JP2024089479A - Analysis device, learning device, and program - Google Patents

Abstract

[Problem] The objective is to accurately estimate the amount of charge at a drilling position. [Solution] The analysis device includes an acquisition unit that acquires a support pattern determined in advance for the face position of the natural ground to be excavated in a tunnel, a full face image photographed by a camera of the face position, and acquires, for each drilling position for charging the tunnel face, sensor data obtained when drilling at the drilling position and a partial image of the full face image including the drilling position, and an estimation unit that estimates the amount of charge at the drilling position using a pre-trained model that inputs the support pattern, the sensor data, and the partial image, and outputs the amount of charge at the drilling position. [Selected Figure] Figure 2

Description

The present invention relates to an analysis device, a learning device, and a program.

Generally, the amount of powder charged in tunnel construction is determined by the following steps (i) to (iv).

(i) Preliminary surveys such as elastic wave exploration will be used to determine the tunnel geological profile and standard values for support patterns at each location.

(ii) Based on the above data, create standard drilling/charge arrangements for each support pattern.

(iii) An experienced engineer will observe the entire face on-site during construction and make a comprehensive judgment based on the visual condition of the drilling position, the feel of the percussion drill when drilling holes using a jumbo, and the construction record of existing faces (such as the amount of over-excavation visually confirmed), and determine the amount of powder to be charged.

(iv) When hard rock is present in some areas, the number of drilling charges may be increased arbitrarily compared to the standard pattern.

As mentioned above, the decision on the amount of powder to be charged is heavily dependent on the intuition of experienced technicians, and as the population continues to age and decline in the birthrate, it will be necessary to decide the amount of powder to be charged without involving the judgment of experienced technicians.

In addition, in recent years, computer jumbos have been developed that automatically acquire various data corresponding to the feel of percussion drills in jumbos (drilling speed, damping pressure, rotation pressure, and drilling energy calculated from three items, etc.) and the coordinates of the drilling position in numerical form, and various companies are working on methods for determining the amount of powder using this data (for example, non-patent document 1).

"Construction record of large cross-section tunnel using Japan's first 4-boom fully automatic computer jumbo" <URL: https://jcmanet. or. jp/bunken/symposium/2018/2018r18. pdf>

However, most conventional techniques use various data and separately measured overcutting of the face as explanatory variables to determine the amount of charge that minimizes overcutting.

It is difficult to determine rock type, cracks, interlayer inclusions, etc. using only the various data obtained from computer jumbos. Currently, these items are judged by experienced engineers through visual inspection and are used as a factor in selecting the amount of powder charge, but it is difficult to take this influence into account with the conventional method of determining the amount of powder charge using only the various data obtained from computer jumbos. Furthermore, in order to measure the amount of overexcavation at the face, it is necessary to approach the unstable face immediately after blasting and directly obtain data using a three-dimensional laser scanner, etc., but this is a dangerous task with a high possibility of the face collapsing.

Taking the above facts into consideration, the present invention aims to be able to accurately estimate the charge amount at the drilling position.

The analysis device according to the present invention includes an acquisition unit that acquires a support pattern determined in advance for the face position of the natural ground to be excavated for a tunnel, a panoramic face image photographed by a camera of the face position, and acquires, for each drilling position for charging the tunnel face, sensor data obtained when drilling the hole at the drilling position and a partial image of the panoramic face image that includes the hole drilling position; and an estimation unit that estimates the amount of charging at each drilling position using a pre-trained model that inputs the support pattern, the sensor data, and the partial image, and outputs the amount of charging at the drilling position.

According to the analysis device of the present invention, the acquisition unit acquires a support pattern determined in advance for the face position of the natural ground to be excavated for the tunnel, and a panoramic image of the face taken by a camera of the face position, and also acquires, for each drilling position for the tunnel face, sensor data obtained when drilling the hole at the drilling position and a partial image of the panoramic image of the face that includes the drilling position. Then, the estimation unit estimates the amount of charge at each drilling position using a pre-trained model that inputs the support pattern, the sensor data, and the partial image and outputs the amount of charge at the drilling position.

In this way, the amount of charge at each drilling position can be estimated with high accuracy by using a pre-trained model that inputs the support pattern, the sensor data, and the partial image for each drilling position and outputs the amount of charge at the drilling position.

