JP7376374B2 - Method for estimating soil particle size distribution - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act - Okumura Gumi Technical Research Annual Report No. 45 (Published by Okumura Gumi Co., Ltd. on September 1, 2020)

The present invention relates to a method for estimating the particle size distribution of soil, and particularly to a method for estimating the particle size distribution of soil, which estimates the particle size distribution of sampled soil using a trained model obtained by machine learning using artificial intelligence.

The particle size distribution of soil is an important indicator for determining the mechanical properties and compaction characteristics of the ground, but it is difficult to frequently measure the particle size distribution of soil that is generated in large quantities at the site. . An estimation method has also been proposed in which the contours of particles are measured by image analysis, but this method cannot be applied to soil that contains small particles with unclear contours. If the particle size distribution of soil, including small particles, can be estimated using a simple method, it is thought that it will lead to improved productivity in construction work.

In addition, the particle size distribution of soil is one of the important indicators for estimating properties such as soil compaction characteristics, water permeability, and liquefaction strength. To measure particle size distribution, it is necessary to perform sieve analysis and sedimentation analysis, but it is difficult to frequently perform these analyzes on the large amounts of soil generated during the construction process at construction sites. Therefore, if the particle size distribution of soil can be easily estimated, it will lead to improved productivity in construction work and improved quality of buildings and structures.

Against this background, there are many examples of research aimed at automatically estimating particle size distribution from images of soil surfaces.A typical method is to extract the contours of soil particles on the soil surface through image processing. There are methods for estimating the particle size and particle size distribution from this (for example, see Non-Patent Document 1). However, this method cannot be applied to small particles whose contours cannot be captured from the image, and is limited to estimating the particle size distribution of large particles.

In addition, in recent years, related research has been carried out using convolutional neural networks (hereinafter referred to as CNN), which is a type of deep learning that is known to exhibit high performance in the field of image classification. A research example of estimating the properties of soil from an image of generated soil has been reported (see, for example, Non-Patent Document 2).

Japan Society of Civil Engineers 67th Annual Academic Conference Summaries, August 2012, published by Japan Society of Civil Engineers, “Rock material particle size analysis system using digital camera images”, Kenichi Kawano et al. Proceedings of the 32nd National Conference of the Japanese Society for Artificial Intelligence, July 2018, published by the Japanese Society for Artificial Intelligence, “Method for determining soil properties by image analysis using deep learning”, Shinichi Homma et al.

In the soil property determination method based on image analysis using CNN in Non-Patent Document 2, properties such as fluidity of the excavated soil are determined from images of the excavated soil discharged through a belt conveyor, such as hard, soft, no target object, etc. , it is only possible to make a discrimination using the created trained model, and cannot estimate the particle size distribution of soil. On the other hand, CNN has the advantage of being able to automatically extract various features necessary for image discrimination. It is thought that if the system is trained to extract specific features of soil containing small particles whose contours are difficult to extract, it may be possible to estimate the particle size distribution.

An object of the present invention is to provide a method for estimating the particle size distribution of soil that can accurately estimate the particle size distribution of soil containing small particles using machine learning using artificial intelligence.

The present invention connects the surface of single-grained soil sieved from raw material soil or the surface of unsieved sampled soil whose particle size distribution is to be estimated, connected via a known wired or wireless communication network. The method is implemented in a particle size distribution estimation system that includes an imaging device that captures an image at a predetermined imaging distance, and a server that functions as a learning/estimation device that can perform machine learning, and each of the aforementioned units is classified into a plurality of particle sizes . The particle size distribution of the sampled soil is estimated by a trained model obtained by a machine learning algorithm using artificial intelligence built into the server , using image data using a large number of images of one particle size soil as training data. In the method for estimating particle size distribution, each image of the single-grained soil is obtained by capturing the surface of the single-grained soil, which is spread evenly to a thickness that does not create voids when viewed from above , using the imaging device. Image data using a large number of images of the single-grained soil, which is imaged and serves as training data for obtaining the learned model, includes additional images, The additional image is an image of the same size as the single particle size soil image by combining multiple divided images obtained by dividing one image, and each divided image is , any of the images of the single-grained soil are divided and extracted, and the proportion of each divided image of the single-grained soil in each additional image is As the proportion of single-grained soil of each particle size, these proportions are linked to the additional image containing the plurality of divided images of the plurality of single-grained soils, and when viewed in plan. An image taken by the photographing device at the predetermined photographing distance of the surface of the sampled soil whose particle size distribution is to be estimated , which is spread evenly to a thickness that does not create voids, is obtained by the server via the communication network. Provided is a soil particle size distribution estimation method for estimating the particle size distribution of sampled soil by inputting the sampled soil into the trained model and causing the server to output the particle size ratio of each class of sampled soil. Thus, the above objective was achieved.

