Tunnel face evaluation support device, tunnel face evaluation support program, and tunnel face evaluation support method

Abstract

To evaluate a tunnel face with high accuracy.SOLUTION: A tunnel face evaluation support device comprises an image acquisition unit 2 that acquires tunnel face images, a first face evaluation category inference unit 6 that infers a provisional evaluation category from the tunnel face image using a first trained face evaluation model 11 that is trained to input training images related to the tunnel face and infer an evaluation category corresponding to the training image, a drilling data acquisition unit 4 that acquires drilling data obtained by a rock drill when the tunnel face image is obtained, and a second face evaluation category inference unit 7 that inputs training drilling data obtained when the training image is acquired and the provisional evaluation category inferred by the first face evaluation category inference unit 6 for the training image as training input data and infers an evaluation category, taking into account both the tunnel face image and the drilling data, from the drilling data and the provisional evaluation category inferred by the first face evaluation category inference unit 6 for the tunnel face image, using a second trained face evaluation model 12 that is trained to infer an evaluation category corresponding to the training input data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a tunnel face evaluation support device, a tunnel face evaluation support program, and a tunnel face evaluation support method. [Background technology]

[0002] In mountain tunnel construction, the following process is repeated: drilling a hole with a rock drill, charging the hole with a charge, detonating the charge, removing the resulting debris from the tunnel, and then spraying concrete onto the wall. In order to confirm and improve the workability and safety of this series of operations, the condition of the tunnel face that appears after the debris is removed is sometimes evaluated.

[0003] Such tunnel face evaluations are generally performed by workers visually observing the tunnel face and evaluating multiple items such as compressive strength, which indicates the strength of the rock mass, and weathering alteration, which indicates the degree of alteration of the rock mass due to weathering, etc. However, such evaluations are qualitative and may depend on the worker's senses and experience, so some workers may not be able to perform the evaluation accurately.

[0004] In this regard, Patent Document 1 discloses a tunnel face evaluation support device that evaluates a tunnel face based on image data of a target area including the tunnel face and point cloud data of measured three-dimensional shapes. In such a configuration, the tunnel face is evaluated by the face evaluation support device based on image data and point cloud data, which may allow for more accurate evaluation than when workers evaluate the tunnel face. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2023-34740 Summary of the Invention [Problem to be solved by the invention]

[0006] Here, compressive strength, which indicates the strength of the rock mass and is exemplified above as an item for evaluating a tunnel face, is evaluated, for example, by the ease with which the rock mass cracks when struck with a hammer. In this regard, in the configuration of Patent Document 1, both the image data and point cloud data used for evaluating the tunnel face are data related to the external shape of the tunnel face. In such a configuration, it is not easy to properly evaluate evaluation items such as the compressive strength, which are difficult to evaluate based on the external shape of the tunnel face alone. For this reason, there are limits to how accurately a tunnel face can be evaluated.

[0007] The problem to be solved by the present invention is to provide a tunnel face evaluation support device, a tunnel face evaluation support program, and a tunnel face evaluation support method that evaluate a tunnel face with high accuracy. [Means for solving the problem]

[0008] In order to solve the above problems, the present invention employs the following means. In other words, the present invention provides a tunnel face evaluation support device that supports the evaluation of a tunnel face, and includes: an image acquisition unit that acquires a tunnel face image; a first face evaluation category inference unit that inputs a learning image related to the tunnel face and infers a provisional evaluation category from the tunnel face image using a first trained face evaluation model that has been trained to infer an evaluation category corresponding to the learning image; a drilling data acquisition unit that acquires drilling data obtained by a rock drill when the tunnel face image was obtained; and a second face evaluation category inference unit that inputs the training drilling data when the training image was obtained and the provisional evaluation category inferred by the first face evaluation category inference unit for the training image as learning input data and infers the evaluation category taking into account both the tunnel face image and the drilling data from the drilling data and the provisional evaluation category inferred by the first face evaluation category inference unit for the tunnel face image using a second trained face evaluation model that has been trained to infer the evaluation category corresponding to the learning input data.

[0009] According to the above configuration, the first trained face evaluation model is trained to input a training image related to a tunnel face and infer an evaluation category corresponding to the training image. That is, the first trained face evaluation model is trained to evaluate a tunnel face based on its external shape. When a tunnel face image is input to such a first trained face evaluation model, a provisional evaluation category is inferred as an evaluation result based on the external shape of the tunnel face. The second trained tunnel face evaluation model is trained to input the provisional evaluation category inferred by the first tunnel face evaluation category inference unit for the training image and the training drilling data when the training image was obtained as training input data, and to infer the evaluation category corresponding to the training input data. In other words, the second trained tunnel face evaluation model is trained to evaluate the tunnel face taking into account both the external shape and properties of the tunnel face by using both the provisional evaluation category, which is the evaluation result based on the external shape of the tunnel face, and the drilling data obtained by a rock drill. When the drilling data and the provisional evaluation category inferred by the first tunnel face evaluation category inference unit for the corresponding tunnel face image are input to the second trained tunnel face evaluation model, an evaluation category is inferred that takes into account both the tunnel face image, i.e., the external shape, and the drilling data, i.e., the properties. In this way, the final assessment category takes into account both the external shape and properties of the tunnel face, which makes it possible to improve the accuracy of assessments for items that are difficult to assess sufficiently from the external shape of the tunnel face alone. In this way, it is possible to provide a tunnel face evaluation support device that evaluates the tunnel face with high accuracy.

[0010] In one aspect of the present invention, when an image of the entire tunnel face is input, the image acquisition unit generates the tunnel face image by cutting out each of the top, right shoulder, and left shoulder of the tunnel.

[0011] According to the above configuration, the tunnel face image to be input to the first trained tunnel face evaluation model is generated by cutting out the respective portions of the tunnel crown, right shoulder, and left shoulder from an image of the entire tunnel face. In this way, the tunnel face can be appropriately evaluated.

[0012] In another aspect of the present invention, the drilling data is any one or any combination of hole depth, drilling speed, impact pressure, feed pressure, damper pressure, rotation speed, rotation pressure, water volume, and water pressure.

[0013] According to the above configuration, the tunnel face can be evaluated in detail from the perspective of its properties, which is different from its external shape, using drilling data input as any one or a combination of the hole depth, drilling speed, impact pressure, feed pressure, damper pressure, rotational speed, rotational pressure, water volume, and water pressure.

[0014] In another aspect of the present invention, the tunnel face evaluation support device further includes a pre-processing unit that corrects the drilling data, and the second face evaluation category inference unit inputs the drilling data corrected by the pre-processing unit into the second learned face evaluation model, and the pre-processing unit performs one or a combination of the following: missing value processing that excludes error data in the drilling data; item selection processing that selects some of the items from multiple items included in the drilling data and deletes the other items; dimensionality reduction processing that reduces the number of items by extracting features from the multiple items through statistical analysis and creating new items; and averaging processing that averages multiple data obtained according to the hole depth.

[0015] According to the above configuration, when the preprocessing unit executes missing value processing to remove erroneous data from the punching data, the erroneous data is prevented from being reflected in the evaluation. In addition, when the pre-processing unit performs an item selection process in which some items are selected from multiple items contained in the drilling data and other items are deleted, if there is an item in the drilling data that has a negative impact on the evaluation, it can be deleted and excluded from the input to the second trained face evaluation model. Furthermore, when the preprocessing unit performs dimension reduction processing, it extracts features from multiple items of drilling data using statistical analysis such as principal component analysis, and creates new items. The new items generated in this way more clearly reflect the characteristics of the drilling data as a whole than the items before the dimension reduction processing was performed. Therefore, the evaluation accuracy of the second trained tunnel face evaluation model may be improved. At the same time, the dimension reduction processing reduces the total number of items. Therefore, the configuration of the entire tunnel face evaluation support device, including the second trained tunnel face evaluation model, may be simplified. When the preprocessing unit performs an averaging process that averages multiple pieces of data obtained according to the depth of the hole, the data obtained for each depth for one hole is aggregated into a single value, thereby reducing the number of drilling data corresponding to one hole. This can simplify the configuration of the tunnel face evaluation support device, including the second trained face evaluation model.