The model according to the present invention is a neural network model that inputs the support pattern, the sensor data, and the partial image, and outputs the charge amount at the drilling position.

The neural network model of the present invention may include a convolutional layer.

The sensor data according to the present invention may include the drilling speed, drilling energy, or damping pressure when drilling a hole.

The model according to the present invention can further input the full-view image of the face and output the charge amount at the drilling position.

The learning device according to the present invention includes a learning data acquisition unit that acquires learning data including a support pattern determined in advance for the face position of the natural ground to be excavated for a tunnel, sensor data obtained when drilling holes at each drilling position for the tunnel face, a partial image including the drilling position in a full-view image of the face captured by a camera of the face position, and an optimal amount of charge for the drilling position determined from the blasting results at the face position, and a learning unit that learns a model that inputs the support pattern, the sensor data, and the partial image based on the learning data, and outputs the amount of charge for the drilling position.

According to the learning device of the present invention, the learning data acquisition unit acquires learning data including a support pattern determined in advance for the face position of the natural ground to be excavated for the tunnel, sensor data obtained when drilling holes at each drilling position for the tunnel face, a partial image including the drilling position in a full face image captured by a camera of the face position, and the optimal amount of charge for the drilling position determined from the blasting results at the face position. Then, based on the learning data, the learning unit learns a model that inputs the support pattern, the sensor data, and the partial image, and outputs the amount of charge for the drilling position.

In this way, by learning a model that inputs the support pattern, the sensor data, and the partial image and outputs the amount of charge at the drilling position, it is possible to accurately estimate the amount of charge at the drilling position.

The analysis program of the present invention is a program that causes a computer to function as an acquisition unit that acquires a support pattern that is predetermined for the face position of the natural ground to be excavated for a tunnel, a panoramic face image taken by a camera of the face position, and, for each drilling position for charging the tunnel face, acquires sensor data obtained when drilling the hole at the drilling position and a partial image of the panoramic face image that includes the hole drilling position, and an estimation unit that estimates the amount of charging at each drilling position using a pre-trained model that inputs the support pattern, the sensor data, and the partial image, and outputs the amount of charging at the drilling position.

The learning program of the present invention is a program for making a computer function as a learning data acquisition unit that acquires learning data including a shoring pattern determined in advance for the face position of the natural ground to be excavated for a tunnel, a full-view image of the face position photographed by a camera, sensor data obtained when drilling holes at each drilling position for the tunnel face, a partial image of the full-view image of the face that includes the drilling position, and the optimal amount of charge for the drilling position determined from the blasting results at the face position, and a learning unit that learns a model that inputs the shoring pattern, the sensor data, and the partial image, and outputs the amount of charge for the drilling position, based on the learning data.

As described above, the analysis device and analysis program of the present invention have the effect of estimating the charge amount at each drilling position with high accuracy by using a pre-trained model that inputs the support pattern, the sensor data, and the partial image for each drilling position and outputs the charge amount at the drilling position.

In addition, the learning device and learning program of the present invention can provide the effect of accurately estimating the amount of charge at the drilling position by learning a model that inputs the support pattern, the sensor data, and the partial image and outputs the amount of charge at the drilling position.

1 is a block diagram showing an analysis device according to an embodiment of the present invention; 1 is a functional block diagram showing an analysis device according to an embodiment of the present invention. FIG. 11 is a diagram illustrating an example of input data. FIG. 1 is a diagram illustrating an example of a configuration of a neural network. 4 is a flowchart showing the contents of an analysis processing routine of the analysis device according to the embodiment of the present invention. FIG. 13 is a diagram showing experimental results using the method of the present embodiment. FIG. 13 is a diagram showing experimental results using the method of the present embodiment. FIG. 13 is a diagram showing experimental results using the method of the present embodiment. FIG. 13 is a diagram showing experimental results using the method of the present embodiment. FIG. 13 is a diagram showing experimental results using the method of the present embodiment. FIG. 13 is a diagram showing the results of applying the method of the present embodiment to a field.

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

<Overview of the embodiment of the present invention>
In an embodiment of the present invention, drilling energy data obtained from the drill navigation of the computer drill jumbo and local images of the drilling locations are used to train a neural network and predict the amount of charge at each drilling location.

