and intuition to control operation so that the deviation from the excavation plan extension line would not become too large.

Furthermore, in recent years, the use of large amounts of data and AI based on technological innovations in IoT (Internet of Things) and AI (Artificial Intelligence) is expected to bring about the Fourth Industrial Revolution, and in the field of shield tunneling, for example, a shield tunneling planning support system that utilizes AI has been developed (see, for example, Non-Patent Document 1). In the shield tunneling planning support system of Non-Patent Document 1, a reinforcement learning method is used in which the AI self-learns through trial and error to arrive at an optimal solution, making it possible to derive planned values for the operation of the shield machine according to the linear shape of the shield tunnel and a segment placement plan.

Developing a shield tunneling planning support system using AI / Corporate information / Shimizu Corporation, May 25, 2018, [Retrieved June 10, 2019], Internet (URL: https://www.shimz.co.jp/company/about/news-release/2018/2018005/html)

Patent Publication No. 2021-14727

However, the shield tunneling planning support system in Non-Patent Document 1 merely performs a preliminary simulation of the operation method of the shield machine for the planned linear alignment and the allocation method of multiple segments of different shapes prior to actually carrying out shield tunneling work at the construction site, and sets these planned values. Furthermore, the data items (parameters) of the training data for creating the trained model are limited and specific, so it is difficult to obtain an accurate prediction model that reflects the characteristics of the shield machine, known as its "habits," which are due to various uncertain factors at each construction site, such as changes in the geology of the surrounding ground, changes in environmental conditions, or the characteristics of each shield machine, and to properly manage the shield machine so that it excavates along the planned extension line set underground.

Meanwhile, shield tunneling methods often use shield tunneling management systems to get a detailed understanding of the tunneling status as the shield machine progresses. Shield tunneling management systems are well-known systems that collect various data, such as survey data from shield tunneling work and measurement data from various sensors installed on the shield machine, and centralize the management of the shield machine. They are able to estimate the condition of the excavated soil in the ground and the load on the shield machine based on the results of changes over time in the measurement data during construction and statistical processing of the data, and by inputting the survey results, it is possible to calculate the position of the shield machine and segments and determine the deviation from the tunneling plan extension line.

In addition, in the shield tunnel excavation management system, when excavation is carried out corresponding to each ring of the lining body made of segments that form the shield tunnel, data on numerous parameters (data items) is collected, for example, at time intervals of about 5 seconds, or for each jack stroke of about 10 mm by the shield jack, and stored as a large amount of data. By analyzing this large amount of data using AI, it is thought that it will be possible to appropriately manage the shield tunneling machine so that it excavates along the excavation plan extension line set underground, reflecting the characteristics of the shield tunneling machine, known as its "habits," which are due to various uncertain factors resulting from changes in the geology of the surrounding ground at each construction site, changes in environmental conditions, or the characteristics of each shield tunneling machine.

For this reason, the applicant of the present application has proposed, for example, in Patent Document 1, a construction management method for a shield machine that uses the results of machine learning by artificial intelligence using construction measurement data sent from a shield excavation management system as teacher data to manage the excavation of the shield machine. In the construction management method for a shield machine in Patent Document 1, a learning area is set on the excavation plan extension line of the shield machine at the construction site, and in the set learning area, a learned model is created by machine learning with construction measurement data when constructing one or more rings on the rear side of the excavation direction as explanatory variables and the deviation of the tip of the shield machine when constructing a specified ring on the front side of the excavation direction as the objective variable. In the main excavation area forward of the excavation direction from the learning area, the created learned model is used as a prediction model, and construction measurement data when constructing one or more rings on the rear side of the excavation direction that have been constructed is input, and the deviation of the tip of the shield machine when constructing a specified ring on the front side of the excavation direction is output, so that the excavation of the shield machine is managed while predicting the deviation of the tip of the shield machine from the excavation plan extension line.

