Soil identity determination device, soil identity determination method, and model generation device

Abstract

To provide a soil identity determination device and a model generation device that realize highly accurate similarity determination and soil estimation.SOLUTION: A soil identity determination device 1 according to one aspect of an embodiment includes an obtaining unit 101 and an estimation unit 102. The obtaining unit 101 obtains two processing object images regarding sediment excavated during pile construction. The estimation unit 102 estimates similarity information between the two processing object images by inputting the two processing object images obtained by the obtaining unit 101 into a learned model that is trained to receive two images regarding sediment and to output similarity information of the images.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a soil identity determination device, a soil identity determination method, and a model generation device. [Background technology]

[0002] The classification of each type of soil excavated in a boring survey is determined by workers, and there is variation depending on experience, etc. In order to reduce variation, for example, the method of Patent Document 1 can be used to determine the classification of the excavated soil by computer. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2004-219166 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, in pile construction, workers also judge whether the "two pieces of soil" are the same (i.e., identical), and there is variation depending on experience, etc. For this reason, there is a demand for a computer to be able to determine whether two pieces of soil are identical or not.

[0005] An object of the present invention is to realize highly accurate similarity determination and soil type estimation. [Means for solving the problem]

[0006] (1) According to one aspect of the present invention, a soil identity determination device includes an acquisition unit and an estimation unit. The acquisition unit acquires two processing target images related to soil excavated during pile construction. The estimation unit estimates similarity information between the two processing target images by inputting the two processing target images acquired by the acquisition unit into a trained model that is trained to receive the two images related to soil and output similarity information between the images. According to the above configuration (1), for example, a method of determining whether two soils are the same using a computer focuses on how similar the two soils are (similarity), and determines that they are the same if the similarity information indicating the degree of similarity is sufficiently similar. For example, by estimating whether the images to be processed are similar using a trained model, it is possible to estimate with high accuracy whether the soils are of the same soil quality.

[0007] (2) In some embodiments, in (1) above, the acquisition unit acquires two images as the two images to be processed: an image of soil collected in a boring survey, an image of soil collected during excavation of the test pile, and an image of soil excavated during construction of the test pile. According to the above configuration (2), for example, by estimating the similarity of soil between two images using a trained model, a user can understand whether the soil has the same soil quality.

[0008] (3) In some embodiments, in the above (1), the acquisition unit acquires, as the two images to be processed, two images of each of the soils and sands collected in a boring survey at different locations, two images of each of the soils and sands collected during excavation of a test pile at different locations, or two images of each of the soils and sands excavated during construction of a main pile at different locations. According to the above configuration (3), for example, by estimating the similarity of soil between two images using a trained model, a user can understand whether the soil has the same soil quality.

[0009] (4) In some embodiments, in (1) to (3) above, the estimation unit uses the trained model to extract features for each of the two images to be processed, and if the similarity of the features is equal to or greater than a threshold, it determines that the soil and sand in the two images are similar. According to the above configuration (4), for example, by estimating the similarity of soil between two images using a trained model, a user can understand whether the soil has the same soil quality.

[0010] (5) In some embodiments, the soil identity determination device according to (1) above further includes a display control unit that displays the similarity information. According to the above configuration (5), the user can easily know whether the soil and sand are of the same soil quality or not.

[0011] (6) In some embodiments, in the above (4), the estimation unit estimates soil type classifications for the two processing target images from the feature amounts. The soil identity determination device further includes a display control unit that displays the determination result and the soil type classifications of the processing target images. According to the above configuration (6), for example, by displaying the similarity of soil and estimating and displaying the soil type classification, the user can understand whether the soil has the same soil type.

[0012] (7) One embodiment of the present invention provides a soil identity determination method, which acquires two images to be processed related to soil excavated during pile construction, and inputs the two acquired images to be processed into a learned model that is trained to receive the two images of soil and output similarity information of images, thereby estimating similarity information between the two images to be processed. According to the above configuration (7), for example, a method of determining whether two soils are the same by a computer focuses on how similar the two soils are (similarity), and if they are sufficiently similar (if similarity information indicating the degree of similarity is equal to or greater than a threshold), it is determined that they are the same. For example, by estimating whether the images to be processed are similar using a trained model, it is possible to estimate with high accuracy whether the soils are of the same soil quality.