Further, in the soil particle size distribution estimation method of the present invention, it is preferable that the divided images constituting the additional image are obtained by dividing one image into 4 equal parts or 16 equal parts.

Furthermore, in the soil particle size distribution estimation method of the present invention, it is preferable that the learned model is created using a convolutional neural network as a machine learning algorithm.

Furthermore, in the soil particle size distribution estimation method of the present invention, it is preferable that the trained model is created by applying transfer learning.

Further, in the soil particle size distribution estimation method of the present invention, the image of the single particle size soil or the image of the collected soil is the single particle size soil movably spread on a conveyor to a predetermined thickness. Alternatively, it is preferable that the surface of the collected soil be imaged by the imaging device disposed above the conveyor at a predetermined interval.

Furthermore, in the soil particle size distribution estimation method of the present invention, it is preferable that the raw material soil and the collected soil are locally generated soil.

According to the method for estimating the particle size distribution of soil of the present invention, the particle size distribution of soil containing small particles can be estimated with high accuracy by machine learning using artificial intelligence.

FIG. 1 is a schematic diagram illustrating the system configuration of a particle size distribution diagnosis system in which a soil particle size distribution estimation method according to a preferred embodiment of the present invention is implemented. FIG. 2 is an explanatory diagram of estimation processing by the soil particle size distribution estimation method according to a preferred embodiment of the present invention. FIG. 2 is an explanatory diagram of a process of creating a trained model by supervised learning. It is an explanatory diagram using images of characteristic changes due to stirring of mixed soil and single-grained soil. FIG. 2 is an explanatory diagram of a configuration example of a perceptron and calculation. FIG. 2 is an explanatory diagram of a configuration example and calculation of a convolutional neural network. FIG. 2 is an explanatory diagram of the configuration of the VGG 16 and the configuration of transfer learning. (a) to (g) are explanatory diagrams illustrating sample images of single-grained soil serving as training data, and (h) and (i) are explanatory diagrams illustrating additional images. It is a flowchart of the experimental flow of a demonstration experiment. (a) is a perspective view illustrating raw material soil made of silicate clay, (b) is a surface diagram illustrating mixed soil (sandy soil), and (c) is a surface diagram illustrating mixed soil (fine-grained soil). be. This is a chart showing the estimation results of Model 1 (left: sandy soil, right: fine-grained soil). This is a chart showing the estimation results of Model 2 (left: sandy soil, right: fine-grained soil). This is a chart showing the estimation results of Model 3 (left: sandy soil, right: fine-grained soil). This is a chart comparing the errors of each model.

A method for estimating the particle size distribution of soil according to a preferred embodiment of the present invention is to use an image of the particle size distribution of locally generated soil containing small particles collected at a construction site of civil engineering work, as the soil whose particle size distribution is to be estimated. By making it easy to estimate through analysis, the estimated particle size distribution of locally generated soil can be used as an important index when determining the mechanical properties and compaction characteristics of the local ground, such as the permeability of the local ground, This method was adopted as a method to improve productivity at construction sites and improve the quality of buildings and structures by knowing the liquefaction strength in advance.

Furthermore, according to conventional methods of measuring soil particle size distribution using sieve analysis and sedimentation analysis, it is difficult to frequently perform these analyzes on the large amounts of soil generated during the construction process at construction sites. For this reason, a method for estimating particle size distribution from images taken of the soil surface has been developed, but conventional methods for estimating particle size distribution using image analysis do not identify small particles whose contours cannot be captured from images. It is difficult to apply in the case of soil containing soil. The soil particle size distribution estimation method of this embodiment uses, for example, the particle size distribution estimation system 10 shown in FIG. It is possible to easily and accurately estimate the particle size distribution of soil, thereby solving the technical problems described above.

As shown in Figures 1 to 3, the method for estimating the particle size distribution of soil according to this embodiment is based on each single grain that has been sieved from the raw material soil (locally generated soil) and classified into multiple grain sizes. The trained model 20 (see FIG. 3) obtained by machine learning using artificial intelligence built into the server 11 uses a large number of images of the slope (see FIGS. 8(a) to (i)) as training data. This is a soil particle size distribution estimation method that estimates the soil particle size distribution. Images of each single-grained soil serving as training data are obtained by capturing the surface of the single-grained soil, which is leveled to a thickness T that does not create voids in plan view, at a predetermined imaging distance D using an imaging device 12. (see FIG. 1), and is also linked to the ratio of particle size for each class. An image taken by the photographing device 12 at a predetermined photographing distance D of the surface of collected soil (locally generated soil) whose particle size distribution is to be estimated, which is spread evenly with a thickness T that does not create voids when viewed from above, is trained. The particle size distribution of the sampled soil (see figure) is estimated by inputting the data into the model 20 and outputting the ratio of particle size of each class of sampled soil.