In addition, in Patent Document 1, data related to selected specified parameters is extracted from the construction measurement data measured in one or more rings on the rear side of the excavation direction in the learning area, and a prediction model is created in which the data related to the extracted specified parameters is used as an explanatory variable and the deviation of the tip of the shield machine when constructing a specified ring on the front side of the excavation direction is used as the objective variable. In this excavation area, construction measurement data related to the selected specified parameters when constructing one or more rings on the rear side of the excavation direction is input to the created prediction model, and the deviation of the tip of the shield machine when constructing a specified ring on the front side of the excavation direction is output.

Furthermore, in Patent Document 1, the selected predetermined parameters are created by replacing the specific parameters from the construction measurement data when constructing one or more rings on the rear side of the excavation direction in the learning area with data related to specific parameters deleted as the confirmation explanatory variable, and creating multiple parameter confirmation models with the deviation of the tip of the shield machine when constructing a specified ring on the front side of the excavation direction as the confirmation objective variable by replacing the specific parameters from which the data is deleted, and including confirmed parameters selected by comparing the output data of the deviation of the tip of the shield machine from the trained model with the output data of the deviation of the tip of the shield machine from each parameter confirmation model. This makes it possible to create a prediction model using teacher data that removes parameters that have a negative effect on the accuracy of a prediction model that uses the deviation of the tip of the shield machine as the objective variable, and parameters that do not affect the accuracy, and to more efficiently and accurately manage the shield machine so that it excavates along the excavation plan extension line set in the ground.

In the construction management method for a shield tunneling machine in Patent Document 1, the accuracy of the prediction model can be improved by narrowing down the parameters that are explanatory variables of the training data used to create the prediction model to those that are thought to have an effect on the accuracy of the prediction model. However, in order to further improve the accuracy of the prediction model, it is desirable to create an excavation prediction model that predicts the deviation of the tip of the shield tunneling machine through machine learning using artificial intelligence, with appropriate weighting set on the parameters that are the selected explanatory variables.

However, even when the parameters that serve as explanatory variables are narrowed down to those that are thought to affect the accuracy of the predictive model, the number of parameters is still considerable, and therefore the number of confirmation models that need to be created to determine the appropriateness of the weighting of each parameter becomes enormous, which makes machine learning using artificial intelligence require a lot of time and effort. For this reason, there is a need for the development of technology that can more efficiently determine the appropriateness of the weighting set for the parameters that are sent from the shield tunneling management system and serve as explanatory variables in the shield tunneling method, thereby further improving the accuracy of predictive models created by machine learning using artificial intelligence.

The present invention aims to provide a method for searching weightings for an excavation prediction model that can more efficiently determine the appropriateness of weightings set for parameters sent from a shield excavation management system and serving as explanatory variables in a shield excavation method, thereby further improving the accuracy of a prediction model created by machine learning using artificial intelligence.

The present invention is a method for searching for weighting of an excavation prediction model for determining the weighting of explanatory variables, which is used in creating an excavation prediction model that predicts the deviation of the tip of a shield machine by machine learning using artificial intelligence, using construction measurement data sent from a shield excavation management system as teacher data in a shield tunneling method. The excavation prediction model is created as a learned model in which the construction measurement data when constructing one or more rings on the rear side of the excavation direction is used as the explanatory variable, and the deviation of the tip of the shield machine when constructing a specified ring on the front side of the excavation direction is used as the objective variable. The search for the weighting of explanatory variables includes a step of setting a weighting base value and a weighting coefficient, and a step of creating multiple weighting search models in which multiple data sets using data weighted by the set weighting base value and weighting coefficient are used as explanatory variables, comparing the output data of the deviation of the tip of the shield machine from each of the created multiple weighting search models, and searching for the weighting of explanatory variables from the data set used in the weighting search model to be adopted. In the step of setting the weighting base value and the weighting coefficient, The above-mentioned objective has been achieved by providing a weighting search method for an excavation prediction model, which divides predetermined selection data related to the direction prediction of a shield tunneling machine, selected from the data, into a plurality of search groups according to the degree of contribution of the direction prediction, sets the basic weighting value for each selection data of each grouped search group, and sets a predetermined number of patterns of the weighting coefficients for each search group, which are predetermined values in the search range of 0.0 to 1.0. In the process of searching for the weighting of the explanatory variables, a plurality of weighting search models are created using the data sets in a number obtained by multiplying the number of patterns of the weighting coefficients in the search range set for each search group, and compares the output data of the deviation of the tip of the shield tunneling machine by the plurality of weighting search models using the data sets in a number obtained by multiplying the number of patterns of the weighting coefficients in the search range set for each search group with actual measurements, and determines the weighting of the explanatory variables from the data sets used in the weighting search model to be adopted.