<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 (Central Processing Unit) 12, a graphics card 13, a GPU (Graphics Processing Unit) 14, a RAM (Random Access Memory) 16, a HDD (Hard Disk Drive) 18, a communication interface (I/F) 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 accepts as input a support pattern determined in advance for the face position of the ground to be excavated for the tunnel, and a panoramic image of the face position captured by a camera.

For each drilling position for loading the tunnel face with explosives, the input unit 10 receives as input the sensor data obtained when drilling the hole at that position and a partial image of the entire tunnel face image that includes that drilling position.

The sensor data includes the drilling speed, drilling energy, or damping pressure when the drill is used to drill a hole at the drilling position. For example, the sensor data includes the drilling speed, drilling energy, and damping pressure obtained from the computer Jumbo's drill navigation when the drill is used to drill a hole at the drilling position.

The calculation unit 20 includes an acquisition unit 22, an estimation unit 24, a model memory unit 26, a learning data acquisition unit 28, and a learning unit 30.

The acquisition unit 22 acquires the support pattern that has been determined in advance for the face position of the ground to be excavated for the tunnel, which is input to the input unit 10, and a panoramic image of the face position captured by a camera. It also acquires, for each drilling position for the tunnel face loading, sensor data obtained when drilling at that position, and a partial image of the panoramic image of the face that includes that position.

For example, as shown in Figure 3, the distance from the tunnel entrance, drilling length, average energy, and soil cover are acquired as sensor data, and the support pattern is acquired as either "CI," "CII," "DI," or "DII," and the storage location of the partial image is acquired. Note that Figure 3 shows an example in which the charge amount is also acquired as learning data.

The estimation unit 24 estimates the charge amount at each drilling position using a pre-trained model that inputs the support pattern, sensor data, and partial image for each drilling position and outputs the charge amount at the drilling position.

Specifically, the model is a neural network model that takes the support pattern, sensor data, and partial images as input, and outputs the charge amount at the drilling position. More specifically, the neural network model includes a convolutional layer.

For example, as shown in FIG. 4, the neural network model 32 includes an input layer 321 that receives as input 520 parameters representing the support pattern, the sensor data for the drilling position, and a partial image of the drilling position, an intermediate layer 322 that receives as input 520 parameters representing the output of the input layer 321, and an output layer 323 that outputs one parameter representing the charge amount for the drilling position.

The neural network model may also take a panoramic image of the face as input.

The model memory unit 26 stores the trained model.

As shown in FIG. 3 above, the learning data acquisition unit 28 acquires learning data including a support pattern determined in advance for the face position of the ground to be excavated for the tunnel, sensor data obtained when drilling holes at each drilling position for the tunnel face, a partial image including the drilling position in a full-view image of the face taken by a camera, and the optimal amount of powder for the drilling position determined from the blasting results at the face position.

The learning unit 30 learns a neural network model that inputs the support pattern, sensor data, and partial images based on the learning data acquired by the learning data acquisition unit 28, and outputs the amount of explosives at the drilling position, and stores the learned model in the model storage unit 26.

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

First, for each face position of the natural ground to be excavated for the tunnel, multiple drilling positions on the cross section of the face position and the amount of charge for each drilling position are determined manually for the first predetermined number of times (e.g., five times), and the input unit 10 of the analysis device 100 accepts as input the amount of charge for each manually determined drilling position for the first predetermined number of times.

In addition, the input unit 10 of the analysis device 100 accepts as input the support pattern, sensor data for each drilling position, and partial images of each drilling position within the overall face image for the first predetermined number of times.

The learning unit 30 of the analysis device 100 learns a model using the support pattern, the sensor data for each drilling position, the partial image for each drilling position, and the charge amount for each drilling position obtained in the first specified number of times as learning data, and stores the learned model in the model storage unit 26.

Then, after the first predetermined number of times, the input unit 10 of the analysis device 100 receives as input the support pattern, the sensor data for each drilling position, and a full-view image of the face for each face position of the natural ground to be excavated for the tunnel, and the analysis processing routine shown in Figure 5 is executed by the analysis device 100.