[0016] The present invention also provides a tunnel face evaluation support program for supporting the evaluation of a tunnel face, the program comprising: an image acquisition means for acquiring a tunnel face image; a first face evaluation category inference means for inputting a learning image relating to a tunnel face and inferring a provisional evaluation category from the tunnel face image using a first trained face evaluation model that has been trained to infer an evaluation category corresponding to the learning image; a drilling data acquisition means for acquiring drilling data by a rock drill when the tunnel face image is obtained; and a first face evaluation category inference means for inputting a learning image relating to the tunnel face and inferring a provisional evaluation category from the tunnel face image using a first trained face evaluation model that has been trained to infer an evaluation category corresponding to the learning image. A tunnel face evaluation support program is provided for realizing a second face evaluation category inference means that inputs the provisional evaluation category inferred by the first face evaluation category inference means for the learning image as learning input data, and infers the evaluation category taking into account both the tunnel face image and the drilling data from the drilling data and the provisional evaluation category inferred by the first face evaluation category inference means for the tunnel face image using a second trained face evaluation model that has been trained to infer the evaluation category corresponding to the learning input data.

[0017] According to the above-described configuration, as already explained with respect to the tunnel face evaluation support device, the tunnel face can be evaluated with high accuracy.

[0018] The present invention also provides a tunnel face evaluation support method for supporting the evaluation of a tunnel face, comprising: an image acquisition step for acquiring a tunnel face image; a first face evaluation category inference step for inputting a learning image related to the tunnel face and inferring a provisional evaluation category from the tunnel face image using a first trained face evaluation model trained to infer an evaluation category corresponding to the learning image; a drilling data acquisition step for acquiring drilling data obtained by a rock drill when the tunnel face image was obtained; and a second face evaluation category inference step for inputting the training drilling data when the training image was obtained and the provisional evaluation category inferred for the training image by the first face evaluation category inference step as learning input data and inferring the evaluation category taking into account both the tunnel face image and the drilling data from the drilling data and the provisional evaluation category inferred for the tunnel face image by the first face evaluation category inference step using a second trained face evaluation model trained to infer the evaluation category corresponding to the learning input data.

[0019] According to the above-described configuration, as already explained with respect to the tunnel face evaluation support device, the tunnel face can be evaluated with high accuracy. [Effects of the Invention]

[0020] According to the present invention, it is possible to provide a tunnel face evaluation support device, a tunnel face evaluation support program, and a tunnel face evaluation support method that evaluate a tunnel face with high accuracy. [Brief explanation of the drawings]

[0021] [Figure 1] 1 is a block diagram of a tunnel face evaluation support device according to an embodiment of the present invention. [Figure 2]FIG. 1 shows an image of the entire tunnel face. [Figure 3] FIG. 10 is an explanatory diagram of the process of extracting learning images and tunnel face images from an image of the entire tunnel face. [Figure 4] FIG. 2 is an explanatory diagram showing the flow of data when learning the first face evaluation model and the second face evaluation model. [Figure 5] FIG. 1 is an explanatory diagram showing the process of drilling a hole using a rock drill. [Figure 6] FIG. 1 is an explanatory diagram showing an image of the operation of the rock drill's drifter and drilling rod. [Figure 7] FIG. 10 is an explanatory diagram of punching data. [Figure 8] This is an explanatory diagram showing the flow of data when inferring an evaluation category using the first trained face evaluation model and the second trained face evaluation model. [Figure 9] 10 is a flowchart showing the processing flow when learning the first face evaluation model and the second face evaluation model. [Figure 10] This is a flowchart showing the processing flow when inferring an evaluation category using the first trained face evaluation model and the second trained face evaluation model. [Figure 11] FIG. 10 is an explanatory diagram showing the evaluation results obtained by the tunnel face evaluation support device of the above embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0022] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The tunnel face evaluation support device supports the evaluation of tunnel faces. The tunnel face evaluation support device infers the evaluation category for various evaluation items related to the tunnel face based on drilling data obtained when a hole is drilled with a rock drill and images of the tunnel face taken after the hole created by drilling is charged and then blasted to remove the resulting debris outside the tunnel.

[0023] Evaluation items related to the tunnel face include, for example, compressive strength, weathering, crack spacing, crack condition, running inclination, amount of spring water, and deterioration due to spring water. Compressive strength is an evaluation item related to the strength of the rock mass. Weathering is an evaluation item that indicates the extent to which the rock mass has been altered by weathering and water. Crack spacing is an evaluation item related to the spacing of cracks that have appeared at the tunnel face. Crack condition is an evaluation item for evaluating whether the cracks are tightly packed or open. Running inclination is an evaluation item related to the direction in which the boundaries between multiple rock masses appear at the tunnel face extend and the inclination angle of the boundaries. The amount of spring water is an evaluation item related to the amount of water seeping out of the rock mass. Deterioration due to spring water is an evaluation item that indicates the extent to which the rock mass has been altered by the seeping water.

[0024] For each of these evaluation items, multiple evaluation categories are set. For example, with regard to compressive strength, six evaluation categories, e.g., 1 to 6, can be set depending on the strength of the rock mass. In this case, for example, if a rock fragment placed on the ground is difficult to break even when struck with a hammer, evaluation category 1 indicates the strongest rock mass. As the value increases, the rock mass becomes weaker, e.g., if the rock fragment can be crushed with a fingertip, evaluation category 6 indicates the weakest rock mass. Similarly, with regard to weathering alteration, four evaluation categories, e.g., 1 to 4, can be set depending on the state of alteration of the rock mass. In this way, the number of evaluation categories can vary depending on the evaluation item. In general, in the evaluation of tunnel faces, the evaluation category corresponding to each evaluation item is determined by striking rock fragments with a hammer or visually inspecting the tunnel face.

[0025] FIG. 1 is a block diagram of a tunnel face evaluation support device according to an embodiment of the present invention. The tunnel face evaluation support device 1 includes an image acquisition unit 2, a learning unit 3, a drilling data acquisition unit 4, a pre-processing unit 5, a first face evaluation category inference unit 6, a second face evaluation category inference unit 7, and an evaluation category aggregation unit 8.

[0026] The tunnel face evaluation support device 1 is a computer equipped with hardware such as a processor and memory, and by executing the tunnel face evaluation support program according to this embodiment, it is functionally equipped with the above-mentioned configuration. That is, by executing the tunnel face evaluation support program, the tunnel face evaluation support device 1 causes the computer to realize an image acquisition function, a learning function, a drilling data acquisition function, a preprocessing function, a first face evaluation category inference function, a second face evaluation category inference function, and an evaluation category aggregation function, which correspond to the image acquisition unit 2, the learning unit 3, the drilling data acquisition unit 4, the preprocessing unit 5, the first face evaluation category inference unit 6, the second face evaluation category inference unit 7, and the evaluation category aggregation unit 8, respectively.

[0027] As will be explained later, the first tunnel face evaluation category inference unit 6 infers a provisional evaluation category when a tunnel face image is input. Furthermore, the second tunnel face evaluation category inference unit 7 infers an evaluation category that takes into account both the tunnel face image and the drilling data when the provisional evaluation category and drilling data are input. To effectively perform these inferences, the first tunnel face evaluation category inference unit 6 and the second tunnel face evaluation category inference unit 7 are equipped with a first trained tunnel face evaluation model 11 and a second trained tunnel face evaluation model 12, respectively, trained by the training unit 3. Therefore, the tunnel face evaluation support device 1 performs two main operations: training the evaluation category and estimating the evaluation category. For simplicity's sake, the following will first describe each component of the tunnel face evaluation support device 1 during training of the evaluation category, and then describe the behavior of each component during estimation of the evaluation category.