[0013] (8) According to an aspect of the present invention, a model generation device includes an acquisition unit and a training unit. The acquisition unit acquires images of two types of soil and sand. The training unit uses the images of the two types of soil and sand as input data, and trains a machine learning model to output similarity information between the two input images using learning data in which similarity information including at least one of whether the two types of images are similar and the similarity of the two types of images is used as correct answer data, thereby generating a trained model. According to the above configuration (8), for example, a trained model capable of estimating soil similarity with high accuracy can be generated. Effect of the Invention

[0014] According to the present invention, it is possible to realize highly accurate similarity determination and soil type estimation. [Brief description of the drawings]

[0015] [Figure 1] FIG. 1 is a block diagram showing a soil identity determination device according to the first embodiment. [Diagram 2] FIG. 2 is a flowchart showing the first estimation process according to the first embodiment. [Diagram 3] FIG. 3 is a diagram illustrating an example of input and output to a trained model related to the first estimation process. [Figure 4] FIG. 4 is a flowchart showing the second estimation process according to the first embodiment. [Diagram 5] FIG. 5 is a diagram illustrating an example of input and output to a trained model related to the second estimation process. [Figure 6] FIG. 6 is a flowchart showing the third estimation process according to the first embodiment. [Figure 7] FIG. 7 is a diagram illustrating an example of input and output to a trained model related to the third estimation process. [Figure 8] FIG. 8 is a block diagram showing a model generating device according to the second embodiment. [Figure 9] FIG. 9 is a flowchart showing a trained model generation process of the model generation device according to the second embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of input and output to a machine learning model that executes a similarity determination process. [Figure 11] FIG. 11 is a diagram illustrating an example of input and output to a machine learning model that executes a soil quality estimation process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, a soil identity determination device, a soil identity determination method, and a model generation device according to the embodiments of the present invention will be described in detail with reference to the drawings. Note that in the following embodiments, parts with the same numbers perform similar operations, and redundant explanations will be omitted.

[0017] (First embodiment) The soil identity determination device according to the first embodiment will be described with reference to the block diagram of FIG. The soil identity determination device 1 shown in the first embodiment includes a processing circuit 10, a storage unit 11, and a communication interface 12, which are connected to each other via a bus.

[0018] The processing circuit 10 is a processor and is realized by, for example, any one or a combination of a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), etc. The processing circuit 10 includes an acquisition unit 101, an estimation unit 102, and a display control unit 103.

[0019] The acquisition unit 101 acquires two images to be processed, which are images relating to soil excavated for construction of piles. The estimation unit 102 estimates similarity information between the two images to be processed by inputting the two images to be processed acquired by the acquisition unit into a trained model that is trained to receive two images of soil and to output similarity information between the two images. The similarity information is information that includes at least one of whether the two images are similar or not and the degree of similarity between the two images. A display control unit 103 displays the processing target image and the similarity information estimated by the estimation unit 102 on, for example, a display.

[0020] The storage unit 11 is configured with a storage medium such as a hard disk drive (HDD) or a solid state drive (SSD), and stores the image to be processed and the trained model after training. Note that, without being limited to these, other data related to pile construction and information obtained by boring survey (columnar diagram, N value), etc. may be stored in association with each other.

[0021] The communication interface 12 is an interface for transmitting and receiving data between devices that comply with a predetermined communication standard. The communication standard may be a wireless network such as a wireless LAN (Local Area Network) that complies with Wi-Fi (registered trademark) or Bluetooth (registered trademark), or a wired network using a LAN cable or the like.

[0022] Next, the first estimation process of the soil identity determining device 1 according to the first embodiment will be described with reference to the flowchart of FIG.

[0023] In step SA1, the acquisition unit 101 acquires two types of images as images to be processed. For example, the acquisition unit 101 may acquire two images from among an image of soil collected by a boring survey, an image of soil collected during excavation of a test pile, and an image of soil excavated by construction of the main pile. Alternatively, the acquisition unit 101 may acquire two images of soil collected by a boring survey at different points, two images of soil collected during excavation of a test pile at different points, or two images of soil excavated by construction of the main pile at different points. In other words, the comparison can be made using images from any stage of the boring survey, the construction of the test pile, and the construction of the main pile.

