Construction method of grain size distribution measurement learning model of soil particle

Abstract

To achieve expansion of a feature amount of soil particles and efficient construction of a learning model by executing mathematical processing to small scaled data.SOLUTION: The invention is configured so that learning input data for constructing a learning model for correcting an error of a particle size distribution of image analysis comprises: a particle size distribution A of image analysis of classified soil, in addition to a particle size of screen analysis and particle size of image analysis; a particle size ratio distribution B indicating a particle size ratio of image analysis of disposed soil with respect to a particle size of screen analysis; a particle size ratio distribution C for indicating a particle size ratio of the image analysis of the disposed soil to the particle size of the image analysis; a DFT conversion value F obtained by performing Fourier transformation to the particle size of the image analysis of the disposed soil for replacing the particle size with intensity and a phase of a frequency component; DFT evaluation values G, H indicating a ratio of the DFT conversion value F to a DFT conversion value obtained from the particle size of screen analysis or the particle size of the image analysis; and component coefficients D, E indicating a component ratio of combination of an element particle size distribution on the basis of a relation between an element particle size distribution obtained from the image analysis and screen analysis from the classified soil in each rank where original soil is screened, and the DFT conversion value determined from the particle size of the screen analysis or particle size of the image analysis.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present invention relates to a method for constructing a learning model for measuring the particle size distribution of soil particles in order to evaluate the characteristics and properties of soil according to particle size in the treatment of fine particle removal as one of the volume reduction methods for soil containing radioactive materials, and in the treatment of soil contaminated with harmful substances such as heavy metals. [Background technology]

[0002] Generally, to evaluate soil properties, particle size distribution, which expresses the content of each particle size of soil particles as a mass percentage, is measured in accordance with JISA1204. However, since the target soil must be sieved and weighed each time, the measurement is a time-consuming and labor-intensive process. In particular, at contaminated soil treatment sites, rapid measurement of particle size distribution is desirable to evaluate the properties of the target soil and the results of treatment such as the removal of specific particle sizes. Conventionally, Patent Document 1 discloses a particle size distribution measuring device that irradiates a dispersed particle group to be measured with laser light, measures the spatial intensity distribution of diffracted and scattered light, and converts the measurement result of the spatial intensity distribution into the particle size distribution of the particle group to be measured using a coefficient matrix that takes into account the refractive index of the particle group to be measured. However, when evaluating soil, which is a compound of various particle sizes and shapes, there are problems with measurement accuracy due to the non-uniformity of particle size specific to soil particles and the intensity of the laser light. When the soil particles are non-spherical particles, there is a tendency to evaluate them toward fine particles, which may deviate from the actual situation. Patent Document 2 discloses a method of irradiating a sample with a laser, acquiring scattered light in a time series with a detector, and confirming and correcting the presence or absence of large particles from the detection results. However, the measurement targets are fine particles of 0.3 nm to 20 nm such as proteins, and particles exceeding 100 nm among the measurement targets are treated as large particles. Therefore, this method is difficult to apply to the measurement of soil that contains particles larger than 0.1 μm and has a wide particle size range. In contrast, Patent Document 3 performs pretreatment to disperse soil particles while leveling them by passing them through a slit-shaped sieve, but the target that can be evaluated is limited to gravel and sand among the soil components, and even if overlapping of soil particles is prevented, there is a limit to improving the accuracy of the particle size distribution analysis results by image analysis when finer silt and clay are included. Furthermore, Patent Document 4 discloses a measurement method in which analysis result data is calculated by machine learning for image data of soil particle groups acquired by image analysis. This measurement method shortens the time required to identify soil characteristics such as particle size distribution, but due to the characteristics of machine learning, it is necessary to prepare a large amount of analysis result data for learning and training in order to build a learning model. In addition, padding and adding features are given as ways to increase the analysis result data, but since it is difficult to distinguish the characteristics from the appearance of soil and soil particles, padding by simple processing of image data or extraction of features by processing analysis result data including image data is not effective. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2010-101653 A [Patent Document 2] Patent Publication No. 2021-183983 [Patent Document 3] JP 2019-113429 A [Patent Document 4] Patent Publication No. 2021-067468 Summary of the Invention [Problem to be solved by the invention]