In the weighting search method for an excavation prediction model of the present invention, the weighting base value and the weighting coefficient set in the step of setting the weighting base value and the weighting coefficient are values set based on the knowledge of a worker who is a shield engineer, and it is preferable that the greater the contribution to the directional prediction, the closer the weighting base value and the weighting coefficient are to 1.0.

In addition, it is preferable that the weighting search method for the excavation prediction model of the present invention is such that the predetermined selection data related to the directional prediction of the shield tunneling machine selected from the construction measurement data is grouped into three search groups, the first group, the second group, and the third group, according to the degree of contribution of the directional prediction, and the weighting basic value is set to be larger in the order of the first group, the second group, and the third group.

Furthermore, in the weighted search method for excavation prediction models of the present invention, five weighting coefficients, namely, 0.2, 0.4, 0.6, 0.8, and 1.0, are set for the search groups of the first group, three weighting coefficients, namely, 0.4, 0.7, and 1.0, are set for the search groups of the second group, and two weighting coefficients, namely, 0.5 and 1.0, are set for the search groups of the third group, and a plurality of weighted search models are created using 30 of the data sets, the number being the product of the number of patterns of the weighting coefficients in the search range set for each search group.

Furthermore, in the weighted search method for an excavation prediction model of the present invention, it is preferable that the machine learning is machine learning using a neural network that uses a support vector machine as an algorithm.

In addition, in the weighted search method for the excavation prediction model of the present invention, it is preferable that the deviation of the tip of the shield tunneling machine, which is the objective variable, is a horizontal deviation, a vertical deviation, and a directional deviation.

The weighting search method for an excavation prediction model of the present invention makes it possible to more efficiently determine the appropriateness of the weighting applied to parameters (selected data) that serve as explanatory variables sent from the shield excavation management system in a shield excavation method, thereby further improving the accuracy of the prediction model created by machine learning using artificial intelligence.

1 is an explanatory diagram of a system network in which a weighted search method for an excavation prediction model according to a preferred embodiment of the present invention is implemented. FIG. FIG. 1 is an explanatory diagram of the trained model creation process. FIG. 13 is an explanatory diagram of a ring report. FIG. 1 is an explanatory diagram illustrating a plurality of variables (parameters) that serve as training data for creating a trained model. FIG. 4 is an explanatory diagram of the positional relationship of construction data.

A weighted search method for an excavation prediction model according to a preferred embodiment of the present invention is implemented in a system network having a configuration as shown in FIG. 1. The system network shown in FIG. 1 uses a shield excavation management system 10 that manages the construction of a shield tunneling machine 20 (see FIG. 2) at a shield construction site, and preferably a cloud server 11 that is connected via a LAN 12 and has a known machine learning tool (software) that can implement a neural network with a support vector machine as an algorithm as AI (artificial intelligence), to create a prediction model that uses the deviation of the tip of the shield tunneling machine 20 as a target variable, for example, so that the shield tunneling machine 20 can be appropriately managed so that it excavates along an excavation plan extension line 21 (see FIG. 2) set underground. The weighting search method for an excavation prediction model of this embodiment is preferably implemented by machine learning using AI in a system network with the configuration shown in FIG. 1, using a large amount of data (construction measurement data) related to a large number of parameters sent from a shield excavation management system 10 at a shield construction site and stored in a cloud server 11, and enables more efficient determination of the appropriateness of the weighting set for selected, predetermined parameters (selected data) that serve as explanatory variables, thereby further improving the accuracy of the prediction model created by machine learning using artificial intelligence.