Further, in this embodiment, the trained model 20 obtained by the artificial intelligence built into the server 11 is preferably created using a convolutional neural network as a machine learning algorithm based on deep learning.

In this embodiment, the method for estimating the particle size distribution of soil can be implemented, for example, in the particle size distribution estimation system 10 shown in FIG. The particle size distribution estimation system 10 includes a server 11 that functions as a learning/estimation device capable of performing machine learning, and locally generated soil 13, which is made up of raw material soil with a single particle size obtained by sifting (single particle size soil) or unsifted soil. A belt conveyor 14 serves as a conveying device capable of transporting collected soil that is to be spread evenly to a predetermined thickness T, and soil generated on site is placed above the belt conveyor 14 and transported. It is configured to include an imaging device 12 that images the surface of soil 13 at a predetermined imaging distance D, and a locally generated soil storage container 15 disposed at the downstream end of a belt conveyor 14. The photographing device 12 and the server 11 are connected via a known wired or wireless communication network, and images of the locally generated soil 13 taken by the photographing device 12 can be sent to the server 11 as appropriate. It has become.

The server 11 constituting the particle size distribution estimation system 10 is composed of a computer, and is equipped with a known convolutional neural network capable of deep learning as a machine learning algorithm using artificial intelligence. It functions as a learning/estimation device that can perform Furthermore, the server 11 is equipped with a known image analysis program (installed therein), so that it can perform various types of image processing and image analysis.

Here, the computer constituting the server 11 includes a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), I/F (Interface), HDD (Hard Disk Drive), storage means, input means, display means, output means, etc. The CPU is configured to control image processing using an image analysis program and machine learning using artificial intelligence, while using the RAM as a work area, in accordance with various programs incorporated in the ROM. In addition, the CPU functions as a storage means, input means, display means, output means, etc. by having various computer programs incorporated in the ROM, and also functions as a storage means, an input means, a display means, an output means, etc., as well as data sent from the photographing device 12, analysis results by artificial intelligence, etc. can be stored in, for example, a database, displayed on a display, or output from a printer.

The photographing device 12 constituting the particle size distribution estimation system 10 can be any of various known digital cameras capable of photographing digital images. More specifically, as the photographing device 12, for example, the product name "CANON Power Shot S120" (manufactured by Canon Inc.) can be preferably used. The photographing device 12 is fixed above the belt conveyor 14 while maintaining a predetermined photographing distance D of, for example, about 200 to 2000 mm from the surface of the leveled locally generated soil 13. The surface of the locally generated soil 13 being conveyed can be photographed continuously at predetermined intervals, for example, within a photographing range of about 100 to 1000 mm in length and width. Each photographed image is sent to the server 11, which is a learning/estimation device, and is shaped into an image of, for example, 224 x 224 pixels and a size of 50 mm square (pixel resolution 0.22 mm) through image processing. The information is stored in the storage unit.

In this embodiment, a large number of images of each single-grained soil that has been sieved from the raw material soil (locally generated soil) and classified into multiple particle sizes, and images of these single-grained soils are provided. The trained model 20 is created by machine learning using artificial intelligence built into the server 11 using additional images (sample images, see FIGS. 8(a) to 8(i)) as training data. .

In this embodiment, each single particle diameter soil classified into a plurality of particle sizes is classified according to the particle size classification range of sieve analysis specified in JISA1204:2009 "Soil particle size test method" as shown in Table 1, for example. can be divided. In the sieve analysis specified by JIS, the particle size of soil particles less than 75 mm can be classified into a maximum of 13 categories. Soil particles smaller than 0.075 mm can be classified into more detailed categories by sedimentation analysis, but from the perspective of practical use, particle size estimation at the level of sedimentation analysis is not necessary. I can think.

In addition, soil properties include water content ratio, particle composition, etc., but these properties can be excluded. In this embodiment, locally generated soil in an air-dried state is used as the raw material soil, and this raw material soil is sieved to obtain single particle size soil of each classed particle size that becomes training data. It is now possible to estimate the particle size distribution by using the naturally dried raw material soil, which is unsieved locally sourced soil with mixed particle size soil, as the collected soil (locally generated soil) whose particle size distribution should be estimated. It has become.