First, in step S100, the acquisition unit 22 acquires a pre-determined support pattern for the face position of the ground to be excavated for the tunnel, which is input to the input unit 10, and a panoramic image of the face position captured by a camera.

Then, in step S102, the acquisition unit 22 acquires, for each drilling position on the cross section of the face position that was manually determined for the tunnel face charging, sensor data obtained when drilling the hole at the drilling position and a partial image including the drilling position within the full face image.

In step S104, the estimation unit 24 inputs the support pattern, sensor data, and partial image for each drilling position, and estimates the charge amount for that drilling position using the trained model.

In step S106, the estimated charge amount for each drilling position in step S104 is output by the output unit 50.

The user determines the charge amount for each drilling position based on the estimated charge amount for each drilling position. Then, blasting is actually performed at the face position with the determined charge amount for each drilling position, and the user determines the optimal charge amount for each drilling position from the blasting results. The input unit 10 of the analysis device 100 then accepts the optimal charge amount for each drilling position as input.

In step S108, the learning data acquisition unit 28 acquires the average value of the optimal charge amount for each drilling position as the correct answer data.

In step S110, the learning data acquisition unit 28 acquires learning data including a support pattern determined in advance for the face position of the ground to be excavated for the tunnel, sensor data obtained when drilling holes at each drilling position for the tunnel face loading, a partial image including the drilling position in a full-view image of the face taken by a camera, and the average value of the optimal loading amount for each drilling position acquired as correct answer data.

In step S112, the learning unit 30 learns the trained model by updating the trained model so as to optimize a predetermined loss function based on the training data.

In step S114, it is determined whether or not the face work on the natural ground has been completed. If the face work on the natural ground has not been completed, the process returns to step S100 and the process of steps S100 to S114 is repeated for the next face position.

<Example>
Using construction data for 640 m of a tunnel, the effectiveness of the method described in the above embodiment was verified by comparing with two comparative methods.

Comparative Example 1 is a method that uses multiple regression using only numerical data. The method of Comparative Example 1 is a method that has low expressiveness but is less prone to overlearning when analyzing numerical data.

Comparative Example 2 is a method that uses LightGBM regression using only numerical data. This method is highly expressive when used to analyze numerical data, but is prone to overlearning. The method in Comparative Example 2 is, for example, a decision tree-based method that uses random forest and gradient boosting. After trying three or four patterns of hyperparameters, we selected one that seemed to be problem-free.

The method described in the above embodiment is a method that uses MLP regression using numerical data and image data. Note that because random numbers are used for learning, the results will not be the same every time.

Figure 6 shows the experimental results in which data from six face positions among the first 95th to 110th face positions was used as learning data, and data from 65 face positions among the 118th face positions onwards was used as validation data. Figure 6(A) shows the experimental results of Comparative Example 1, Figure 6(B) shows the experimental results of Comparative Example 2, and Figure 6(C) shows the experimental results of the method of the above embodiment. In Figure 6, the open circles indicate the average value of the estimated charge amount at each drilling position by the model, and the dotted circles indicate the average charge amount at each drilling position obtained as the correct answer data. The vertical axis indicates the charge amount, and the horizontal axis indicates the distance from the wellhead.

At the stage of five pieces of training data, it can be seen that prediction is difficult for any method. However, while Comparative Example 1 is extremely negative and Comparative Example 2 only produces almost the same prediction, it can be said that the method of the above embodiment produces a relatively realistic prediction.

Figure 7 shows the experimental results in which data from 15 face positions from the first 95th to 168th face positions was used as learning data, and data from 56 face positions from the 172nd face positions onwards was used as validation data. Figure 7(A) shows the experimental results of Comparative Example 1, Figure 7(B) shows the experimental results of Comparative Example 2, and Figure 7(C) shows the experimental results of the method of the above embodiment.

In Comparative Example 1, the point cloud clusters were captured to a certain extent, but there were still many problems, such as predictions outside the data range. However, it can be seen that the method of the above embodiment is able to keep the deviation in the predictions immediately afterwards small.

Figure 8 shows the experimental results in which data from 40 face positions from the 95th to 313th cut face positions from the beginning was used as learning data, and data from 31 face positions from the 317th cut face positions onwards was used as validation data. Figure 8(A) shows the experimental results of Comparative Example 1, Figure 8(B) shows the experimental results of Comparative Example 2, and Figure 8(C) shows the experimental results of the method of the above embodiment.