Here, in this embodiment, for example, the product name "Arigataya" (manufactured by Calculation Studio Co., Ltd.) can be preferably used as a known shield excavation management system 10. The cloud server 11 is configured to include, for example, a computer. The computer includes a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), an I/F (Interface), a HDD (Hard Disk Drive), a storage means, an input means, a display means, an output means, etc. The CPU controls machine learning by AI while using the RAM as a work area according to various programs built into the ROM. In addition, the CPU, by having various computer programs built into the ROM, functions the storage means, input means, display means, output means, etc., and stores large amounts of data sent from the shield excavation management system 10 and analysis results by AI, for example, in a database unit, or displays them as specified information on, for example, a display 14 at the site via a LAN 12 or a personal computer 13 installed at the site, or outputs them from a printer.

The weighting search method for the excavation prediction model of this embodiment is a search method for determining the weighting of explanatory variables used in creating an excavation prediction model that predicts the deviation of the tip of the shield tunneling machine 20 by machine learning using AI, using construction measurement data sent from the shield tunneling management system 10 as teacher data in a shield tunneling method. The excavation prediction model preferably uses construction measurement data from the construction of one or more rings N-1 to N-5 (see Figure 2) on the rear side of the excavation direction as explanatory variables, and preferably uses the construction measurement data from the construction of a predetermined ring N to N+5 on the front side of the excavation direction as explanatory variables. A trained model is created with the deviation of the tip as the objective variable, and the search for weighting of the explanatory variables includes the steps of setting a base weighting value (see Table 1) and a weighting coefficient (see Table 2), creating multiple weighting search models with multiple data sets using data weighted by the set base weighting value and weighting coefficient as explanatory variables, comparing the output data of the deviation of the tip of the shield tunneling machine 20 from each of the multiple weighting search models created, and searching for the weighting of the explanatory variables from the data set used in the weighting search model to be adopted.

In the process of setting the basic weighting value and the weighting coefficient, the predetermined selection data related to the directional prediction of the shield tunneling machine 20 selected from the construction measurement data (see Figures 3 and 4) is grouped into a number of search groups A, B, and C according to the degree of contribution to the directional prediction (see Table 1), and a basic weighting value is set for each of the selection data of each of the grouped search groups A, B, and C (see Table 1). In addition, a predetermined number of weighting coefficients are set for each of the search groups A, B, and C, which are predetermined numerical values in the search range of 0.0 to 1.0 (see Table 2).

In the process of searching for weights of explanatory variables, multiple weighting search models are created using data sets in which the number of weighting coefficients in the search range set for each search group A, B, C is multiplied by the number of patterns, and the output data of the deviation of the tip of each shield machine from the multiple weighting search models using data sets in which the number of weighting coefficients in the search range set for each search group A, B, C is multiplied by the number of patterns is compared with the actual measured value, and the weighting of the explanatory variables is determined from the data set used for the weighting search model to be adopted.

In addition, in this embodiment, the predetermined selection data related to the directional prediction of the shield tunneling machine 20 selected from the construction measurement data (see Figures 3 and 4) is grouped into three search groups, the first group A, the second group B, and the third group, according to the degree of contribution of the directional prediction, as shown in Table 1, and the larger weighting basic values are set in the order of the first group A, the second group B, and the third group C.

Furthermore, in this embodiment, as shown in Table 2, four weighting coefficient patterns, 0.2, 0.4, 0.6, 0.8, and 1.0, are set for the first search group A, three weighting coefficient patterns, 0.4, 0.7, and 1.0, are set for the second search group B, and two weighting coefficient patterns, 0.5 and 1.0, are set for the third search group C. In this embodiment, multiple weighted search models are created using 5 x 3 x 2 = 30 data sets, which is the product of the number of weighting coefficient patterns in the search range set for each search group A, B, and C.

Furthermore, in this embodiment, the weighting base value and the weighting coefficient set in the process of setting the weighting base value and the weighting coefficient are values set based on the knowledge of the shield engineer worker, and the greater the contribution to the directional prediction, the closer the value is to 1.0.