[0024] In step SA2, the estimation unit 102 extracts features for each of the two images using a trained model. For example, features obtained by feature extraction processing in a convolutional neural network may be used, and a trained model capable of extracting features from an image may be used.

[0025] In step SA3, the estimation unit 102 determines whether the similarity of the feature amounts between the two images is equal to or greater than a threshold. The similarity of the feature amounts may be determined, for example, by using at least one of the following as an index: the magnitude of the feature amount value, the closeness of the value, or the ratio of the types of the feature amount. If the feature amount is expressed as a vector, a general method for calculating the similarity, such as the cosine distance, may be used. If the similarity is equal to or greater than the threshold, the process proceeds to step SA4. If the similarity is less than the threshold, the process proceeds to step SA5.

[0026] In step SA4, the estimation unit 102 determines that the soil types are the same since the similarity is equal to or greater than the threshold value. In step SA5, the estimation unit 102 determines that the soil types are different because the similarity is less than the threshold value. In step SA6, the display control unit 103 displays the similarity information of the processing target image by the estimation unit 102 as a determination result. Note that estimated soil information of the soil (hereinafter, estimated soil information) may be displayed in addition to the determination result. Furthermore, in the example of FIG. 2, a comparison of the similarity between two images has been described, but a similar process may be used to compare the similarity between three or more images. The processes from step SA2 to step SA5 may be performed by a trained model. That is, two images may be input to the trained model as processing target images, and the trained model may output similarity information for the two images.

[0027] Next, an example of input and output to the trained model related to the first estimation process will be described with reference to the conceptual diagram of FIG. In the example of Figure 3, two images to be processed are input to a trained model 301: a test pile image 302, which is an image of soil collected during excavation of a test pile, and a main pile image 303, which is an image of soil excavated during the construction of the pile. The image to be processed is an image taken by an optical imaging device with the soil and sand in a plate. It is also assumed that the image is taken with a scale attached so that the size of the soil and sand can be measured. It is not limited to a scale, and any reference object such as a pen or notebook can be taken with the soil and sand as long as it is possible to understand that the size of the soil and sand to be measured is understood. Furthermore, if the size of the plate used in the image as the learning data of the trained model and the plate used in the image to be processed are clear, the grain shape of the sand and gravel can be understood in common during training and inference of the machine learning model, so an image of the soil and sand alone can be taken without using a scale. The trained model 301 extracts the features of the test pile image 302 and the main pile image 303, respectively, and calculates the similarity of the features. Furthermore, here, estimated soil information is also estimated. The trained model 301 outputs a judgment result 304.

[0028] The judgment result 304 includes similarity information 305. Here, it is assumed that the threshold for judging that the soil types are the same is 80%, and since the similarity is 85%, a message such as "Same soil type" is output as information on whether the two images of soil and sand have the same soil type as one of the judgment results 304, along with the numerical value of the similarity "85%" as similarity information 305. In other words, the judgment result 304 shows that the test pile image 302 and the actual pile image 303 have the same soil type. Furthermore, the estimated soil type information 306 "80% gravel, 15% sand" may be output together. This allows the user to understand the extent to which the soils are similar and the proportions of the soil types that are similar, as the basis for determining that the soil types are the same. Similarly, if the soil types are determined to be different, the user can understand the extent to which they are dissimilar and the differences in the proportions of the soil types.

[0029] Next, the second estimation process of the soil identity determination device 1 according to the first embodiment will be described with reference to the flowchart in Fig. 4. The second estimation process is a process in which, for example, an image of a test pile near a bearing layer is used as a reference image, the trained model is made to learn the features of the reference image, and the similarity between the image to be processed and the reference image is estimated.

[0030] In step SB1, one image to be processed is obtained. In step SB2, a trained model that has trained a certain reference image is used to estimate the similarity between the processing target image and the reference image, and similarity information is generated. Specifically, since the trained model has trained the features of the reference image, it is sufficient to compare the features of the reference image with the features of the processing target image and output whether they are similar or not. In step SB3, the display control section 103 displays the similar information on a display or the like.