[0004] Image analysis can shorten the time and reduce the labor required to identify the particle size distribution of soil particles, but it contains errors compared to actual measurements of particle size distribution, so it is necessary to use machine learning, including deep learning, to correct the errors in the image analysis method. In order to build a learning model for machine learning, a large amount of analysis result data for learning and training must be collected, and the time and labor required for preparations at the soil treatment site, from taking samples of the soil to be treated to processing the data, becomes excessive. Therefore, an objective of the present invention is to provide a method for constructing a learning model for measuring the particle size distribution of soil particles, which efficiently expands the features of soil particles and optimizes the learning model by applying mathematical processing to small-scale learning and training data. [Means for solving the problem]

[0005] In the present invention, a method is adopted in which a large number of pieces of input data are input, including an image analysis particle size distribution obtained by image analysis of soil particles in image data of photographed soil used for learning, a sieving particle size distribution obtained by sieving the soil used for learning, and distribution data linking the image analysis particle size with the sieving particle size, and machine learning is performed to construct a learning model for measuring particle size distribution of soil particles by image analysis in order to correct errors contained in the image analysis particle size distribution. The learning input data for the learning model are: particle size distribution A of raw soil obtained by sieving raw soil collected from soil into arbitrary classes; particle size distribution A of raw soil obtained by photographing the same raw soil and analyzing the image of the soil particles into the same classes; particle size distribution B of treated soil obtained by image analysis of treated soil classified into a predetermined particle size range; particle size distribution C of treated soil obtained by image analysis of treated soil classified into a predetermined particle size range; particle size distribution D of treated soil obtained by image analysis of treated soil to particle size distribution C of raw soil; The DFT calculation includes a DFT transformation value F which uses a Fourier transform to convert the particle size determined by image analysis of the soil into the intensity and phase of frequency components; DFT evaluation values ​​G and H which indicate the ratio of the DFT transformation value F to the DFT transformation value obtained from the particle size determined by sieving the raw soil or the particle size determined by image analysis of the raw soil; and composition coefficients D and E which indicate the composition ratio of the combination of element particle size distributions from the relationship between the element particle size distribution obtained by sieving and image analysis of classified soil of single particle size in each class obtained by sieving the raw soil into any class and the DFT transformation value obtained from the particle size determined by sieving the raw soil or the particle size determined by image analysis of the raw soil. Effect of the Invention

[0006] By applying mathematical processing to small-scale learning and training data, the present invention can expand the features of soil particles and properly and efficiently construct a learning model, thereby reducing errors in the particle-size distribution of soil particles obtained by image analysis methods. In particular, it has the effect of shortening the time required to identify the particle-size distribution of soil particles by image analysis methods and reducing the workload. [Brief description of the drawings]