[0028] The learning unit 3 uses learning data 20 to train the first face evaluation category inference unit 6 and the second face evaluation category inference unit 7. The learning data 20 has multiple combinations of learning images, learning drilling data, and correct values ​​of evaluation categories as teacher data. The learning data 20 is prepared in advance, for example, by being acquired during past mountain tunnel construction work, before the first face evaluation category inference unit 6 and the second face evaluation category inference unit 7 learn. The learning images are images of the tunnel face. The learning drilling data is drilling data acquired by a rock drill when the tunnel face photographed as the learning image is obtained (strictly speaking, the drilling data has been processed by the pre-processing unit 5). In other words, after a hole is drilled and drilled by a rock drill to obtain the learning drilling data, an image of the tunnel face photographed in a state where the drilled hole has been charged, blasted, and debris removed is prepared as the learning image corresponding to the learning drilling data. Furthermore, the evaluation results of the tunnel face, for example by a worker, are prepared as correct values ​​for the evaluation category in correspondence with the learning image and the learning drilling data.

[0029] Fig. 2 is a diagram showing an image of the entire tunnel face. Fig. 3 is an explanatory diagram of the process of extracting learning images and tunnel face images from an image of the entire tunnel face. In this embodiment, learning images 101L are generated by extracting the top T, right shoulder R, and left shoulder L from each of images 100 of the entire tunnel face acquired in past mountain tunnel construction work.

[0030] FIG. 4 is an explanatory diagram showing the flow of data when the first face evaluation model and the second face evaluation model are trained. The learning unit 3 learns the first face evaluation model 11L to generate the first trained face evaluation model 11, and then learns the second face evaluation model 12L to generate the second trained face evaluation model 12. In this embodiment, the first trained face evaluation model 11 (first face evaluation model 11L) is, for example, a CNN (convolutional neural network), and the second trained face evaluation model 12 (second face evaluation model 12L) is a DNN (deep neural network). As described above, in this embodiment, the first trained face evaluation model 11 and the second trained face evaluation model 12 are realized by artificial neural networks, but they may also be realized by other methods such as SVM (support vector machine), RF (random forest), and logistic regression.

[0031] The first face evaluation model 11L and the second face evaluation model 12L are prepared for each evaluation item and trained for each evaluation item. As a result, the first trained face evaluation model 11 and the second trained face evaluation model 12 are generated individually for each evaluation item. In other words, in this embodiment, the number of first trained face evaluation models 11 and second trained face evaluation models 12 matches the number of evaluation items. Figure 4 and Figure 8, which will be used in the following explanation, are limited to illustrating one evaluation item, and therefore Figures 4 and 8 show only one first trained face evaluation model 11 and one second trained face evaluation model 12.

[0032] When the first face evaluation model 11L is machine-learned, the training image 101L is used as input to the first face evaluation model 11L. The first face evaluation model 11L includes a convolution processing unit 11a and a full connection unit 11b. The convolution processing unit 11a includes a plurality of convolution layers arranged in series. When the training image 101L is input, the convolution processing unit 11a performs convolution filtering using filters provided in each of the convolution layers, and further performs batch normalization processing and pooling processing as necessary to generate a feature map of the training image 101L.

[0033] The fully connected unit 11b includes an input layer 11bi and an output layer 11bo. Although not shown in FIG. 4, the fully connected unit 11b may include an intermediate layer between the input layer 11bi and the output layer 11bo. In the fully connected unit 11b, multiple layers are arranged in series, and the nodes constituting each layer are fully connected to each other between adjacent layers. The feature map generated in the final layer of the convolution processing unit 11a is input to the input layer 11bi. Then, calculations are sequentially performed based on weights set on the connection lines between nodes, tracing each layer constituting the fully connected unit 11b in order, and the final calculation results are stored in each output node 11n of the output layer 11bo. The output nodes 11n are provided according to each evaluation category. If the first face evaluation model 11L corresponds to an evaluation item having, for example, five evaluation categories, five output nodes 11n are provided, and the inference result 11LS during learning stored in the output node 11n has five numerical values ​​corresponding to these five output nodes 11n, as shown in Figure 4.

[0034] The inference result 11LS during training is expressed as a softmax function, which of the evaluation categories, for example, 1 to 5, the first face evaluation model 11L has inferred for the evaluation of the input training image 101L. For example, in the example of FIG. 4, it is inferred that the probability that the evaluation of the input training image 101L is the first evaluation category is 0.02, and the probability that it is the second evaluation category is 0.51. Similarly, the probabilities that it is the third, fourth, and fifth evaluation categories are 0.28, 0.16, and 0.03, respectively, and the sum of all these is 1. As a result, in this case, it is inferred that the evaluation of the input training image 101L corresponds to the second evaluation category, which is the evaluation category with the largest value among them.

[0035] This inference result 11LS during learning is compared with training data 120. The training data 120 is the evaluation result for the evaluation item corresponding to the first tunnel face evaluation model 11L when, for example, an experienced worker evaluates the tunnel face at the time when the training image 101L input to the first tunnel face evaluation model 11L was obtained, and is the correct value of the evaluation category for that evaluation item prepared in the training data 20 corresponding to the training image 101L. In the example of FIG. 4, the first tunnel face evaluation model 11L corresponds to an evaluation item having five evaluation categories and performs five-class classification, so the training data 120 also consists of five values. Among these, the one corresponding to the evaluation category evaluated by the worker, in this case the second evaluation category, is set to a value of 1, and the others are set to a value of 0.

[0036] As described above, machine learning of the first face evaluation model 11L is performed by adjusting the values ​​of the filter parameters, weights, etc. of the first face evaluation model 11L by the backpropagation method or the gradient descent method so that the inference result 11LS during learning is compared with the teacher data 120 and the inference result 11LS during learning becomes closer to the teacher data 120. As a result, when a learning image 101L is input, the first face evaluation model 11L is trained to output an inference result 11LS that is close to the corresponding teacher data 120.

[0037] In this way, the learning unit 3 generates the first trained tunnel face evaluation model 11 as a trained model used as a program module, which is part of artificial intelligence software, with learning parameters such as filter parameters and weights. The inference result 11S output by the first trained tunnel face evaluation model 11 after training is converted into a dummy variable that takes a variable of either 0 or 1. That is, the value corresponding to the evaluation category inferred to have the highest probability is converted to 1, and other values ​​are converted to 0. In the case of Figure 4, the resulting dummy variable is [0, 1, 0, 0, 0]. In this way, the inference result 11S is treated as the result of a provisional evaluation of the tunnel face performed by the first trained tunnel face evaluation model 11 based only on the image, i.e., the external shape of the tunnel face, i.e., as a provisional evaluation category 11T. For example, in the case of Figure 4, the first trained tunnel face evaluation model 11 infers that the provisional evaluation category 11T is 2.

[0038] In the manner described above, the first trained face evaluation model 11 is trained and generated in accordance with each evaluation item so as to input a training image 101L relating to a tunnel face and infer a provisional evaluation category 11T corresponding to the training image 101L.

[0039] When the learning of the first trained face evaluation model 11 is completed, the learning unit 3 trains the second face evaluation model 12L to generate the second trained face evaluation model 12. Before explaining this learning, the drilling data that is input to the second face evaluation model 12L and the second trained face evaluation model 12 will be explained below.