In this embodiment, the construction measurement data during tunnel excavation at the shield construction site, which serves as teacher data sent from the shield excavation management system 10, is data on a large number of parameters, for example, about 187 items shown in Figure 4, selected as those related to directional control of the shield tunneling machine 20, preferably based on the knowledge of a shield engineer, from a large number of data items that can be output by the shield excavation management system 10 as "ring reports" (see Figure 3). These data are collected as a set of data for each ring of the lining that makes up the shield tunnel, and are sent to the cloud server 11 preferably every 5 seconds and stored, for example, in a memory unit. That is, in the shield tunneling method, the shield tunneling machine 20 excavates the tunnel face with the rotary cutter 20a at the tip of the shield tunneling machine 20 while extending the shield jack 20b for an excavation length corresponding to the length of one ring, and then the excavation work of the shield tunneling machine 20 is temporarily stopped and one ring's worth of tunnel lining is assembled from the segments 20e using the erector device 20d at the rear of the skin plate 20c. Since excavation work is carried out ring by ring, the construction measurement data is preferably processed as a set of data for each ring.

In this embodiment, the selected construction measurement data of preferably 187 data items (parameters) is classified into a total of 44 pieces of excavation data (related to the front body, middle body, rear body, gyrocompass, coordinates, deviation, displacement, etc.) 50A related to grasping the position of the shield tunneling machine 20, a total of 120 pieces of excavation data (related to the shield jack, center bending jack, copy cutter, etc.) 50B related to the directional control of the shield tunneling machine 20, and a total of 23 pieces of excavation data (related to the recording date and time, cutter bit acceleration, tail clearance, etc.) 50C related to soil quality, etc. These classified data items are collected as data that changes over time by being sent to the cloud server 11 via the shield tunneling management system 10, for example, every 5 seconds, as described above, and are stored in, for example, a memory unit. Note that the horizontal deviation of the tip of the shield tunneling machine 20, the vertical deviation of the tip, and the azimuth deviation (directional deviation) indicated by "*" in FIG. 4 are the objective variables.

These construction measurement data are preferably collected with the ring number of the assembly position of the segment 20e as the title. That is, as shown in FIG. 5, depending on the data item, the position where the data is collected is more than one ring length away from the position of the face at the tip of the shield tunneling machine 20 where the rotary cutter 20a is placed, the position where the backfill material is injected behind the skin plate 20c, and the assembly position of each ring by the segment 20e placed in the expansion and contraction part of the shield jack 20b. The data for positions more than one ring length are also sent to the shield tunneling management system 10 and collected as data during excavation work on the face to ensure the ring width at the assembly position of the ring number, and data during assembly work on the ring at the assembly position of the ring number, with the ring number of the segment assembly position as the title.

Furthermore, in this embodiment, prerequisite data regarding the structure of the shield tunneling machine or the performance of the mechanical equipment is input and registered in advance in the shield tunneling management system 10 or the cloud server 11. Examples of prerequisite data regarding the structure of the shield tunneling machine include the diameter of the cutter face plate, the number and arrangement of shield jacks and bending jacks, etc. These prerequisite data regarding the structure of the shield tunneling machine are not directly input as explanatory variables, but can be combined with construction measurement data during tunnel excavation sent from the shield tunneling management system 10 and normalized by performing feature engineering before being input. Examples of prerequisite data regarding the performance of the mechanical equipment of the shield tunneling machine include the maximum pressure, maximum speed of extension and contraction, and maximum stroke of each shield jack and bending jack. These prerequisite data regarding the performance of the mechanical equipment of the shield tunneling machine are input with the maximum value converted to 1.0, and can be used to normalize and evaluate the construction measurement data during tunnel excavation sent from the shield tunneling management system 10 with values between 0.0 and 1.0.

Furthermore, in this embodiment, the measurement data processed by feature engineering and normalization and input from the construction measurement data during tunnel excavation sent from the shield tunneling management system 10 includes, for example, data related to the operation of the operator, such as data related to the pressure and stroke of the shield jack and the center bending jack, and data related to the cutter rotation. Also, for example, data related to the position and direction of the shield tunneling machine includes data related to the horizontal displacement amount, vertical displacement amount, directional displacement amount, etc. Furthermore, for example, data related to the surrounding environment affecting the shield tunneling machine includes data related to the face water pressure, tail clearance, etc.