[0007] [Figure 1]FIG. 1 is a schematic diagram of a measurement system for measuring the particle size distribution of soil particles. [Diagram 2] FIG. 2 is an explanatory diagram showing an overview of data processing using a learning model for measuring the particle size distribution of soil particles according to the present invention. [Diagram 3] FIG. 2 is an explanatory diagram of a trained model for measuring the particle size distribution of soil particles according to the present invention. [Figure 4] FIG. 1 is a flow diagram of the pre-processing process of input data for a learning model. [Diagram 5] FIG. 11 is a flow diagram of a process for acquiring configuration coefficients of input data. [Figure 6] FIG. 1 is a flow diagram of the process of constructing a trained model. [Figure 7] FIG. 1 is a schematic explanatory diagram of a learning model. [Figure 8] 1 is a graph showing a comparison of the particle size distribution measurement results of three types of samples A, B, and C by sieving and image analysis. [Figure 9a] 1 is a graph and a table showing a comparison of the element particle size distribution measurement results for sample A obtained by sieving with sieve openings of 850 to 2000 μm and by image analysis. [Figure 9b] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample A by sieving with sieve openings of 425 to 850 μm and by image analysis. [Figure 9c] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample A by sieving with sieve openings of 212 to 425 μm and by image analysis. [Figure 9d] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample A by sieving with sieve openings of 106 to 212 μm and by image analysis. [Figure 9e] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample A by sieving with 0 to 106 μm mesh and by image analysis. [Figure 10a] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample B by sieving with sieve openings of 850 to 2000 μm and by image analysis. [Figure 10b]1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample B by sieving with sieve openings of 425 to 850 μm and by image analysis. [Figure 10c] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample B by sieving with sieve openings of 212 to 425 μm and by image analysis. [Figure 10d] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample B by sieving with sieve openings of 106 to 212 μm and by image analysis. [Figure 10e] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample B by sieving with 0 to 106 μm mesh and by image analysis. [Figure 11a] 1 is a graph and a table showing a comparison of the element particle size distribution measurement results for sample C obtained by sieving with sieve openings of 850 to 2000 μm and by image analysis. [Figure 11b] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample C by sieving with sieve openings of 425 to 850 μm and by image analysis. [Figure 11c] 1 is a graph and a table showing a comparison of the element particle size distribution measurement results for sample C obtained by sieving with sieve openings of 212 to 425 μm and by image analysis. [Figure 11d] 1 is a graph and a table showing a comparison of the measurement results of element particle size distribution for sample C by sieving with sieve openings of 106 to 212 μm and by image analysis. [Figure 11e] 1 is a graph and a table showing a comparison of the element particle size distribution measurement results for sample C obtained by sieving with 0 to 106 μm mesh size and by image analysis. [Figure 12a] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving raw soil sample A with sieve meshes of 850 to 2000 μm and by image analysis. [Figure 12b] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving raw soil sample A with sieve meshes of 425 to 850 μm and by image analysis. [Figure 12c]1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving sample A with sieve openings of 212 to 425 μm and by image analysis. [Figure 12d] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 106 to 212 μm and by image analysis for sample A. [Figure 12e] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with 0 to 106 μm mesh size and by image analysis for sample A. [Figure 13a] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 850 to 2000 μm and by image analysis for sample B. [Figure 13b] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 425 to 850 μm and by image analysis for sample B. [Figure 13c] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 212 to 425 μm and by image analysis for sample B. [Figure 13d] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 106 to 212 μm and by image analysis for sample B. [Figure 13e] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with 0 to 106 μm mesh size and by image analysis for sample B. [Figure 14a] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 850 to 2000 μm and image analysis for sample C. [Figure 14b] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 425 to 850 μm and by image analysis for sample C. [Figure 14c] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 212 to 425 μm and by image analysis for sample C. [Figure 14d]1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with sieve openings of 106 to 212 μm and by image analysis for sample C. [Figure 14e] 1 is a graph and a table showing a comparison of DFT transformation values ​​of element particle size distribution obtained by sieving with 0 to 106 μm mesh size and by image analysis for sample C. [Figure 15a] 1 is a photograph for image analysis showing an example of image data of large grain size of raw soil. [Figure 15b] 1 is a photograph for image analysis showing an example of image data of medium-sized grain size of raw soil. [Figure 