[0040] Fig. 5 is an explanatory diagram showing the process of drilling (boring) using a rock drill. Fig. 6 is an explanatory diagram showing the operation of the rock drill's drifter and drilling rod. In mountain tunnel construction, a hole H for loading explosives used for blasting is drilled by a rock drill 200 such as a computer jumbo.

[0041] The rock drill 200 is equipped with multiple booms 203. Each boom 203 is equipped with a guide shell 204, a drifter 205, and a drilling rod 206. The boom 203 is provided so as to be rotatable in any direction and to be extendable and retractable. The guide shell 204 is attached to the tip of the boom 203. The drifter 205 is a hydraulic impact device that is arranged at the rear of the guide shell 204 and is configured to move forward and backward while being guided by the guide shell 204. The drilling rod 206 is a rod-shaped member for drilling a hole in the rock mass G and is attached to the drifter 205 so as to move on the guide shell 204. A bit 207 for excavating the rock mass G is attached to the tip of the drilling rod 206.

[0042] The drifter 205 contains a hydraulically driven piston 205a. When the piston 205a strikes the rear end of the drill rod 206, an elastic stress wave is generated. This stress wave propagates through the drill rod 206 and reaches the bit 207, crushing the rock mass G. A feed pressure P is applied to the drifter 205 so that the front of the bit 207 abuts against the rock mass G and so that the energy of the stress wave is efficiently transmitted to the rock mass G. When the feed pressure P is applied in this manner, the force from the rock mass G acts on the drill rod 206 as a reaction force, but this reaction force is absorbed by the damper 208. Furthermore, rotation is applied to the drill rod 206 via the spline portion 206a to maintain high crushing efficiency. During drilling, water W is pumped into the drill rod 206 and the bit 207 to discharge cuttings from the hole H.

[0043] The drilling data is any one or a combination of the hole depth, drilling speed, impact pressure, feed pressure, damper pressure, rotational speed, rotational pressure, water volume, and water pressure. FIG. 7 is an explanatory diagram of the drilling data. In this embodiment, the drilling data includes the drilling speed during drilling, the impact pressure N of the drifter 205, the feed pressure P, the damper pressure (pressure acting on the damper 208), the rotational speed, rotational pressure, water volume, and water pressure of the drilling rod 206. The drilling data is acquired for each hole H. The drilling data can be acquired when the bit 207 reaches a certain depth from the surface of the rock mass G. In the example of FIG. 7, the drilling data is acquired at each position (depth) at approximately 10 cm intervals from the surface of the rock mass G.

[0044] In this embodiment, as will be described later, multiple data obtained for each hole according to depth are averaged by depth, so the hole depth is not used to infer the evaluation category. From this perspective, in this embodiment, the drilling data consists of the drilling speed, impact pressure, feed pressure, damper pressure, rotation speed, rotation pressure, water volume, and water pressure.

[0045] The drilling data expresses what kind of action the rock drill 200 had to exert on the rock mass when drilling or boring a hole with the rock drill 200, and the degree of that action is thought to vary depending on the properties of the rock mass. From this perspective, the drilling data can be said to indicate the properties of the tunnel face.

[0046] The drilling data as described above is corrected by the pre-processing unit 5 to generate the training drilling data 111L. Immediately after the bit 207 is placed against the surface of the rock mass G and drilling is started, drilling is performed at low pressure and low speed to prevent the bit 207 from vibrating and shifting from the planned drilling position. In order to reduce the influence of such exceptional behavior immediately after drilling is started, the pre-processing unit 5 excludes drilling data at positions shallower than a predetermined depth, such as 10 cm, from the surface of the rock mass G.

[0047] In addition to the above, the preprocessing unit 5 performs one or a combination of missing value processing, item selection processing, dimension reduction processing, and averaging processing. If an abnormality occurs in the sensor that acquires each piece of drilling data during drilling, the acquired values ​​may be inappropriate. Alternatively, if drilling is stopped midway, the hole will not be used for blasting, so it is not desirable to use the data related to that hole as drilling data. As a missing value processing method, the preprocessing unit 5 treats the drilling data acquired in such cases as error data and excludes the error data from the drilling data.

[0048] In addition, for example, when evaluating a certain evaluation item, it may be possible to obtain better evaluation results by using only some of the items in the punching data and not using other items. In such cases, the pre-processing unit 5 performs item selection processing by selecting some of the items from the multiple items included in the punching data and deleting the other items.

[0049] Furthermore, the pre-processing unit 5 reduces the number of items by extracting features from multiple items through statistical analysis such as principal component analysis to create new items as a dimension reduction process. In this way, the features of the drilling data can be more prominently expressed in fewer items, which may allow for efficient training of the second drill face evaluation model 12L and inference by the second trained drill face evaluation model 12.

[0050] Furthermore, the preprocessing unit 5 averages the multiple data obtained according to the depth of the hole by depth as an averaging process. In the state shown in FIG. 7, for example, data is acquired for each of eight items, from drilling speed to water pressure, at each of 14 different depths, resulting in a total of 112 data points for one hole. However, by averaging each item by hole depth, the number of data points corresponding to one hole can be reduced to eight, the same number as the number of items. In this way, by determining a representative value for each hole for each item of drilling data, the total number of drilling data points can be reduced, and the training of the second face evaluation model 12L and inference by the second trained face evaluation model 12 can be performed efficiently. Note that the representative value for each hole is not limited to the average value; the maximum value, minimum value, variance, etc. of each item may also be used as the representative value.

[0051] Finally, the preprocessing unit 5 calculates one value for each item of the drilling data by calculating the average across holes for each item of the drilling data. For example, if eight items, namely drilling speed, impact pressure, feed pressure, damper pressure, rotation speed, rotation pressure, water volume, and water pressure, are used as the drilling data and dimension reduction processing and item selection processing are not performed, the drilling data will have eight values ​​as a whole. In this way, through the processing by the pre-processing unit 5, learning punching data 111L is generated.

[0052] In the training data 20, the training drilling data 111L and the training image 101L are prepared to correspond to each other. For this reason, for example, if the training image 101L is an image from which the top T portion has been cut out, the pre-processing unit 5 selects only the holes corresponding to the top T portion of the drilling data, and generates training drilling data 111L corresponding to the top T portion, targeting only the selected holes. Similarly, for example, if the training image 101L is an image from which the right shoulder R or left shoulder L portion has been cut out, the pre-processing unit 5 selects only the holes corresponding to the right shoulder R or left shoulder L portion of the drilling data, and generates training drilling data 111L corresponding to the right shoulder R or left shoulder L portion, targeting only the selected holes.

[0053] As shown in Figure 4, the second face evaluation model 12L includes an input layer 12i, multiple intermediate layers 12m, and an output layer 12o. The second face evaluation model 12L is configured such that multiple layers are arranged in series, and the nodes constituting each layer are fully connected to each other between adjacent layers. The drilling data processed by the pre-processing unit 5 as described above is input to the input layer 12i of the second face evaluation model 12L by the learning unit 3 as training drilling data 111L. Therefore, for example, if the training drilling data 111L has eight values ​​as described above, the input layer 12i has eight input nodes 12m corresponding to each of these eight values.

[0054] In addition, the learning unit 3 inputs the learning image 101L corresponding to the learning drilling data 111L input to the second drill face evaluation model 12L in the learning data 20 to the first trained drill face evaluation model 11, and inputs the resulting provisional evaluation category 11T to the input layer 12i. As shown in Figure 4, if the number of evaluation categories corresponding to the evaluation items is, for example, five, and the first trained drill face evaluation model 11 performs five-class classification, the number of values ​​of the provisional evaluation category 11T is also five. Therefore, the input layer 12i has five input nodes 12m corresponding to each of these five values. As a result, in this case, the input layer 12i has, for example, a total of 13 (= 8 + 5) input nodes 12m.