Furthermore, in this embodiment, supervised learning, which is one of the machine learning methods, can be used to estimate the particle size distribution of soil. As shown in Figure 3, supervised learning performs machine learning based on a data set (training data) whose input and output are known, and creates a trained model (computation model) 20 that converts input values into output values. Can be automatically generated. Input data of a dataset (test data) with known input and output is input to the generated trained model 20, output an estimated value, and compare the estimated output data with the known output value (true value). By comparing , the estimation performance of the 20 learned models can be evaluated. This makes it possible to verify the trained model 20 and retrain it as necessary. In this embodiment, training data and test data can be considered as follows.

In order to appropriately generate the learned model (computation model) 20, the training data must contain accurate information. Figure 4 shows images of mixed soil made from locally generated soil and single-grained soil (soil composed only of grain size particles within a single grain size category) after three stirrings. . In the case of mixed soil, the composition of particles distributed on the surface changes each time it is stirred. Particles on the surface of mixed soil do not always have a composition that corresponds to the particle size distribution of the soil, and exhibit various characteristics. On the other hand, single-grained soil is a collection of only particles within a certain range of sizes, so it maintains its characteristic uniformity even after repeated agitation. From the above observations, it is thought that it is preferable to use soil with a single grain size, in which the correspondence between image features and grain sizes is clear, as the training data.

Furthermore, when estimating the particle size distribution of mixed soil in which particles of multiple particle sizes are mixed, it is not appropriate to estimate the entire particle size distribution of the mixed soil from an image of the surface of a single piece of mixed soil. do not have. Therefore, surface images of the mixed soil in multiple states after stirring are prepared as test data, and the estimation results of the particle size distribution of each test data are combined using an arithmetic average, and this is used as the estimation result of the particle size distribution of the entire soil. This is considered preferable. When handling soil at construction sites, the surface of the soil often changes constantly during transportation by conveyor belts and during the process of being deposited in storage containers, so the concept of estimating particle size distribution by integrating multiple images is used. is considered to be applicable to the soil conditions handled at these construction sites.

In addition, in this embodiment, a convolutional neural network capable of deep learning, which is known to exhibit high performance in the field of image classification, is used as a machine learning algorithm to create the trained model 20 by machine learning using artificial intelligence. A Convolutional Neural Network (hereinafter referred to as CNN) can be preferably used.

CNN is a technology that has developed a neural network, which is one of the machine learning algorithms, to enable advanced image recognition, and has the same basic structure as a neural network. Neural networks perform inference by connecting operators called perceptrons. An example of the configuration and calculation of a perceptron is, for example, as shown in Figure 5, the inner product of a one-dimensional array input value (Equation 1) ((1) feature amount) and a two-dimensional array (Equation 2) ((2) weight) ((3) Addition of inner product and bias). The obtained (Equation 4) is input to the transfer function f, and its output is passed to the next perceptron as a new feature amount (Equation 5) ((4) Transfer function). A structure in which these perceptrons are connected in parallel is called a layer, and a structure in which a plurality of layers are connected in series is called a multilayer perceptron (MLP). By transmitting the input value through a plurality of perceptrons, an estimated value y _PRED,C of class C is finally obtained. In model learning, a minimization algorithm is used to search for weights and biases so that the difference between estimated values and true values becomes smaller.

CNN is an application of the concept of convolution, which is one of the image processing techniques, to a neural network, and uses convolution coefficients as weights. An example of the CNN configuration and its calculations are shown in FIG. Most CNNs have three layers: a convolution layer, a pooling layer, and a fully connected layer. The convolution layer has a structure similar to a neural network, and uses input features (Equation 6) ((1) feature map) of a three-dimensional array that is image data (multiple two-dimensional arrays) and weights of a four-dimensional array ( Convolution integral is applied to Equation 7) ((2) kernel) ((3) convolution). The sum of all channels, the aggregated maps, and the addition of bias (Equation 8) (layer (4) is bias addition) is input to the transfer function (input to (5) transfer function), and as its output, the following The layer feature map formula 9) can be obtained ((6) Next layer feature map).

The effect of the kernel is to increase the number of channels in the feature map as the layer deepens. The pooling layer can extract only important features and reduce the map size ((7) After maximum pooling (Equation 10). By increasing the number of channels by the convolution layer and reducing the map by one pooling layer, A wide variety of features can be extracted from the entire image.The kernel acts as a feature filter, and by comprehensively translating the image using a convolution operation, it is possible to extract local features of the entire image using a small number of weights. The finally output feature map is converted (flattened) into a one-dimensional array and passed through MLP called a fully connected layer to obtain the output y _PRED,C . ((8) Fully connected layer (MLP) and output).