15c] 1 is a photograph for image analysis showing an example of image data of small grain size of raw soil. [Figure 16a] 1 is a table showing particle size distribution of sample A by image analysis. [Figure 16b] 1 is a table showing the particle size ratio distribution of sample A by sieving. [Figure 16c] 1 is a table showing particle size ratio distribution by image analysis of sample A. [Figure 16d] 1 is a table showing DFT transformation values ​​of particle size ratio distribution by image analysis of sample A. [Figure 16e] 1 is a table showing DFT evaluation values ​​of particle size ratio distribution by sieving of sample A. [Fig. 16f] 1 is a table showing DFT evaluation values ​​of particle size ratio distribution by image analysis of sample A. [Figure 16g] 1 is a table showing the composition coefficients of the particle size ratio distribution by sieving sample A. [Figure 16h] 1 is a table showing the composition coefficients of the particle size ratio distribution by image analysis of sample A. [Figure 17] 1 is a table showing the configuration of a connection layer of a learning model. [Figure 18] 1 is a table showing the configuration of a learning model network. [Figure 19] 1 is a graph showing a comparison of particle size distributions obtained by sieve measurement and image analysis using a learning model before learning. [Figure 20a] 1 is a graph showing a history of the mean square error of particle size distribution by image analysis versus the number of learning times of the learning model of Example 1. [Figure 20b] 1 is a graph showing the transition of the mean square error of particle size distribution by image analysis versus the number of learning times of the learning model of Example 2. [Figure 20c] 13 is a graph showing the change in mean square error of particle size distribution by image analysis versus the number of learning times of the learning model of Example 3. [Figure 20d] 1 is a graph showing the transition of the mean square error of particle size distribution by image analysis versus the number of learning times of the learning model of Example 4. [Figure 21a] 1 is a graph showing a comparison of particle size distributions obtained by sieving and image analysis when the learning model after learning in Example 1 is used. [Figure 21b] 13 is a graph showing a comparison of particle size distributions obtained by sieving and image analysis when the learning model after learning in Example 2 is used. [Figure 21c] 13 is a graph showing a comparison of particle size distributions obtained by sieving and image analysis when using the learning model after learning in Example 3. [Figure 21d] 13 is a graph showing a comparison of particle size distributions obtained by sieving and image analysis when the learning model after learning in Example 4 is used. [Figure 22] 1 is a table showing the configuration of a network of a learning model in which convolutional layers are parallelized. [Figure 23] FIG. 1 is a schematic explanatory diagram showing the network structure of a learning model in which convolutional layers are parallelized. [Figure 24] 13 is a graph showing the transition of the mean square error of the particle size distribution with respect to the number of times of learning of a learning model in which a convolution layer is parallelized. [Diagram 25] 13 is a graph showing a comparison of particle size distributions obtained by sieve measurement and image analysis using a learning model after learning in which convolutional layers are parallelized. [Figure 26] 13 is a graph showing the transition of the mean square error of the granularity distribution with respect to the number of learning iterations of a learning model in which the convolutional layer is parallelized and early truncated. [Figure 27] 13 is a graph showing a comparison of particle size distributions obtained by sieve measurement and image analysis using a trained model with parallelized convolutional layers and early truncation. [Figure 28]13 is a graph showing a comparison of particle size distribution in a case where the error reduction rate is low in the training data and particle size distribution by sieving, as a result of image analysis using a learning model. [Figure 29] 13 is a graph showing a comparison of particle size distribution in cases where the error reduction rate is low in test data and particle size distribution by sieving, as a result of image analysis using a learning model. [Diagram 30] FIG. 1 is a schematic diagram illustrating a structure of a neural network. [Diagram 31] 1 is a table showing an example of input data settings for a neural network. [Diagram 32] 1 is a table showing the configuration of a learning model network. [Diagram 33] 1 is a graph showing the transition of the mean square error of particle size distribution by image analysis versus the number of times of learning of the learning model. [Diagram 34] 1 is a graph showing a comparison of particle size distribution by image analysis using a learning model and particle size distribution by sieving. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] An embodiment of the present invention will be described with reference to the drawings. As shown in FIG. 1, a measurement system for carrying out the particle-size distribution measurement method using the learning model of the present invention includes an imaging device 1 that captures image data by photographing raw soil S that has been collected from the soil to be measured and leveled on a flat pallet P, and an arithmetic device 2 consisting of a computer that operates the imaging device 1 and measures the particle-size distribution from the image data.