[0055] When the training drilling data 111L and the provisional evaluation category 11T are input to the input layer 12bi, calculations are performed sequentially based on the weights set for the connecting lines between the nodes, tracing each layer in order from the input layer 12i to the intermediate layer 12m and the output layer 12o, and the final calculation results are stored in each of the output nodes 12n of the output layer 12bo. The output nodes 12n are provided corresponding to each of the evaluation categories. If the second face evaluation model 12L corresponds to an evaluation item having, for example, five evaluation categories, five output nodes 12n are provided, and the inference results 12LS during training stored in the output nodes 12n have five numerical values ​​corresponding to these five output nodes 12n.

[0056] The inference result 12LS during learning, like the inference result 11LS during learning, is expressed as a softmax function, which indicates which evaluation category, for example, from 1 to 5, the second face evaluation model 12L has inferred from the evaluation of the input learning drilling data 111L and the provisional evaluation category 11T. This inference result 12LS during learning is compared with the training data 120. This training data 120 is the correct value of the evaluation category for each evaluation item, prepared in correspondence with the learning image 101L input to the first face evaluation model 11L and the learning drilling data 111L input to the second face evaluation model 12L.

[0057] As described above, the second face evaluation model 12L is machine-learned by adjusting the weights and other values ​​of the second face evaluation model 12L by the backpropagation method or the gradient descent method so that the inference result 12LS during learning is compared with the teacher data 120 and the inference result 12LS during learning becomes closer to the teacher data 120. As a result, when the second face evaluation model 12L is input with the training drilling data 111L and the provisional evaluation category 11T estimated by inputting the corresponding training image 101L to the first trained face evaluation model 11, the second face evaluation model 12L is trained to output the inference result 12LS that is close to the teacher data 120 corresponding to the training drilling data 111L and the training image 101L.

[0058] In this way, the learning unit 3 generates a second trained face evaluation model 12 as a trained model that is used as a program module that is part of artificial intelligence software, with learning parameters such as weights trained.

[0059] As described above, the second trained face evaluation model 12 is trained and generated corresponding to each evaluation item by inputting the provisional evaluation category 11T inferred by the first face evaluation category inference unit 6 for the training image 101L and the training drilling data 111L at the time the training image 101L was obtained as training input data, and inferring the evaluation category corresponding to the training input data.

[0060] Next, the behavior of each component element when estimating the evaluation category will be described. When estimating the evaluation category, first, a tunnel face image is acquired by the image acquisition unit 2. For example, a worker evaluating a tunnel face takes an image 100 of the entire tunnel face of a tunnel under construction, as shown in Fig. 2. When image 100 of the entire tunnel face is input, image acquisition unit 2 generates and acquires tunnel face image 101 by cutting out each of the top T, right shoulder R, and left shoulder L, as explained with reference to Fig. 3 during learning.

[0061] Figure 8 is an explanatory diagram showing the flow of data when inferring an evaluation category using the first trained face evaluation model and the second trained face evaluation model. In this embodiment, the inference of the evaluation category using the first trained face evaluation model 11 and the second trained face evaluation model 12 is performed for each evaluation item: compressive strength, weathering, crack spacing, crack condition, running slope, amount of spring water, and deterioration due to spring water.

[0062] The first face evaluation category inference unit 6 inputs the tunnel face images 101 generated for each of the crown T, right shoulder R, and left shoulder L to each of the first trained face evaluation models 11 generated corresponding to each evaluation item. The first trained face evaluation model 11 has a configuration similar to that of the first face evaluation model 11L. When the tunnel face image 101 is input to the first trained face evaluation model 11, calculations are performed sequentially within the first trained face evaluation model 11 in the same manner as described during learning, and inference results 11S are output to each of the output nodes 11n of the output layer 11bo of the full connection unit 11b.

[0063] The inference result 11S, like the inference result 11LS during learning, is expressed as a softmax function, indicating which evaluation category, for example, from 1 to 5, the first trained tunnel face evaluation model 11 has inferred for the input tunnel face image 101. As already explained, the inference result 11S is converted into a dummy variable that takes on a variable of 0 or 1. That is, the value of the inference result 11S corresponding to the evaluation category inferred to have the highest probability is converted to 1, and other values ​​are converted to 0. In this way, the inference result 11S is treated as the result of a provisional evaluation of the tunnel face by the first trained tunnel face evaluation model 11 based only on the image, i.e., the external shape of the tunnel face, i.e., as a provisional evaluation category 11T.

[0064] As described above, the first tunnel face evaluation category inference unit 6 infers a provisional evaluation category 11T corresponding to each evaluation item from the tunnel face image 101 using the first trained tunnel face evaluation model 11 generated corresponding to each evaluation item. The first tunnel face evaluation category inference unit 6 infers a provisional evaluation category 11T for each of the top T, right shoulder R, and left shoulder L.

[0065] Next, the drilling data acquisition unit 4 acquires drilling data by the rock drill 200 when the above-mentioned tunnel face image 101 was obtained. More precisely, in order to obtain the tunnel face photographed as the above-mentioned tunnel face image 101, the hole is drilled by the rock drill 200, and then charged and blasted, and the drilling data during this drilling is acquired.

[0066] The acquired drilling data is corrected by the pre-processing unit 5, as in the case of learning. First, the pre-processing unit 5 excludes drilling data at positions shallower than a predetermined depth from the surface of the bedrock G, for example. Then, the pre-processing unit 5 performs one or a combination of missing value processing, item selection processing, dimension reduction processing, and averaging processing. In this embodiment, only the same processing as that performed by the pre-processing unit 5 during learning is performed. Finally, the pre-processing unit 5 calculates one value for each item of the drilling data by calculating the average between holes for each item of the drilling data.

[0067] When inferring the evaluation category, drilling data is generated for each of the top T, right shoulder R, and left shoulder L, just as during learning. The pre-processing unit 5 selects only the holes corresponding to the top T portion so as to correspond to the tunnel face image 101 of the top T, and generates drilling data corresponding to the top T portion for only the selected holes. Similarly, for the right shoulder R or left shoulder L portion, only the holes corresponding to the right shoulder R or left shoulder L portion are selected, and drilling data corresponding to each of the right shoulder R or left shoulder L portion is generated for only the selected holes.

[0068] The second face evaluation category inference unit 7 inputs the drilling data 111 corrected by the preprocessing unit 5 and the provisional evaluation category 11T inferred by the first face evaluation category inference unit 6 for the tunnel face image 101 to each of the second trained face evaluation models 12 generated corresponding to each evaluation item. The second trained face evaluation model 12 has a configuration similar to that of the second face evaluation model 12L. When the drilling data 111 and the provisional evaluation category 11T are input to the second trained face evaluation model 12, calculations are performed sequentially within the second trained face evaluation model 12 in the same manner as described during learning, and inference results 12S are output to each of the output nodes 12n of the output layer 12o.

[0069] The inference result 12S, like the inference result 12LS during learning, is expressed as a softmax function indicating which evaluation category, for example, 1 to 5, the second trained tunnel face evaluation model 12 has inferred for the input drilling data 111 and provisional evaluation category 11T. In the example of Figure 8, as a result of inferring the evaluation, for example, the probabilities that the evaluation categories are 1, 2, 3, 4, and 5 out of the five evaluation categories are 0.02, 0.81, 0.08, 0.06, and 0.03, respectively. Therefore, the second tunnel face evaluation category inference unit 7 infers that the evaluation category "2," which has the highest probability value of 0.81, is the evaluation category that takes into account the tunnel face image 101 and the drilling data 111.