[0031] Next, an example of input and output to the trained model related to the second estimation process will be described with reference to the conceptual diagram of Fig. 5. Here, as the reference image, an image of a test pile near the bearing layer, as shown in Fig. 3, is assumed. 5, a processing target image 502, for example, a real pile image, is input to a trained model 501. The trained model 501 extracts the feature amount of the processing target image 502 and calculates the similarity with the feature amount of the test pile image. The trained model 501 outputs similarity information 503 related to the calculated similarity.

[0032] Here, as an example of similarity information 503, information such as "Similarity with the soil quality of the soil collected at the time of construction of the test pile: 90%" is output. Note that a soil quality judgment result such as "Same soil quality" may be output, as in judgment result 304 in FIG. 3. This makes it clear that the test pile image and the processing target image have the same soil quality, and therefore it can be determined that the supporting layer has been reached. Furthermore, by also outputting the similarity value, it is possible to grasp the degree to which the soil is similar. Next, the third estimation process of the soil identity determining device 1 according to the first embodiment will be described with reference to the flowchart of FIG.

[0033] In step SC1, the acquisition unit 101 acquires an image to be processed. Specifically, in the construction of a cast-in-place pile, an image of soil and sand excavated with a drilling bucket is acquired by an optical imaging device such as a camera.

[0034] In step SC2, the estimation unit 102 inputs the processing target image into the trained model, and generates estimated soil information from the feature amounts in the processing target image.

[0035] In step SC3, the display control unit 103 outputs the processing target image and the estimated soil property information to a display or the like.

[0036] Next, an example of input and output to the trained model related to the third estimation process will be described with reference to the conceptual diagram of FIG. The processing target image 502 is input to the trained model 701. The trained model 701 outputs estimated soil quality information 702, which estimates the type of soil of the soil in the processing target image 502. As the estimated soil quality information 702, information such as clay, silt, sand, and gravel is output, and in this example, "soil quality: gravel" is output. The estimated soil quality information 702 may be output as text information, or may be output as an image in which the estimated soil quality information 702 is superimposed on the processing target image 502. Furthermore, for example, the probability value of each class in the output layer of the trained model may be output as the probability of the estimated soil quality. For example, the soil quality classification and a numerical value representing the accuracy of the soil quality estimation as a percentage, such as "estimated soil quality (probability) gravel 80%, sand 15%, other 5%" may be output as the estimated soil quality information 702. Note that the "soil quality classification" corresponds to the classification name or symbol of soil material in a boring log diagram, or the classification name or symbol in an engineering classification system of soil material.

[0037] According to the first embodiment described above, by inputting two images as processing target images into the trained model, it is possible to estimate the similarity between the two images and determine whether or not they have the same soil type. In addition, by outputting the estimated soil type information together, it is possible to present the reason why the soil type is determined to be the same, and it is possible to output an estimation result that is convincing to the user. In addition, by inputting the image to be processed into a trained model that has learned from borehole images of soil and sand near the supporting layer, and estimating the similarity using the extracted features, the similarity between the image to be processed and the borehole image can be determined, and can also be used to determine the supporting layer. In addition, by inputting an image of the soil to be processed as the image to be processed into the trained model, estimated soil quality information can be generated in which the soil quality of the soil in the image to be processed is estimated. A geological stratum is a layer of soil of multiple types, and there may be multiple soils of the same soil type at different depths. Existing trained models for soil type estimation (soil type estimation AI) judge these to be the same soil, but when constructing cast-in-place piles, it is necessary to judge them as different soils. For this reason, the present invention is an AI that judges the identity of two soils based on similarity information that includes more than just soil type features. By using a trained model in this way, it is possible to achieve highly accurate similarity determination and soil type estimation.

[0038] Second embodiment In the second embodiment, a configuration and operation of a model generation device that generates a trained model used in the soil identity determination device according to the first embodiment will be described.

[0039] The model generating device according to the second embodiment will be described with reference to the block diagram of FIG. The model generating device 2 according to the second embodiment includes a processing circuit 20, a storage unit 21, and a communication interface 22, which are connected to each other via a bus.