[0009] The calculation device 2 includes an image analysis function unit 2a that creates particle size distribution data B1, B2, ... Bn from image data A1, A2, ... An of photographed soil, as shown in Fig. 2. Separately, the particle size distribution results based on mass fractions obtained by sieving such as JIS A 1204 are taken as actual measured values. An error is included between the particle size distribution data, which is the measured value output by image analysis using the image analysis function unit 2a, and the actual measured value, as a difference due to the limits of detection accuracy, etc.

[0010] The calculation device 2 further includes a pre-processing function unit and a learning model function unit 2b that outputs particle size distribution result data C1, C2, ... Cn as a result of correcting and amending the particle size distribution data B1, B2, ... Bn, which includes errors with respect to the actual measured values. Machine learning by the learning model function unit 2b is performed by a neural network (NN), a convolutional neural network (CNN), or the like.

[0011] In constructing the learning model according to this embodiment, there is a limit to the improvement in performance of the learning model even when applying not only ordinary machine learning but also deep learning, which is a NN, because the particle size distribution using only the particle size distribution measured by image analysis contains few features and it is difficult to grasp trends. Therefore, the input data to be input to the learning model is subjected to mathematical pre-processing by the pre-processing function unit / learning model function unit 2b built into the system, thereby expanding and optimizing the features. This processing makes it possible to efficiently build a learning model to be built into a machine learning system using NN. Furthermore, by capturing the waveform shape of the particle size distribution as image data, deep learning using CNN is applied to build a learning model that outputs a corrected particle size distribution with higher accuracy than NN.

[0012] In constructing the learning model, we used data that contained errors, which was the particle size distribution measured by image analysis from soil image data, and data that was a pair of what was thought to be the correct particle size distribution measured by sieving the same soil. As shown in Figure 3, data in which the features of the particle size distribution measured by image analysis had been expanded in advance by mathematical preprocessing was input to the system. In addition, all the data was divided into training data with teacher data, in which the actual measured values ​​from sieving were used as teacher data, and validation data, in which the actual measured values ​​were treated as unknown data. Training data is input and corrected through a system incorporating a learning model, and learning is repeated to reduce the error between the corrected output value and the teacher data, building a learning model. Furthermore, validation data is used to verify the generalization ability of the learning model as its ability to handle unknown data. If the validation results in a large error between the output corrected value and the validation data, accuracy is improved by repeating learning while being careful not to overfit.

[0013] The input data to be input to the learning model is expanded by pre-processing, and as shown in Figure 4, particle size distribution data A obtained by image analysis, particle size ratio distribution B obtained by sieving, particle size ratio distribution C obtained by image analysis, DFT transformation value F of the particle size distribution obtained by image analysis, DFT evaluation values ​​G and H, and composition coefficients D and E are used.

[0014] In order to carry out pretreatment, a calibration process must be carried out on the soil to be evaluated. As the first calibration process, the raw soil, which is untreated soil before classification processes such as sieving, is subjected to measurement of the particle size distribution measured by sieving in a given particle size range. In addition, the same soil is subjected to measurement of the particle size distribution of the same number of classes as the sieving measured by image analysis.

[0015] The next step in the calibration process is the process of acquiring calibration data, which is shown in Figure 5. The raw soil is classified into a given particle size range by sieving (S1), and image data is acquired for each class of the single-particle-size classified soil by photographing it with a camera or the like (S2). At the same time, the single-particle-size classified soil for each class is sieved again with the same number of classes (S3). From the image data, the particle-size distribution, which is the measured value by image analysis, and the particle-size distribution, which is the actual measured value by sieving, are treated as a pair of data to construct analysis result data (S4). By performing mathematical processing on the constructed analysis result data, the composition coefficients obtained by image analysis and composition coefficients obtained by sieving are obtained from the single-particle-size classified soil as calibration data (S5).

[0016] Note that the particle size distribution measured by image analysis of image data of raw soil and classified soil varies greatly depending on the threshold value set for the contours and boundaries of soil particles. For this reason, particle size distribution data may be obtained at different threshold values, and the particle size distribution may be corrected taking into account the variance in the particle size distribution data. First, the difference between the calculated particle size distribution's D50 average particle size and standard values ​​such as the sieve mesh (lower limit of each class) or the D50 average particle size of the particle size distribution obtained by sieving is calculated (Equation 1), Difference = D50 average particle size - reference value (1) The standard deviation σ may be calculated from the variance or unbiased variance of this difference, and particle size distributions that may occur within a specific deviation range such as 1σ or 3σ may be selected. These particle size distributions may then be summed and averaged to be used as the standard particle size distribution in image analysis.