[0070] As described above, the second face evaluation category inference unit 7 infers an evaluation category for each evaluation item that takes into account both the tunnel face image 101 and the drilling data 111 from the drilling data 111 and the provisional evaluation category 11T inferred by the first face evaluation category inference unit 6 for the tunnel face image 101, using the second learned face evaluation model 12 generated corresponding to each evaluation item. The second face evaluation division inference unit 7 performs the above-described processing for each of the top T, right shoulder R, and left shoulder L to infer the evaluation division.

[0071] The evaluation category tallying unit 8 tally the evaluation results of the second face evaluation category inference unit 7 for each evaluation item. For each evaluation item, the evaluation category tallying unit 8 calculates a value obtained by, for example, doubling the evaluation category inferred for the crest T, the sum of the evaluation category inferred for the right shoulder R, and the evaluation category inferred for the left shoulder L, divided by 4, and outputs this as the final result to a display device not shown.

[0072] Next, a tunnel face evaluation support method using the above-mentioned tunnel face evaluation support device will be described with reference to Figures 1 to 8, 9 and 10. Figure 9 is a flowchart showing the processing flow when learning the first face evaluation model and the second face evaluation model. Figure 10 is a flowchart showing the processing flow when inferring an evaluation category using the first trained face evaluation model and the second trained face evaluation model.

[0073] First, the process flow during learning of the evaluation category will be explained. The learning process of the first face evaluation model 11L and the second face evaluation model 12L shown in Fig. 9 is executed for each evaluation item, namely, compressive strength, weathering, crack spacing, crack condition, running slope, amount of spring water, and deterioration due to spring water.

[0074] When learning the evaluation category, learning data is prepared as shown in FIG. 9 (step S1). As the training data 20, a plurality of combinations of training images 101L, training drilling data 111L, and correct values ​​of the evaluation category as teacher data 120 are prepared. The training images 101L are images of a tunnel face. The training drilling data 111L is obtained by performing missing value processing, item selection processing, dimension reduction processing, averaging processing, etc. on the drilling data acquired by a rock drilling machine when the tunnel face photographed as the training image 101L was obtained in the pre-processing unit 5. Furthermore, the evaluation results of the tunnel face by, for example, a worker are prepared as correct values ​​of the evaluation category corresponding to the training images 101L and the training drilling data 111L.

[0075] The learning images 101L are generated by cutting out the top T, right shoulder R, and left shoulder L from each of the images 100, which were taken of the entire tunnel face during past mountain tunnel construction work. For each evaluation item, learning drilling data 111L and correct answer values ​​for the evaluation category are also prepared for each of the top T, right shoulder R, and left shoulder L.

[0076] Next, the learning unit 3 learns the first face evaluation model 11L for each evaluation item to generate the first learned face evaluation model 11 (step S2). For each evaluation item, the learning unit 3 inputs the learning image 101L into the first face evaluation model 11L corresponding to that evaluation item, and compares the output inference result 11LS during learning with the teacher data 120, i.e., the correct value of the evaluation category of that evaluation item corresponding to the learning image 101L. Then, the values ​​of the filter parameters, weights, etc. of the first face evaluation model 11L are adjusted by the backpropagation method or the gradient descent method so that the inference result 11LS during learning becomes close to the teacher data 120, thereby performing machine learning of the first face evaluation model 11L corresponding to each evaluation item.

[0077] Thereafter, the learning unit 3 learns the second face evaluation model 12L to generate the second trained face evaluation model 12 (step S3). For this purpose, the learning unit 3 inputs the training image 101L into the corresponding first trained cutting face evaluation model 11 for each evaluation item, and obtains the resulting provisional evaluation category 11T. The learning unit 3 inputs the training drilling data 111L and the provisional evaluation category 11T for the evaluation item obtained for the training image 101L corresponding to the training drilling data 111L into the second cutting face evaluation model 12L corresponding to each evaluation item. By adjusting the values ​​of the weights and the like of the second cutting face evaluation model 12L using the backpropagation method or the gradient descent method so that the inference result 12LS output during learning becomes a value close to the teacher data 120, machine learning of the second cutting face evaluation model 12L corresponding to each evaluation item is performed.

[0078] Next, the processing flow when estimating the evaluation category will be explained. In this embodiment, the inference of the evaluation category using the first trained face evaluation model 11 and the second trained face evaluation model 12 as shown in Fig. 10 is performed for each evaluation item, namely, compressive strength, weathering, crack spacing, crack condition, running slope, amount of spring water, and deterioration due to spring water.

[0079] When estimating the evaluation category, as shown in FIG. 10, first, the image acquisition unit 2 acquires a tunnel face image 101 (step S11). A worker who evaluates a tunnel face takes an image 100 of the entire tunnel face of a tunnel under construction. When image 100 of the entire tunnel face is input, image acquisition unit 2 generates and acquires tunnel face image 101 by cutting out each of the top edge T, right shoulder R, and left shoulder L, as explained with reference to FIG. 3 during learning.

[0080] Next, the first face evaluation category inference unit 6 infers a provisional evaluation category 11T corresponding to each evaluation item using the first trained face evaluation model 11 generated corresponding to each evaluation item (step S12). The first tunnel face evaluation category inference unit 6 inputs the tunnel face images 101 generated for each of the top T, right shoulder R, and left shoulder L into each of the first trained tunnel face evaluation models 11 generated corresponding to each evaluation item. The first tunnel face evaluation category inference unit 6 infers a provisional evaluation category 11T for each of the top T, right shoulder R, and left shoulder L for each evaluation item.

[0081] Next, the drilling data acquisition unit 4 acquires drilling data by the rock drill 200 when the tunnel face image 101 is obtained (step S13). The acquired drilling data is corrected by the pre-processing unit 5, as in the learning process. First, the pre-processing unit 5 excludes drilling data at positions shallower than a predetermined depth from the surface of the bedrock G, for example. Then, the pre-processing unit 5 performs one or a combination of missing value processing, item selection processing, dimension reduction processing, and averaging processing. The pre-processing unit 5 generates drilling data for each of the top T, right shoulder R, and left shoulder L.

[0082] Then, the second face evaluation category inference unit 7 inputs the drilling data 111 corrected by the preprocessing unit 5 and the provisional evaluation category 11T inferred by the first face evaluation category inference unit 6 corresponding to the evaluation category for the tunnel face image 101 to each of the second learned face evaluation models 12 generated corresponding to each evaluation item, and obtains an inference result 12S corresponding to each evaluation item. The second face evaluation category inference unit 7 infers that the evaluation category with the highest value among the inference results 12S for each evaluation item is the evaluation category that takes into account the tunnel face image 101 and the drilling data 111 (step S14). The second face evaluation category inference unit 7 performs the above-described processing for each of the top T, right shoulder R, and left shoulder L for each evaluation item to infer the evaluation category.

[0083] Finally, the evaluation category counting unit 8 counts and outputs the evaluation results of the second face evaluation category inference unit 7 for each evaluation item (step S15).

[0084] Next, a description will be given of the results of evaluating the accuracy of the above-described tunnel face evaluation support device 1. FIG. In Figure 11, the left and right graphs show the results of evaluating the crack spacing and crack condition as evaluation items, respectively. In each graph, the graph labeled "Photo AI" shows the accuracy rate, i.e., the percentage of results that match the evaluation category determined by the worker, when a trained model is constructed to evaluate each evaluation item based only on the external shape of the tunnel face and the trained model estimates the evaluation category. Similarly, the graph labeled "Drilling AI" shows the accuracy rate, when a trained model is constructed to evaluate each evaluation item based only on drilling data and the trained model estimates the evaluation category. Furthermore, the graph labeled "Photo & Drilling AI" shows the accuracy rate when a trained model is constructed to evaluate each evaluation item based on both tunnel face images and drilling data, as in the tunnel face evaluation support device 1 of this embodiment.