In addition, in this embodiment, the construction measurement data during tunnel excavation sent from the shield excavation management system 10 includes data that takes positive and negative values. Examples of such data that takes positive and negative values include data related to the pitching angle and rolling angle of the front body, the pitching angle and rolling angle of the rear body, cutter rotation, cutter torque, etc. For example, in the case of cutter torque, where there is a positive side (right rotation) and a negative side (left rotation), the minimum value after normalization can be -0.5 and the maximum value can be +0.5 (the difference between the minimum and maximum is 1.0). If the maximum value of the cutter torque is 300 kN·m and it rotates right at 150 kN·m, the normalized value can be +0.25.

In this embodiment, at a given shield construction site, the initial construction section set on the excavation plan extension line 21 of the shield tunneling machine 20, for example, an extension of 70 rings (1 ring = 1.5 m), is set as the learning area for the first week (for example, 125 m), and as shown in Figure 2, the construction measurement data from 30 data sets is used as the explanatory variable, which is the number obtained by multiplying the number of patterns of weighting coefficients in the search range set for each search group A, B, C when constructing preferably five rings N-1 to N-5 on the rear side of the excavation direction, and each of the three deviations (horizontal deviation, vertical deviation, directional deviation) of the tip of the shield tunneling machine 20 when constructing each of the six rings, preferably six rings N to N+5 on the front side of the excavation direction, is used as the objective variable. A total of 540 (30 x 6 x 3) trained models are created as weighted search models. Here, among the construction measurement data serving as explanatory variables sent from the shield tunneling management system 10 when constructing preferably five rings N-1 to N-5 on the rear side in the tunneling direction, it is preferable that certain data relating to the structure of the shield tunneling machine 20 or the performance of the mechanical equipment includes data that has been normalized or converted into meaningful synthetic data by characteristic engineering.

In addition, in this embodiment, the three deviations of the tip of the shield tunneling machine 20, which are the objective variables, namely the horizontal deviation, vertical deviation, and directional deviation, can be calculated according to a predetermined formula from data related to predetermined surveying data, such as measurement data from a gyrocompass attached to the shield tunneling machine 20, based on the construction measurement data obtained when constructing each of the six rings N to N+5 on the forward side of the excavation direction, and can be values of the horizontal deviation, vertical deviation, and directional deviation from the planned extension line 21 of the shield tunnel, preferably the center of the cutter at the tip of the shield tunneling machine 20.

Here, in this embodiment, in the process of setting the weighting base value and the weighting coefficient, the predetermined selection data related to the direction prediction of the shield tunneling machine 20 selected from the construction measurement data (see Figures 3 and 4) is grouped into multiple search groups A, B, and C according to the degree of contribution to the direction prediction, and the first group, search group A, belongs to seven data, preferably the horizontal deviation amount (front), horizontal deviation amount (rear), bend angle (left and right), horizontal deviation (rear end), horizontal inclination of the cutter face, horizontal moment, and horizontal overexcavation amount, as shown in Table 1. For these selection data, 0.8 is set as the weighting base value for the horizontal deviation amount (front) and horizontal deviation amount (rear), and 0.5 is set for the bend angle (left and right), horizontal deviation (rear end), horizontal inclination of the cutter face, horizontal moment, and horizontal overexcavation amount, respectively.

Furthermore, of the predetermined selection data related to the direction prediction of the shield tunneling machine 20, the data belonging to the second search group B are preferably four data items: shield jack pressure, thrust, cutter bit acceleration 1X axis, and cutter bit acceleration 1Y axis. In the process of setting the weighting base value and weighting coefficient, 0.3 is set as the weighting base value for each of these four items of selection data.

Furthermore, among the predetermined selection data related to the direction prediction of the shield tunneling machine 20, the data belonging to the third search group C preferably consists of three data items: azimuth deviation, azimuth displacement, and rear body azimuth. In the process of setting the weighting base value and weighting coefficient, 0.2 is set as the weighting base value for each of these three items of selection data.