[0017] As shown in Fig. 6, the learning model is constructed by photographing raw soil taken from the soil to obtain image data (S6), calculating the particle size distribution by image analysis (S7), constructing input data by preprocessing using calibration data (S8) (S9), and inputting these input data to the learning model as training data and reflecting them in the learning model (S10). As shown in Fig. 7, the learning model incorporating CNN inputs input data related to particle size distribution based on image data to the convolution layer of the neural network, and outputs particle size distribution result data of the estimated value of the actual measurement of the particle size distribution from the output layer via the intermediate layer. Note that when preparing a learning model incorporating a neural network, there is no problem if a model omitting the convolution layer is constructed.

[0018] As a specific example, FIG. 8 compares the "particle size distribution (sieve)" measured by sieving three types of raw soil (Sample A, B, C) into five particle size ranges, with the "particle size distribution (image)" measured by photographing the raw soil in the same particle size range and analyzing the image. Using sieves with sieve meshes of 2000 μm, 850 μm, 425 μm, 212 μm, and 106 μm, the particle size range is divided into five classes, 850-2000 μm, 425-850 μm, 212-425 μm, 106-212 μm, and 0-106 μm, and the particle size (amount of components) of the particle size range relative to the total amount of raw soil is expressed as a percentage. The particle size distribution by image analysis is processed by adjusting the sensitivity of detection of the contours and boundaries of soil particles. After that, the classification is performed with settings corresponding to the above sieve meshes.

[0019] Grain size ratio distribution B is an evaluation value that evaluates the amount of change from the grain size distribution data A shown in Figure 8 (the measurement result of grain size distribution by sieving) of the original soil. As an evaluation of the amount of change, it is expressed as a grain size ratio based on the grain size of the original soil. The distribution of grain size ratios for each calculated class is treated as a feature included in the grain size distribution data A. The particle size ratio distribution B is calculated by the following formula (2a).

number

number

[0020] The DFT transformed value F is obtained by capturing a curve representing a particle size distribution, such as the particle size distribution data A, as a signal waveform and converting it into the intensity and phase of the frequency components by applying a Fourier transform. In calculating the DFT value F, since the number of classes is set to 5 in this embodiment, the DFT value F is calculated using the following formula (3) using a discrete Fourier transform (DFT), where j represents the imaginary unit.

number

[0021] The DFT evaluation values ​​G and H are values ​​that evaluate the amount of change on the frequency axis with respect to the DFT transformation value F obtained from the particle size distribution data A, relative to the DFT transformation value obtained from the particle size distribution by sieving the raw soil or image analysis. The evaluation of the amount of change is expressed as a ratio based on the DFT evaluation value of the raw soil. The distribution of the calculated DFT evaluation values ​​G and H for each class is treated as a feature included in the particle size distribution data A. The DFT evaluation values ​​G and H are calculated by the following formulas (4a) to (4d).

number

[0022] The composition coefficients D and E are expressed as the composition ratio of the combination of element particle-size distributions, which are particle-size distributions within each class, based on the relationship between the DFT conversion values ​​of the element particle-size distributions and the particle-size distribution data A. Calibration data is obtained according to the acquisition process shown in Figure 5. The analysis result data shows the particle-size distribution obtained by sieving and image analysis of classified soil of each class. Examples are shown in Figures 9a to 11e.

[0023] In order to calculate the composition coefficients D and E, the DFT transformation value of the element particle size distribution obtained by sieving and image analysis is calculated using formula (3). Examples are shown in Figs. 12a to 14e. The particle size distribution obtained by sieving and image analysis can be expressed by superposing element particle size distributions. The following relational equation is derived for the real and imaginary parts of the DFT transformation value.

number

number

number

number

number

number

[0024] The above pre-processing makes it possible to expand the features without padding the data, and to increase the amount of input data to the learning model. As a result of this processing, the data used for learning can be increased by up to eight times compared to the particle size distribution obtained by image analysis. By adopting a particle size distribution of text data, etc., instead of image data, it is possible to avoid data volume bloating and reduce the size of the data. If characteristic data such as the circularity and distortion of soil particles for each class are added when performing image processing on the above input data, the accuracy of the correction can be further improved. Also, instead of the DFT conversion values, coefficients obtained from simultaneous equations consisting of the element grain-size distribution (sieving) and element grain-size distribution (image analysis) in Figures 9a to 11e may be added to the input data in Figure 4.