[0085] For example, with regard to crack spacing, the "Drilling AI" has a certain accuracy rate, but the accuracy rate of the "Photo AI" is low. Conversely, with regard to crack condition, the "Photo AI" has a certain accuracy rate, but the accuracy rate of the "Drilling AI" is low. In contrast, the "Photo & Drilling AI", which corresponds to the tunnel face evaluation support device 1 of this embodiment, takes into account both the tunnel face image 101 and the drilling data 111, and is therefore able to infer evaluation categories with higher accuracy than the "Photo AI" and the "Drilling AI" for both crack spacing and crack condition.

[0086] The tunnel face evaluation support device 1 as described above is a tunnel face evaluation support device 1 that supports the evaluation of a tunnel face, and includes an image acquisition unit 2 that acquires a tunnel face image 101, a first face evaluation category inference unit 6 that inputs a learning image 101L related to the tunnel face and infers a provisional evaluation category 11T from the tunnel face image 101 using a first trained face evaluation model 11 that has been trained to infer an evaluation category corresponding to the learning image 101L, a drilling data acquisition unit 4 that acquires drilling data 111 by a rock drill 200 when the tunnel face image 101 is obtained, and a first face evaluation category inference unit 6 that infers a provisional evaluation category 11T from the tunnel face image 101 using a first trained face evaluation model 11 that has been trained to infer an evaluation category corresponding to the learning image 101L. The tunnel face evaluation model 12 is equipped with a second face evaluation category inference unit 7 that inputs the training drilling data 111L obtained when the tunnel face image 101L is obtained and the provisional evaluation category 11T inferred by the first face evaluation category inference unit 6 for the training image 101L as training input data, and infers an evaluation category taking into account both the tunnel face image 101 and the drilling data 111 from the drilling data 111 and the provisional evaluation category 11T inferred by the first face evaluation category inference unit 6 for the tunnel face image 101.

[0087] According to the above configuration, the first trained face evaluation model 11 is trained to input a training image 101L related to a tunnel face and infer an evaluation category corresponding to the training image 101L. In other words, the first trained face evaluation model 11 is trained to evaluate a tunnel face based on its external shape. When a tunnel face image 101 is input to such a first trained face evaluation model 11, a provisional evaluation category 11T is inferred as an evaluation result based on the external shape of the tunnel face. The second trained tunnel face evaluation model 12 is trained to input the provisional evaluation category 11T inferred by the first tunnel face evaluation category inference unit 6 for the training image 101L and the training drilling data 111L when the training image 101L was obtained as training input data, and to infer an evaluation category corresponding to the training input data. That is, the second trained tunnel face evaluation model 12 is trained to evaluate the tunnel face taking into account both the external shape and properties of the tunnel face based on both the provisional evaluation category 11T, which is an evaluation result based on the external shape of the tunnel face, and the drilling data 111 obtained by the rock drill 200. When the drilling data 111 and the provisional evaluation category 11T inferred by the first tunnel face evaluation category inference unit 6 for the corresponding tunnel face image 101 are input to the second trained tunnel face evaluation model 12, an evaluation category is inferred that takes into account both the tunnel face image 101, i.e., the external shape, and the drilling data 111, i.e., the properties. In this way, the final assessment category takes into account both the external shape and properties of the tunnel face, which makes it possible to improve the accuracy of assessments for items that are difficult to assess sufficiently from the external shape of the tunnel face alone. In this way, it is possible to provide a tunnel face evaluation support device 1 that evaluates the tunnel face with high accuracy.

[0088] When constructing the tunnel face evaluation support device 1 described above, it is possible to consider a configuration in which the evaluation category is inferred based solely on the tunnel face image 101. However, depending on the evaluation category, a sufficient evaluation may not be possible based solely on the external shape of the tunnel face. For example, compressive strength, which indicates the strength of a rock mass, is evaluated based on the ease with which the rock mass cracks when struck with a hammer. However, information regarding the strength of the rock mass is difficult to obtain from the tunnel face image 101 alone, and therefore the accuracy of the inferred evaluation category may not be high. In contrast, in the tunnel face evaluation support device 1 of this embodiment, the inference result by the first trained face evaluation model 11 based on the tunnel face image 101 is input as a provisional evaluation category to the second trained face evaluation model 12 together with the drilling data 111, and the final evaluation category is output. By doing this, even if the inference result by the first trained face evaluation model 11, which is based only on the external shape of the tunnel face, is not valid, the second trained face evaluation model 12 can adjust this inference result based on the information in the drilling data 111 and output a more accurate evaluation category.

[0089] In addition, when an image 100 capturing the entire tunnel face is input, the image acquisition unit 2 generates a tunnel face image 101 by cutting out each of the tunnel's top T, right shoulder R, and left shoulder L.

[0090] According to the above configuration, by cutting out each of the tunnel crown T, right shoulder R, and left shoulder L from image 100, which is a photograph of the entire tunnel face, a tunnel face image 101 is generated as input to first trained face evaluation model 11. In this way, first trained face evaluation model 11 can evaluate the tunnel face for each of the tunnel crown T, right shoulder R, and left shoulder L, making it possible to perform a detailed evaluation of the tunnel face.

[0091] The drilling data 111 is one of the depth of the hole H, drilling speed, impact pressure N, feed pressure P, damper pressure, rotation speed, rotation pressure, water volume, and water pressure, or a combination of any of these.

[0092] According to the above configuration, the tunnel face can be evaluated in detail from the perspective of its properties, which is different from its external shape, using the drilling data 111 input as one or a combination of the hole depth, drilling speed, impact pressure N, feed pressure P, damper pressure, rotation speed, rotation pressure, water volume, and water pressure.

[0093] In addition, the tunnel face evaluation support device 1 further includes a pre-processing unit 5 that corrects the drilling data, and the second face evaluation category inference unit 7 inputs the drilling data 111 corrected by the pre-processing unit 5 into the second learned face evaluation model 12, and the pre-processing unit 5 performs one or a combination of the following: missing value processing that excludes error data in the drilling data; item selection processing that selects some items from multiple items included in the drilling data and deletes the others; dimensionality reduction processing that reduces the number of items by extracting features from multiple items through statistical analysis and creating new items; and averaging processing that averages multiple data obtained according to the hole depth.

[0094] According to the above configuration, when the preprocessing unit 5 executes missing value processing to remove erroneous data from the punching data, the erroneous data is prevented from being reflected in the evaluation. In addition, when the pre-processing unit 5 performs an item selection process in which some items are selected from multiple items contained in the drilling data and other items are deleted, if there is an item in the drilling data that has a negative impact on the evaluation, it can be deleted and excluded from the input to the second trained face evaluation model 12. Furthermore, when the pre-processing unit 5 performs dimension reduction processing, it extracts features from multiple items of the drilling data using statistical analysis such as principal component analysis to create new items. The new items generated in this way more clearly reflect the characteristics of the drilling data as a whole than the items before the dimension reduction processing was performed. Therefore, the evaluation accuracy of the second trained tunnel face evaluation model may be improved. At the same time, the dimension reduction processing reduces the total number of items. Therefore, the overall configuration of the tunnel face evaluation support device 1, including the second trained tunnel face evaluation model 12, may be simplified. When the pre-processing unit 5 performs an averaging process to average multiple pieces of data obtained according to the depth of the hole, the data obtained for each depth for one hole is aggregated into a single value, thereby reducing the number of drilling data corresponding to one hole. This can simplify the configuration of the tunnel face evaluation support device 1, including the second trained face evaluation model 12.