It can be seen that prediction is difficult with the method in Comparative Example 1 due to issues with expressiveness. In Comparative Example 2, prediction is somewhat more successful due to the increased amount of data. However, due to the nature of decision tree methods, there is still a tendency for the same value to appear for unknown data.

It was found that the method of the above embodiment was able to obtain closer predictions than the two comparative examples, both for immediate predictions and for subsequent predictions.

Figure 9 shows the experimental results in which data from 60 face positions from the 95th to 586th cut face positions from the beginning was used as learning data, and data from 11 face positions from the 590th cut face position onwards was used as validation data. Figure 9(A) shows the experimental results of Comparative Example 1, Figure 9(B) shows the experimental results of Comparative Example 2, and Figure 9(C) shows the experimental results of the method of the above embodiment.

All of the methods produce similar values for the next prediction after learning is complete. However, while comparative examples 1 and 2 produce values close to the average, which seems to be correct by chance, the predictions made by the method of the above embodiment vary and appear to fluctuate according to the input parameters.

Next, we will explain the results of the correlation evaluation. Figure 10 shows the results of evaluating the correlation coefficient between the data used for prediction and the correct data for each distance from the mine entrance for each method. The vertical axis is the correlation coefficient, and the horizontal axis is the distance from the mine entrance. If a positive correlation is obtained, it can be said that the prediction is working.

While the regressions in Comparative Examples 1 and 2 do not show a clear positive correlation, the method of the above embodiment shows a positive correlation over a long period.

The correlation is disrupted in the first half where there is insufficient training data, and in the second half where there is less evaluation data, but the results show the effectiveness of only the method of the above embodiment.

As described above, by comparing the method of the above embodiment with the methods of the two comparative examples, it was found that regression using images has a certain degree of effectiveness.

Next, we will explain the results of applying the method at a tunnel construction site.

First, learning data was created based on construction data up to July 28, 2022. In addition, as of July 28, construction was completed up to 948 m from the tunnel entrance. The number of valid sections of image data was 145 sections, and the number of valid data was 3,998. In addition, the model was trained until the correlation coefficient reached 0.99.

As prediction data, the model was used to predict the amount of charge in the support pattern DIII section from August onwards, and as of September 1st, predictions were made for a total of two sections. The results are shown in Figure 11.

In Figure 11, the white circles indicate the average charge amount for each drilling position obtained as actual results, and the dotted circles indicate the average charge amount for each drilling position estimated by the model. The vertical axis is the charge amount (number of charges per hole), and the horizontal axis is the distance from the wellhead.

In the first comparison of prediction and actual results, the prediction was 2.34 pieces/hole and the actual result was 2.64 pieces/hole. The amount of explosives was corrected to an error of 0.30 pieces/hole and blasting was performed, but there were no problems with excavation such as hits.

In addition, in a comparison of the second prediction and the actual results, the prediction was 2.63 pieces/hole and the actual result was 2.64 pieces/hole. The amount of explosives was corrected with an error of 0.01 pieces/hole and blasting was performed, but there were no problems with excavation such as hitting the ground. It was found that blasting and excavation were possible with almost the same amount of explosives as the prediction.

Based on the above, the amount of charge in the support pattern DIII section was predicted and corrected based on the actual amount of charge in the pre-excavation, and blasting was performed. As a result, construction could be carried out without any excavation problems such as hitting the rocks.

Based on this, it is believed that a system has been developed that predicts the amount of charge using drilling energy data from DrillNavi and local face images of the drilling location.

As described above, the analysis device according to the embodiment of the present invention can accurately estimate the amount of charge at each drilling position by estimating the amount of charge at the drilling position using a pre-trained model that inputs the support pattern, sensor data, and partial image for each drilling position and outputs the amount of charge at the drilling position.

In addition, by learning a model that inputs the support pattern, sensor data, and partial images and outputs the amount of charge at the drilling position, it is possible to learn a model that accurately estimates the amount of charge at the drilling position.