Various network structures have been proposed for CNN. FIG. 7 shows the network structure of VGG16, which is one of the typical CNNs. In the VGG16 network structure, for a 224 x 224 pixel, 3 channel image, there are 5 blocks consisting of 2 to 3 convolution layers containing multiple 3 x 3 minute pixel kernels and 1 pooling layer. Through this process, a feature map of 7×7 pixels and 512 channels can be generated. By connecting the 25088 flattened features to three fully connected layers, it is possible to design a design that outputs 1,000 classes. “VGG164” has been one of the widely used CNN models since it was announced in 2015 due to its high image classification performance and network simplicity.

The CNN applied in this embodiment is based on, for example, VGG16, and can output the ratio of particle sizes in all 7 categories shown in Table 2 out of a maximum of 13 particle sizes classified by sieve analysis specified in JIS. Therefore, it is preferable to create a new fully connected layer and connect it to the existing fully connected layer 2 (see FIG. 7, VGG16 fully connected layer and experimental fully connected layer). A softmax function shown in equation (11) can be used as the transfer function of the output layer.

The softmax function is a standard sigmoid function whose variable is the input x _C , translated in logD in the direction of the variable axis. The softmax function has two characteristics: an increasing function that is distributed in the domain (-∞, ∞) and the range (0, 1), and the feature that the sum of the outputs of all classes is 1. It can be used as a transfer function when estimating the probability distribution of each class. In this embodiment, the estimated value by the learned model (calculation model) 20 is the particle size distribution, that is, the mass percentage of particles for each particle size category (particle size) distributed between 0 and 100%, and the sum of these is 100%, it is preferable to apply a softmax function.

Preferably, the high performance of VGG16 is demonstrated by verification using a large amount of image data sets, and the learned weights used in the verification are made public under a commercial license. Convolutional layers trained for the purpose of general-purpose object classification are known to be applicable to various other image recognition problems, and the features extracted by the convolutional layers are partially used. This makes it possible to generate relatively high-performance models with a small amount of training data. This method is called transfer learning, and it is widely used in the field of artificial intelligence development, where the preparation of large amounts of training data and the enormous learning time are challenges. In this embodiment, the trained model 20 is preferably created by applying transfer learning.

In this embodiment, a large number of images of single-grained soil that serve as training data for obtaining the learned model 20 are configured to include additional images obtained using these images of single-grained soil. It is preferable that As shown in Figures 8 (h) and (i), the additional image is created by combining multiple divided images obtained by dividing a single image into an image of the same size as the image of single-grain soil. Each divided image is obtained by dividing and extracting an image of one of the single-grained soils. In addition, the proportion of each divided image of single-grained soil in each additional image is the proportion of single-grained soil of each particle size in the additional image, and these proportions are It is now linked to additional images that include divided images of the road.

For example, as shown in Figures 8(a) to (g), images of soil with only a single particle size are associated with particle size distributions, with the proportion of each applicable particle size classified as 100%. However, when using only these images of single-grained soil and the 100% particle size distribution (proportion of single-grained soil) linked to this as training data, VGG16 can There is a possibility that unity itself may be learned as a feature, and there is a concern that it will not be able to handle the estimation of mixed soils composed of particles of various particle sizes.

For this reason, in this embodiment, in order to improve the performance of the learned model 20 with limited training data, a widely used method that can increase the amount of data by applying various transformations to the training data is used. It is preferable to adopt a method called inflating. In this embodiment, as a method of inflating the training data, in order to learn the characteristics of images containing particles of a plurality of particle sizes, it is preferable to use an image of a single particle size soil in each classification category. An additional image is created by dividing the image into a plurality of images and randomly connecting and combining these divided images. The created extra images are illustrated in FIGS. 8(h) and (i). The divided images constituting the additional image can be, for example, as shown in FIGS. 8(h) and (i), a single image of single-grained soil divided into 4 or 16 equal parts. . The particle size distribution associated with each additional image is defined as the proportion of each divided image of single-grained soil in each additional image, and these proportions are calculated as the proportion of each divided image of single-grained soil in each additional image. It is now linked to premium images that include .

Specifically, the images of only single-grained soil shown in Figures 8(a) to (g) are mapped with the ratio of each classified grain size as 100%, whereas, for example, The additional image shown in Figure 8(h) is an image obtained by dividing a single image into four equal parts. ), one image divided into four of single particle size soil class 4 in Figure 8(e), one image divided into four of single grain size soil class 5 in Figure 8(e), and one image divided into four of single grain size soil class 5 in Figure 8(g). The image is a combination of the four divided images of single particle size soil classification 7, so the additional image in Figure 8 (h) is for particles with a particle size of less than 0.075 mm. The proportion of particles is 25%, the proportion of particles with a particle size of 0.25 to 0.425 mm is 25%, the proportion of particles with a particle size of 0.425 to 0.6 mm is 25%, the particle size is 2 to 4 The proportions of these particle sizes are calculated assuming that the proportion of particles with a particle size of .75 mm is 25%.