[0095] Furthermore, the above-mentioned tunnel face evaluation support program is a tunnel face evaluation support program that supports the evaluation of a tunnel face, and includes a computer including an image acquisition means for acquiring a tunnel face image 101, a first face evaluation category inference means for inferring a provisional evaluation category 11T from the tunnel face image 101 using a first trained face evaluation model 11 that is trained to input a learning image 101L related to the tunnel face and infer an evaluation category corresponding to the learning image 101L, and a drilling data acquisition means for acquiring drilling data 111 by a rock drill 200 when the tunnel face image 101 is obtained. and a second face evaluation category inference means for inferring an evaluation category taking into account both the tunnel face image 101 and the drilling data 111, from the drilling data 111 and the provisional evaluation category 11T inferred by the first face evaluation category inference means for the tunnel face image 101, using a second trained face evaluation model 12 trained to input the training drilling data 111L when the training image 101L was obtained and the provisional evaluation category 11T inferred by the first face evaluation category inference means for the training image 101L as training input data.

[0096] According to the above-described configuration, as already explained with respect to the tunnel face evaluation support device 1, the tunnel face can be evaluated with high accuracy.

[0097] The tunnel face evaluation support method as described above is a tunnel face evaluation support method that supports the evaluation of a tunnel face, and includes an image acquisition step of acquiring a tunnel face image 101, a first face evaluation category inference step of inputting a learning image 101L related to the tunnel face and inferring a provisional evaluation category 11T from the tunnel face image 101 using a first trained face evaluation model 11 that has been trained to infer an evaluation category corresponding to the learning image 101L, a drilling data acquisition step of acquiring drilling data 111 by a rock drill 200 when the tunnel face image 101 is obtained, and and a second face evaluation category inference step in which the training drilling data 111L obtained when the image 101L was obtained and a tentative evaluation category 11T inferred for the training image 101L by the first face evaluation category inference step are input as training input data, and an evaluation category taking into account both the tunnel face image 101 and the drilling data 111 from the drilling data 111 and the tentative evaluation category 11T inferred for the tunnel face image 101 by the first face evaluation category inference step is inferred by a second trained face evaluation model 12 trained to infer an evaluation category corresponding to the training input data.

[0098] According to the above-described configuration, as already explained with respect to the tunnel face evaluation support device 1, the tunnel face can be evaluated with high accuracy.

[0099] The tunnel face evaluation support device, tunnel face evaluation support program, and tunnel face evaluation support method of the present invention are not limited to the above-mentioned embodiments described with reference to the drawings, and various other modifications are possible within the technical scope.

[0100] For example, in the above embodiment, the drilling data is subjected to a certain processing by the pre-processing unit 5, and the corrected results are input as the training drilling data 111L or the drilling data 111 to the second face evaluation model 12L or the second trained face evaluation model 12. However, this is not limited to this. For example, different items may be output from the drilling data depending on the type of rock drill 200 or the manufacturer of the rock drill 200. For this reason, for example, when drilling data is acquired using multiple different types of rock drills 200, it may be necessary to convert or translate the numerical values ​​of one piece of drilling data to match those of the other piece of drilling data. The pre-processing unit 5 may be configured to perform processing to standardize data between such drilling data that may have different items.

[0101] Alternatively, in the above embodiment, the drilling data is configured so that one value for each item is input to the second face evaluation model 12L and the second learned face evaluation model 12 by calculating the average value according to the hole depth and for multiple holes, but this is not limited to this. For example, the tunnel face evaluation support device may be configured not to include the pre-processing unit 5, and the data acquired by the drilling data acquisition unit 4 may be input directly as drilling data to the second face evaluation model and the second learned face evaluation model.

[0102] In addition to this, it is possible to select and discard the configurations given in the above embodiments, or to change them to other configurations as appropriate. [Explanation of symbols]

[0103] 1. Tunnel face evaluation support device 2. Image acquisition unit 4. Drilling data acquisition section 5 Pretreatment section 6 First face evaluation classification inference section 7. Second face evaluation classification inference section 11 First trained face evaluation model 11T Provisional Evaluation Category 12 Second trained face evaluation model 100 Images of the entire tunnel face 101 Tunnel face image 101L Learning Images 111 Drilling Data 111L Training drilling data 200 Jackhammer H hole L left shoulder R right shoulder T top edge

Claims

1. A tunnel face evaluation support device that supports the evaluation of a tunnel face, an image acquisition unit for acquiring a tunnel face image; a first face evaluation category inference unit that receives a learning image of a tunnel face and infers a provisional evaluation category from the tunnel face image using a first trained face evaluation model that has been trained to infer an evaluation category corresponding to the learning image; a drilling data acquisition unit that acquires drilling data by a rock drill when the tunnel face image is obtained; a second face evaluation category inference unit that inputs the training drilling data when the training image was obtained and the provisional evaluation category inferred by the first face evaluation category inference unit for the training image as training input data, and infers the evaluation category taking into account both the tunnel face image and the drilling data from the drilling data and the provisional evaluation category inferred by the first face evaluation category inference unit for the tunnel face image using a second trained face evaluation model that has been trained to infer the evaluation category corresponding to the training input data; A tunnel face evaluation support device equipped with:

2. When an image of the entire tunnel face is input, the image acquisition unit generates the tunnel face image by cutting out each of the top end, right shoulder, and left shoulder of the tunnel. A tunnel face evaluation support device according to claim 1.

3. The tunnel face evaluation support device according to claim 1, wherein the drilling data is one of hole depth, drilling speed, impact pressure, feed pressure, damper pressure, rotation speed, rotation pressure, water volume, and water pressure, or any combination thereof.

4. The drilling data corrected by the preprocessing unit is further corrected by a preprocessing unit, and the second face evaluation category inference unit inputs the drilling data corrected by the preprocessing unit into the second learned face evaluation model, The pre-treatment unit Missing value processing to exclude error data in the punched data; an item selection process for selecting some of the items from among a plurality of items included in the punching data and deleting the other items; A dimension reduction process is performed to reduce the number of items by extracting features from the multiple items through statistical analysis and creating new items. an averaging process for averaging a plurality of data obtained according to the depth of the hole; The tunnel face evaluation support device according to claim 1, wherein the device executes any one of the above or any combination thereof.

5. A tunnel face evaluation support program that supports the evaluation of a tunnel face, On the computer, an image acquisition means for acquiring an image of a tunnel face; a first face evaluation category inference means for inferring a provisional evaluation category from a tunnel face image by inputting a learning image related to the tunnel face and using a first trained face evaluation model that has been trained to infer an evaluation category corresponding to the learning image; a drilling data acquisition means for acquiring drilling data by a rock drill when the tunnel face image is obtained; a second face evaluation category inference means for inferring the evaluation category taking into account both the tunnel face image and the drilling data from the drilling data and the provisional evaluation category inferred by the first face evaluation category inference means for the tunnel face image, using a second trained face evaluation model that is trained to input the training drilling data when the training image is obtained and the provisional evaluation category inferred by the first face evaluation category inference means for the training image as training input data and infer the evaluation category corresponding to the training input data; A tunnel face evaluation support program to achieve this.

6. A tunnel face evaluation support method for supporting evaluation of a tunnel face, comprising: an image acquisition step of acquiring a tunnel face image; a first face evaluation category inference step in which a learning image of a tunnel face is input and a first trained face evaluation model is used to infer a provisional evaluation category from the tunnel face image, the first trained face evaluation model being trained to infer an evaluation category corresponding to the learning image; a drilling data acquisition step of acquiring drilling data by a rock drill when the tunnel face image is obtained; a second face evaluation category inference step in which the training drilling data when the training image was obtained and the provisional evaluation category inferred for the training image by the first face evaluation category inference step are input as training input data, and the second trained face evaluation model is trained to infer the evaluation category corresponding to the training input data, and the second trained face evaluation model infers the evaluation category taking into account both the tunnel face image and the drilling data from the drilling data and the provisional evaluation category inferred for the tunnel face image by the first face evaluation category inference step; A tunnel face evaluation support method including:

Images