Of the selected data belonging to the above-mentioned search groups A, B, and C, the horizontal cutter face inclination, horizontal moment, and horizontal overcut amount are composite data converted using characteristic engineering, and the shield jack pressure, thrust, etc. are normalized data.

Furthermore, in this embodiment, in the process of setting the weighting base value and weighting coefficients, as described above, five weighting coefficients of 0.2, 0.4, 0.6, 0.8, and 1.0 are set for the first group of search group A, which has a large contribution to directional prediction, three weighting coefficients of 0.4, 0.7, and 1.0 are set for the second group of search group B, which has a medium contribution to directional prediction, and two weighting coefficients of 0.5 and 1.0 are set for the third group of search group C, which has a small contribution to directional prediction. In the process of searching for weights for explanatory variables described later, multiple trained models are created as weighted search models using 5 x 3 x 2 = 30 data sets, which is the product of the number of weighting coefficient patterns in the search range set for each of the search groups A, B, and C.

In other words, in this embodiment, in the process of searching for weights of explanatory variables, the above-mentioned 30 data sets, including normalized data and synthetic data converted by feature engineering, are used as explanatory variables, and machine learning is preferably performed using a neural network with a support vector machine algorithm. In this way, a total of 540 trained models (30 x 6 x 3 = 540) can be created as models for weighting search, in which the explanatory variables from the 30 data sets for the five rings N-1 to N-5 on the rear side of the excavation direction are used as input data, and each of the three deviations (horizontal deviation, vertical deviation, and directional deviation) of the tip of the shield tunneling machine 20 when constructing each of the six rings N to N+5 on the front side of the excavation direction is used as output data.

In addition, in this embodiment, the above-mentioned selected data selected from the construction measurement data when one or more rings N-1 to N-5 on the rear side of the excavation direction are constructed, which are measured in the above-mentioned learning area at the shield construction site and sent from the shield excavation management system 10 to the cloud server 11 and stored/accumulated, is input as explanatory variables to each of the total 540 weighted search models created, and the three deviations (horizontal deviation, vertical deviation, and directional deviation) of the tip of the shield tunneling machine 20 when a predetermined ring N to N+5 on the front side of the excavation direction is constructed, which is the objective variable, are output. In this way, by comparing the three deviations output from each weighted search model with the actual measured value of the deviation of the tip of the shield tunneling machine 30 that is actually measured and stored in the cloud server 11, the weighted search model that preferably outputs the horizontal deviation, vertical deviation, and directional deviation closest to the actual measured value is selected from these 540 (horizontal deviation: 180 model, vertical deviation: 180 model, directional deviation: 180 model) models. It becomes possible to search for and determine the optimal weighting to be set for each selected data from the weighting set for each selected data that constitutes the data set used to create the selected predetermined weighting search model.

In addition, these weights can be determined by selecting from 180 weighting search models for horizontal deviations, 180 weighting search models for vertical deviations, or 180 weighting search models for directional deviations the model that outputs the horizontal deviation, vertical deviation, or directional deviation that is closest to the actual measured value of each deviation, and searching for and determining the optimal weighting to be set for each selected data from the weightings set for each selected data that constitute the data set used to create each weighting search model.

As a result, according to the weighting search method for an excavation prediction model of this embodiment, it is possible to pre-group selected data and assign weighting base values and weighting coefficients to the grouped selected data, preferably based on the knowledge of shield engineers, without creating a huge number of confirmation models to determine the appropriateness of the weighting to be set for each parameter, thereby making it possible to more efficiently determine the appropriateness of the weighting set for the parameters (selected data) that are sent from the shield excavation management system and serve as explanatory variables, and further improve the accuracy of the prediction model created by machine learning using artificial intelligence.