In addition, for example, in the additional image obtained by dividing one image into 16 equal parts as shown in FIG. 8(i), two images of the single particle size soil classification 1 shown in FIG. 8(a) are divided into 16, Two images of single particle size soil class 3 divided into 16 in Figure 8(c), three images divided into 16 of single grain size soil class 4 in Figure 8(d), and Figure 8 Two 16-divided images of single-grained soil class 5 in (e), four 16-divided images of single-grained soil class 6 in Figure 8(f), and Figure 8(g) ), the image is a combination of three 16-divided images of single particle size soil classification 7. The proportion of particles with a particle size of 12.5%, the proportion of particles with a particle size of 0.075 to 0.106 mm are 0%, and the proportion of particles with a particle size of 0.106 to 0.25 mm is 12.5%. , the proportion of particles with a particle size of 0.25 to 0.425 mm is 18.8%, the proportion of particles with a particle size of 0.425 to 0.6 mm is 12.5%, the particle size is 0.6 to 2 mm The proportions of these particle sizes are set as 25.0%, and 18.8% of particles with a particle size of 2 to 4.75 mm.

In this embodiment, images of each single-grained soil serving as training data cover the surface of the single-grained soil, which is leveled to a thickness T that does not create voids when viewed from above, at a predetermined level. The image was taken by the imaging device 12 at the imaging distance D, and the image of the collected soil (locally generated soil) whose particle size distribution should be estimated was similarly laid out with a thickness T that does not create voids when viewed from above. The surface of the sampled soil (locally generated soil) is imaged by the imaging device 12 at a predetermined imaging distance D. By inputting the captured image of the collected soil (locally generated soil) into the trained model 20 generated as described above and outputting the particle size ratio of each class of the collected soil, The particle size distribution of soil (see Figures 11 to 13) is estimated.

When taking images of the surface of each single-grained soil that will serve as training data or images of the surface of sampled soil whose particle size distribution is to be estimated, these single-grained soils and sampled soil are viewed in plan. By spreading the soil to a thickness T that does not create voids, it is possible to avoid the background of single-grained soil or sampled soil from being captured in the photographed image, and to create the soil particle size distribution. The learned model 20 enables accurate estimation.

According to the soil particle size distribution estimation method of this embodiment having the above-described configuration, it becomes possible to accurately estimate the soil particle size distribution containing small particles by machine learning using artificial intelligence.

That is, according to the soil particle size distribution estimation method of the present embodiment, a large number of images ( Artificial intelligence using a convolutional neural network (hereinafter referred to as CNN), which is preferably a type of deep learning, incorporated in the server 11, using training data (see FIGS. 8(a) to (i)) as training data. The trained model 20 (see Figure 3) obtained through machine learning is used to estimate the particle size distribution of the sampled soil, and the particle size distribution can be estimated by photographing the surface of the sampled soil using, for example, locally generated soil. The particle size distribution of the soil is estimated by inputting the image captured by the device 12 into the learned model 20 and outputting the particle size ratio of each class of the sampled soil.

As a result, according to this embodiment, for example, even for a large amount of soil generated during the construction process at a construction site, it is possible to estimate properties such as soil compaction characteristics, water permeability, and liquefaction strength. The particle size distribution of soil, which is one of the indicators, can be measured without using methods such as sieve analysis or sedimentation analysis, which are difficult to perform frequently. At the same time, it is preferable to input an image taken by a photographing device of the surface of the sampled soil whose particle size distribution is to be estimated into a trained model obtained by training CNN as a machine learning algorithm. It is possible to estimate the particle size distribution of soil with high accuracy using a simple method that requires only a few steps. Due to this, the soil particle size distribution estimation method of this embodiment can be easily implemented frequently.

Note that the present invention is not limited to the above-described embodiments, and various changes can be made. For example, the divided images that make up the additional image do not necessarily have to be a single image divided into 4 or 16 equal parts, and even if they are divided by a number of divisions other than these numbers. good. Images of single-grained soil and collected soil do not necessarily have to be taken by spreading them movably on a conveyor, but rather images of single-grained soil or collected soil are taken using other methods. It is also possible to take these images by laying it out to a thickness that does not create any voids when viewed from above.

EXAMPLES Hereinafter, the method for estimating the soil particle size distribution of the present invention will be explained in more detail with reference to Examples, but the present invention is not limited in any way by the description of the following Examples.