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 embodiment, the learning process and the analysis process are performed by a single device, but the present invention is not limited to this. For example, the learning process may be performed by a separate learning device, and the analysis device may be configured to perform the analysis process. In this case, the calculation unit of the learning device includes a model storage unit, a learning data acquisition unit, and a learning unit, and the calculation unit of the analysis device includes an acquisition unit, an estimation unit, and a model storage unit.

In addition, while an example has been described in which estimation of the charge amount and model learning are repeated alternately in tunnel construction work at the same site, the present invention is not limited to this. The charge amount in tunnel construction work at the target site may be estimated using a trained model trained using data from tunnel construction work at another site that was previously carried out.

In the above embodiment, the loss function is designed using the average of the charge amounts at each drilling position at the face position as the correct answer data. However, a loss function that incorporates the process and ideas used when determining the charge amount on-site may also be used. For example, the loss function may be designed using the charge amounts at each drilling position at the face position as the correct answer data.

In addition, although the sensor data used herein is the drilling speed, drilling energy, or damping pressure when drilling a hole at a drilling position, the sensor data is not limited to this. Other sensor data that contributes to determining the charge amount may also be used.

In addition, although an example has been described in which the drilling positions for charging are determined manually, this is not limiting. The computer may automatically determine the drilling positions. For example, the drilling positions for charging may be automatically determined based on the drilling energy of a computer drill jumbo and face image information, and the amount of charging may be estimated using the above-mentioned analysis device.

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

10 Input unit 20 Calculation unit 22 Acquisition unit 24 Estimation unit 26 Model storage unit 28 Learning data acquisition unit 30 Learning unit 32 Neural network model 50 Output unit 100 Analysis device 321 Input layer 322 Intermediate layer 323 Output layer

Claims (8)

A support pattern determined in advance for the face position of the ground to be excavated in the tunnel and a panoramic image of the face position photographed by a camera are acquired,
An acquisition unit that acquires, for each drilling position for a tunnel face, sensor data obtained when drilling the hole at the drilling position and a partial image including the drilling position in the face panoramic image;
An estimation unit that estimates the amount of charge at the drilling position using a pre-trained model that inputs the support pattern, the sensor data, and the partial image for each drilling position and outputs the amount of charge at the drilling position;
An analysis device comprising: The analysis device according to claim 1, wherein the model is a neural network model that receives the support pattern, the sensor data, and the partial image as inputs, and outputs the charge amount at the drilling position. The analysis device according to claim 2, wherein the neural network model includes a convolutional layer. The analysis device according to claim 1, wherein the sensor data includes the drilling speed, drilling energy, or damping pressure when drilling a hole. The model further inputs the full-view image of the face and outputs the charge amount at the drilling position. The analysis device according to any one of claims 1 to 4. a learning data acquisition unit that acquires learning data including a support pattern determined in advance for the face position of the natural ground to be excavated in the tunnel, sensor data obtained when drilling holes at each drilling position for the tunnel face for the purpose of charging the tunnel face, a partial image including the drilling position in a full view image of the face taken by a camera, and an optimal amount of charging for the drilling position determined from the blasting results at the face position;
A learning unit that learns a model that inputs the support pattern, the sensor data, and the partial image based on the learning data and outputs the amount of charge at the drilling position;
A learning device comprising: Computer,
A support pattern determined in advance for the face position of the ground to be excavated in the tunnel and a panoramic image of the face position photographed by a camera are acquired,
an acquisition unit that acquires, for each drilling position for loading explosives for the tunnel face, sensor data obtained when drilling at the drilling position and a partial image including the drilling position within the panoramic image of the face; and an analysis program to function as an estimation unit that estimates the amount of explosives at the drilling position using a pre-trained model that inputs the support pattern, the sensor data, and the partial image, and outputs the amount of explosives at the drilling position. Computer,
a learning data acquisition unit that acquires learning data including a shoring pattern determined in advance for the face position of the natural ground to be excavated through a tunnel, a panoramic image of the face position photographed by a camera, and for each drilling position for propellant loading for the tunnel face, sensor data obtained when drilling at the drilling position, a partial image of the panoramic image of the face that includes the drilling position, and an optimal propellant loading amount for the drilling position determined from the blasting results at the face position; and a learning unit that learns a model that uses the shoring pattern, the sensor data, and the partial image as inputs and outputs the propellant loading amount for the drilling position based on the learning data.