The present invention is not limited to the above embodiment and can be modified in various ways. For example, the construction measurement data that is the explanatory variable for creating the trained model does not necessarily have to be obtained from the five rings immediately preceding the rear side of the excavation direction, and one or more rings on the rear side of the excavation direction can be appropriately selected and the construction measurement data obtained from these rings can be used. The ring in which the deviation of the tip of the shield tunneling machine, which is the objective variable, is measured does not necessarily have to be one of the six rings immediately following the front side of the excavation direction, and may be the deviation of the tip of the shield tunneling machine in a ring seven rings or more forward. Machine learning for creating the trained model, parameter confirmation model, and prediction model does not necessarily have to be based on a neural network that uses a support vector machine as an algorithm, and may be based on various other algorithms that are capable of analyzing large amounts of construction measurement data.

10 Shield tunneling management system 11 Cloud server 12 LAN
13 Personal computer 14 Display 20 Shield tunneling machine 20a Rotary cutter 20b Shield jack 20c Skin plate 20d Erector device 20e Segment 21 Excavation plan extension line 22 Learning area N to N+5, N-1 to N-5 Ring

Claims (6)

A method for weighting an excavation prediction model for determining weights of explanatory variables, the method being used in creating an excavation prediction model for predicting the deviation of the tip of a shield machine by machine learning using artificial intelligence using construction measurement data sent from a shield excavation management system as teacher data in a shield tunneling method, comprising:
The excavation prediction model is created as a trained model in which the construction measurement data when constructing one or more rings on the rear side of the excavation direction is used as an explanatory variable, and the deviation of the tip of the shield machine when constructing a predetermined ring on the front side of the excavation direction is used as a target variable,
The search for weighting of explanatory variables includes a step of setting a weighting base value and a weighting coefficient, and a step of creating a plurality of weighting search models using a plurality of data sets using data weighted by the set weighting base value and weighting coefficient as explanatory variables, comparing output data of deviation of the tip of each shield machine from the created plurality of weighting search models, and searching for weighting of explanatory variables from the data sets used in the weighting search model to be adopted,
In the step of setting the weighting basic value and the weighting coefficient, predetermined selection data related to the direction prediction of the shield machine, selected from the construction measurement data, is grouped into a plurality of search groups according to the degree of contribution of the direction prediction, the weighting basic value is set for each selection data of each grouped search group, and a predetermined number of patterns of the weighting coefficients, which are predetermined numerical values in a search range of 0.0 to 1.0, are set for each search group;
In the step of searching for weighting of the explanatory variables, a plurality of weighting search models are created using the data set in a number obtained by multiplying the number of patterns of each of the weighting coefficients in the search range set for each search group, and the output data of the deviation of the tip of each of the shield tunneling machines using the plurality of weighting search models using the data set in a number obtained by multiplying the number of patterns of each of the weighting coefficients in the search range set for each search group is compared with actual measured values, and the weighting of the explanatory variables is determined from the data set used for the weighting search model to be adopted. The weighting search method for an excavation prediction model according to claim 1, wherein the weighting base value and the weighting coefficient set in the step of setting the weighting base value and the weighting coefficient are values set based on the knowledge of a shield engineer, and the greater the contribution to the direction prediction, the closer the weighting base value is to 1.0. A weighting search method for an excavation prediction model according to claim 1 or 2, in which the predetermined selection data related to the direction prediction of the shield machine selected from the construction measurement data is grouped into three search groups, the first group, the second group, and the third group, according to the degree of contribution of the direction prediction, and the weighting basic value is set to be larger in the order of the first group, the second group, and the third group. The weighting search method for an excavation prediction model as described in claim 3, wherein five weighting coefficients of 0.2, 0.4, 0.6, 0.8, and 1.0 are set for the search groups of the first group, three weighting coefficients of 0.4, 0.7, and 1.0 are set for the search groups of the second group, and two weighting coefficients of 0.5 and 1.0 are set for the search groups of the third group, and a plurality of weighting search models are created using 30 of the data sets, the number being the product of the number of patterns of the weighting coefficients in the search range set for each search group. The weighted search method for an excavation prediction model according to any one of claims 1 to 4, wherein the machine learning is machine learning using a neural network with a support vector machine as an algorithm. A weighted search method for an excavation prediction model according to any one of claims 1 to 5, in which the deviation of the tip of the shield tunneling machine, which is the objective variable, is a horizontal deviation, a vertical deviation, and a directional deviation.

Images