〔Example〕
According to the experimental flow shown in FIG. 9, a demonstration experiment of the soil particle size distribution estimation method of the present invention was conducted. In a demonstration experiment based on the experimental flow shown in Figure 9, raw material soil is selected and subjected to sieve analysis to create multiple types of single particle size soil. After creating a dataset of images of single-grained soil and their associated particle size ratios as training data, the single-grained soils were mixed to form a mixed soil, and the particle size distribution was created as test data. Create a test dataset of two different types of soil images and known particle size ratios. Next, the trained model is generated by learning the training data. When generating trained models, multiple trained models are generated for each learning method. Multiple images of mixed soil for testing are input to each of the trained models generated by learning the training data, each outputs an estimated value of the particle size distribution, and the average value of these estimated values is calculated. obtain. The estimated value of the particle size distribution obtained by inputting multiple images of the test soil mixture is compared with the particle size distribution based on the known particle size ratio of the test data, and the error is calculated for each trained model. Comprehensive evaluation of performance. The contents of the demonstration experiment according to the experiment flow shown in FIG. 9 will be described below.

In the demonstration experiment, we selected Soft Silica's purchased soil ``silicate clay'' as the raw material soil, and analyzed it by sieving it by opening. A photograph of the raw material soil made of silicate clay is shown in FIG. 10(a), and the particle size distribution of the raw material soil is shown in Table 2. The raw material soil was composed of soil particles smaller than 4.75 mm, and these were classified into a total of seven particle sizes.

The seven classifications of single-grain size soil obtained by sieve analysis were placed in a container, and images were taken indoors with a camera (CANON Power Shot S120) fixed above. (pixel resolution 0.22 mm). Image data was obtained for 150 samples in each of the 7 categories, for a total of 1,050 samples (see FIGS. 8(a) to (g)). In the images of single-grain size soil from Class 1 to Class 5, it is difficult to visually distinguish the outline of the grain size.

In the demonstration experiment, in order to improve the performance of the trained model created with limited training data, we inflated the training data with the aim of improving estimation performance. As mentioned above, in order to learn the characteristics of mixed soil containing particles of multiple particle sizes, the image of single-grained soil in each category is divided, and these divided images are randomly connected. An image (additional image) was created ((see FIGS. 8(h) and (i))). The number of divisions and combinations was set to two patterns, 4 divisions and 16 divisions, and the particle size distribution of the combined image (additional image) was calculated based on the area ratio of the divided images of each division included in the image. For example, if four of the 16 divided images shown in FIG. 8(i) include images in category 3 in FIG. The mass percentage of one-grain size soil is 25%. We decided to create such a data set, 1500 samples for each division pattern of 4 divisions and 16 divisions, for a total of 3000 samples, and add them to the training data.

Subsequently, the sieved single-grained soil was reconstituted to create a mixed soil. The particle size distribution of the mixed soil is shown in Table 2, and the photographed image samples of the mixed soil are shown in FIGS. 10(b) and (c). When preparing the mixed soil, it was made to contain at least 5% of soil of all seven particle size categories. In addition, as an engineering classification of soil, it is roughly divided into two types: fine-grained soil, which is composed of 50% or more of fine grains, and sandy soil, which is composed of less than 50% of fine grains3). Using this definition as a reference, we created two types of mixed soil for the experiment: sandy soil and fine-grained soil. The sandy soil had a fine grain content of 5%, and the fine grain soil had a fine grain content of 55%. A total of 60 images, 30 images of each of the mixed soils created, were taken in the same manner as for the single-grained soil.

CNN, which exhibits high image recognition performance, was used as the machine learning algorithm for creating the trained model. The basic concept of CNN and the details of CNN applied to the demonstration experiment are as described above.

In addition, in the demonstration experiment, in order to confirm the effectiveness of transfer learning, learning was performed using the following two patterns.
1. When all the kernel weights in Figure 7 are fixed (frozen) using learned weights, and only the weights of the fully connected layer are learned using the training data of the demonstration experiment.
2. In the case of full-scratch learning using only the training data of this experiment, including the kernel weights.
In addition, in order to confirm that the extra image using the split and combined image that was padded for learning could contribute to improving the estimation performance of the trained model 20, learning was also performed without using the padded extra image.

In order to conduct the above study, the three patterns shown in Table 3 were trained, and the performance of each of the obtained trained models was compared.

To estimate the particle size distribution of sandy soil mixed soil and fine-grained soil mixed soil with the particle size distribution shown in Table 2, as described above, the arithmetic mean value of the estimated value y _PRED,C of all 30 images is used. It is calculated using (Equation 12).

The estimation performance of each trained model is determined by the root mean square error of the arithmetic mean value of the estimated value y _PRED,C with respect to the true value y _TRUE of the particle size distribution of the mixed soil with sandy soil and the mixed soil with fine grain soil. (Root Mean Square Error, hereinafter referred to as RMSE) and calculated using (Equation 13).