[0040] The processing circuit 20 is assumed to be a processor similar to the processing circuit 10 of the first embodiment. The processing circuit 20 includes an acquisition unit 201 and a training unit 202.

[0041] The acquiring unit 201 acquires two images of soil excavated in the past for pile construction, and similarity information related to the two images. Examples of the similarity information include at least one of whether the two types of images are similar or not and the similarity (e.g., 0% to 100%) of the two types of images. If estimated soil information is also output as the trained model, the soil information is also acquired. As the soil information, for example, information on the classification of soil determined by visual inspection by a person such as a worker at the depth at which the images were taken may be used. Note that the soil information may be determined by carrying out a soil grain size test.

[0042] The training unit 202 uses the image of soil and sand acquired by the acquisition unit 201 as input data, and trains the machine learning model to output similar information of the input image using learning data in which the corresponding similar information (including soil quality information, if necessary) is used as correct answer data. After the training is completed, a trained model is generated. The machine learning model may be a model of an architecture generally used in machine learning, such as a deep convolutional neural network or Transformer.

[0043] The storage unit 21 is a storage medium similar to the storage unit 11 of the first embodiment, and stores a plurality of pieces of learning data, a machine learning model before training, and a trained model after training. The communication interface 22 is similar to the communication interface 12 of the first embodiment.

[0044] Next, a trained model generation process of the model generation device 2 according to the second embodiment will be described with reference to the flowchart of FIG.

[0045] In step SD1, the acquisition unit 201 acquires learning data from the storage unit 21. In step SD2, the training unit 202 trains the machine learning model using the learning data. Specifically, an image of soil and sand is input to the machine learning model, and a predicted value of similar information is output by the machine learning model. The error between the predicted value of similar information and the similar information that is the correct answer data is calculated as a loss value. After that, the parameters of the machine learning model (network weighting coefficient, bias, etc.) can be updated using a loss function so that the loss value is minimized. Here, since it is assumed that soil types are classified as similar or not, for example, a binary cross entropy error can be used as the loss function of the machine learning model.

[0046] In step SD3, the training unit 202 determines whether the training of the machine learning model has been completed. Whether the training of the machine learning model has been completed may be determined to be completed, for example, when the loss value is equal to or less than a threshold and a predetermined number of epochs of training have been executed. When the training is completed, a trained model is generated. When the training of the machine learning model has been completed, the process proceeds to step SD4, and when the training of the machine learning model has not been completed, the process returns to step SD2 and the same process is repeated.

[0047] In step SD4, the storage unit 21 stores the trained model. The trained model may be transmitted to the outside via the communication interface 22.

[0048] Next, an example of input and output to the machine learning model that executes the similarity determination process will be described with reference to FIG. For the machine learning model 1001, a photographed image 1002 of soil in the learning data and a reference image 1003, which is an image of soil near a support layer collected during construction of a test pile, are input as input data, and similarity information 1004 between the photographed image 1002 and the reference image 1003 is input as correct answer data, and training is performed by the method shown in step SD2. As a result of training, a trained model 301 or a trained model 501 can be generated. In addition, when generating the trained model 301, multiple reference images 1003 are prepared, and training data is designed so that the similarity between the photographed image 1002 and various images can be learned. On the other hand, when generating the trained model 501, data augmentation is permitted in order to assume a specific soil image, such as a support layer judgment, but one representative reference image 1003 is prepared, and training data is designed so that the similarity between the reference image 1003 and the photographed image 1002 can be learned.

[0049] Next, an example of input and output to the machine learning model that executes the soil quality estimation process will be described with reference to FIG. A photographed image 1101 of soil and sand in the learning data is input as input data, and soil information 1102 is input as correct answer data to the machine learning model 1001, and training is performed in the manner shown in step SD2. As a result of the training, a trained model 701 is generated. In addition to the soil information 1102, excavation-related data 1103 such as GPS information, which is position information when the photographed image 1101 was acquired, columnar diagram information obtained by boring survey, and values ​​such as excavation torque and current value during excavation may be further added as correct answer data. In this way, by including information other than the photographed image 1101, such as excavation position information related to the photographed image 1101 and soil quality in layers before and after the stratum related to the photographed image 1101, knowledge related to the photographed image and soil quality under various conditions can be included, and it is expected that the versatility of the machine learning model can be improved.

[0050] Note that, although an example of generating a trained model using supervised learning is shown here, this is not limited to this, and unsupervised learning based only on input data may also be used, or the trained models 301, 501, 701 may be generated by contrastive learning using positive and negative examples.

[0051] According to the second embodiment described above, a trained model is generated by training a machine learning model using learning data including soil quality information determined by an operator on an image of soil and sand as correct answer data. Alternatively, a trained model is generated by training a machine learning model using learning data including a photographed image, a reference image, and a similarity between the photographed image and the reference image. This makes it possible to realize highly accurate soil quality estimation and similarity determination using the trained model.

[0052] Although the soil identity determination device 1 and the model generating device 2 are described as separate entities, this is not limiting, and the soil identity determination device 1 may include the functions of the model generating device 2. In other words, the processing circuit 10 related to the soil identity determination device 1 may include the training unit 202. This allows model learning and inference to be performed by a single device.

[0053] The instructions shown in the processing procedures shown in the above-mentioned embodiments can be executed based on a program, which is software. A general-purpose computer system can store this program in a recording medium in advance and obtain the same effect as that of the above-mentioned identification device by reading the stored program. Furthermore, the recording medium in this embodiment is not limited to a medium independent of a computer or an embedded system, but also includes a recording medium in which a program transmitted via a LAN, the Internet, or the like is downloaded and stored or temporarily stored. [Explanation of symbols]

[0054] 1...soil identity determination device, 2...model generation device, 10,20...processing circuit, 11,21...storage unit, 12,22...communication interface, 101,201...acquisition unit, 102...estimation unit, 103...display control unit, 202...training unit, 301,501,701...trained model, 304...determination result, 305...similar information, 306...estimated soil information, 502...image to be processed, 302...test pile image, 303...main pile image, 503,1004...similar information, 1001...machine learning model, 1003...reference image, 1002,1101...captured image, 1102...soil information, 1103...excavation-related data.

Claims

1. An acquisition unit that acquires two processing target images relating to soil and sand excavated in the construction of piles; An estimation unit that estimates similarity information between the two processing target images by inputting the two processing target images acquired by the acquisition unit into a trained model that is trained to receive two images of soil and sand and output similarity information of the images; A soil identity determination device comprising:

2. The soil identity determination device of claim 1, wherein the acquisition unit acquires two images as the two images to be processed: an image of soil and sand collected in a boring survey, an image of soil and sand collected during excavation of a test pile, and an image of soil and sand excavated during construction of the test pile.

3. The soil identity determination device of claim 1, wherein the acquisition unit acquires as the two images to be processed two images of each of the soils and sands collected in a boring survey at different locations, two images of each of the soils and sands collected during excavation of test piles at different locations, or two images of each of the soils and sands excavated during construction of main piles at different locations.

4. The soil identity determination device according to any one of claims 1 to 3, wherein the estimation unit uses the trained model to extract features for each of the two images to be processed, and if the similarity of the features is equal to or greater than a threshold, determines that the soil and sand related to the two images are similar.

5. The soil identity determination device according to claim 1 , further comprising a display control unit that displays the similarity information.

6. The estimation unit estimates a soil type classification for the two processing target images from the feature amount, 5. The soil identity determination device according to claim 4, further comprising a display control unit that displays a result of the soil similarity determination and a soil classification of the processing target image.

7. Two images to be processed are obtained, which relate to the soil excavated during pile construction. Two images of soil and sand are input, and the two acquired images to be processed are input to a trained model that is trained to output similarity information of the images, thereby estimating similarity information between the two images to be processed. Soil quality estimation method.

8. An acquisition unit for acquiring images relating to two types of soil and sand; A training unit that uses images of two types of soil and sand as input data, trains a machine learning model to output similarity information between the two input images using learning data that includes similarity information including at least one of whether the two types of images are similar and the similarity of the two types of images as correct answer data, and generates a trained model; A model generating device comprising:

Images