Figures 11 to 13 show curves showing the estimation results of the particle size ratios of seven categories of sandy soil mixed soil and fine grained soil mixed soil of Model 1, Model 2, and Model 3 in Table 3. The upper graph of each figure is the ratio of particles within a particle size category (particle size) plotted at the upper limit of the category, and the lower graph is the cumulative particle size curve (the plotted particle size). (percentage of particles smaller than the diameter) is shown. As a reference for discussion, the estimation results for each of the 30 samples are also shown.

In the estimation results of Model 1, the distribution range of estimated values for each sample is large, and there are many curves with significantly high mass fractions. Many images of mixed soil contain particles of various sizes, so this result is thought to be the result of learning the unique distribution of single-grain size soil in the training data. In addition, there are similar estimated curves for sandy soil mixed soil and fine grained soil mixed soil, and it is not possible to classify the characteristics of sandy soil mixed soil and fine grained soil mixed soil using only the estimated curves for each sample. do not have. Furthermore, the estimated curves of Models 2 and 3 have narrower distribution ranges than the estimated curves of Model 1. Furthermore, the estimated curves for the fine-grained soil mixture and the sandy soil mixture do not overlap, and it is possible to distinguish the characteristics of the two types of soil just by looking at the estimated curves for each sample. be.

In any of Models 1 to 3, when all the 30 samples are averaged, a trend close to the actual particle size distribution is obtained. Regarding mixed soil with fine grains, the proportion of fine grains is overestimated in all models. The cause of this is thought to be the adhesion of fine particles. Generally, as the particles become finer, the adhesion force (intermolecular force) exceeds gravity, making it easier for the particles to adhere to surrounding objects.

FIG. 14 shows the comparison results of errors by each model. It is found that the errors become smaller in the order of Model 1, Model 3, and Model 2. Considering the estimated curves of each sample, inflating the training data by combining images used in Models 2 and 3 is considered to be an effective method. In addition, since model 2 had the least error, it is clear that there was also an effect of transfer learning.

10 Particle size distribution estimation system 11 Server 12 Photography device 13 Locally generated soil 14 Belt conveyor (transportation device)
15 Locally generated soil storage container 20 Learned model T Predetermined thickness D that does not create any voids when viewed from above Shooting distance

Claims (6)

Connected via a known wired or wireless communication network, the surface of single-grained soil sieved from raw material soil, or the surface of unsieved sampled soil whose particle size distribution is to be estimated, is captured at a predetermined imaging distance. The particle size distribution estimation system includes a photographing device that captures an image with Estimating the particle size distribution of the soil by estimating the particle size distribution of the sampled soil using a trained model obtained by a machine learning algorithm using artificial intelligence built into the server , using image data using a large number of images as training data. A method,
Each image of the single-grained soil is obtained by capturing an image of the surface of the single-grained soil, which is spread evenly to a thickness that does not create voids when viewed from above , using the imaging device. , image data using a large number of images of the single-grained soil, which is training data for obtaining the learned model, includes additional images,
The additional image is an image of the same size as the single particle size soil image by combining multiple divided images obtained by dividing one image, and each divided image is , any of the images of the single-grained soil are divided and extracted, and the proportion of each divided image of the single-grained soil in each additional image is As the proportion of single-grained soil of each grain size, these proportions are linked to the additional image including the divided images of the plurality of single-grained soils,
An image taken by the photographing device at the predetermined photographing distance of the surface of the sampled soil whose particle size distribution is to be estimated, which is spread evenly to a thickness that does not create any voids when viewed from above, is sent to the server via the communication network. A soil particle size distribution estimation method for estimating the particle size distribution of the sampled soil by inputting the learned model obtained in the above and having the server output the particle size ratio of each class of the sampled soil. 2. The soil particle size distribution estimation method according to claim 1 , wherein the divided images constituting the additional image are obtained by dividing one image into four equal parts or sixteen equal parts. 3. The soil particle size distribution estimation method according to claim 1, wherein the learned model is created using a convolutional neural network as a machine learning algorithm. 4. The soil particle size distribution estimation method according to claim 3 , wherein the trained model is created by applying transfer learning. The image of the single-grained soil or the image of the collected soil is such that the surface of the single-grained soil or the collected soil, which is movably spread to a predetermined thickness on a conveyor, is placed above the conveyor in a predetermined position. The soil particle size distribution estimation method according to any one of claims 1 to 4, wherein the soil particle size distribution estimation method is obtained by imaging with the imaging devices arranged at intervals of . The soil particle size distribution estimation method according to any one of claims 1 to 5, wherein the raw material soil and the collected soil are locally generated soil.

Images