[0025] To improve the efficiency of learning, each piece of data is standardized as shown in Figure 4 by a pre-processing flow for adding input data prior to inputting it into the learning model. For standardization, the data is processed using the following formula (11) so that the data's average is 0 and the standard deviation is 1. When processing this data, the data to be evaluated after learning is evaluated using the following formula (11) using the average and standard deviation used during learning. Also, since standardization is not essential, standardization may be omitted if not necessary.

number

[0026] The activation function of the connection layer from the intermediate layer to the output layer in Fig. 17 is ReLU (Rectified Linear Unit). ReLU is calculated by the following formula.

number

number

[0027] In Fig. 3, the model constructed by learning is evaluated using an error as the difference between the particle size distribution obtained by measurement by sieving and the estimated value calculated. The error function for evaluating the error uses the mean square error and is given by the following formula (14).

number

[0028] The system shown in Figure 17 incorporates a stochastic descent method with mini-batch learning as an optimization algorithm to avoid learning being hindered by local optimum solutions. In addition, to prevent overfitting, AdaGrad (Adaptive Gradient Algorithm) is implemented, which reduces the amount of weight updates as learning progresses.

[0029] In this embodiment, the raw soil shown in Figure 8 was classified using a classification device, and the treated soil was evaluated after being divided into three categories: a large category discharged as gravel, a medium category discharged as sand, and a small category discharged as silt and clay. A total of 116 pieces of data were obtained in this evaluation test. The breakdown of the data is 29 pieces in the large category, 29 pieces in the medium category, and 29 pieces in the small category. The particle size distribution of the original soil can be considered as the sum of the product of the particle size distribution of each of the treated soils discharged into each category and the mass ratio of each category. For this reason, it is possible to apply a process to calculate the total amount as a data padding process, although the amount of data required for machine learning is small. In this evaluation test, the total amount is 29 pieces. When learning using this data, 80% of the data in each category was used as training data for learning, and the remaining 20% ​​was used as test data to confirm generalization performance. Details of the example in which these data were used to evaluate NN and CNN are described below.

[0030] Examples of image data taken of treated soil for each classification are shown in Figures 15a to 15c. Based on these image data, the particle size distribution is evaluated by image analysis. Examples of input data used for machine learning using CNN are shown in Figures 16a to 16h. As shown in FIG. 18, four different examples of machine learning using CNN were performed with different input items. Learning was performed while initializing the update amount, which decreases every 1000 epochs. When the particle size distribution obtained by sieving for each classified soil was compared with the particle size distribution obtained by image analysis before learning, the mean square error was 0.077, as shown in FIG. 19. On the other hand, when the transition of the mean square error against the learning history was examined, it was confirmed that the mean square error after learning was significantly reduced, and the reduction rate was about 30% on average, as shown in FIG. 20a to FIG. 20d. In addition, when the particle size distribution at the end of learning was compared with the particle size distribution obtained by sieving, it was confirmed that not only the teacher data but also the test data was corrected by learning, as shown in FIG. 21a to FIG. 21d.

[0031] As is clear from the figure, the characteristics of the learning model, such as the progress of learning and the difference in correctable data, differ depending on the convolutional layer with a different configuration. FIG. 22 shows the configuration contents selected from the convolutional layers used in the embodiment shown in FIG. 18 and adjusted the number of filters for parallelization, and FIG. 23 shows the configuration diagram of a CNN in which two convolutional layers with different configurations are parallelized. As a result of learning using this CNN, as shown in FIG. 24, the mean square error for the number of learning times decreases for both the teacher data and the test data, and as shown in FIG. 25, there is almost no difference between the estimated value and the actual measured value for the particle size distribution at the end of learning in the learning result using the teacher data, and although there is a difference between the estimated value and the actual measured value in the learning result using the test data, it can be confirmed that the improvement has progressed. It can be confirmed that the error of the evaluation using the teacher data is low in the section between the dashed lines in FIG. 24, but the error of the evaluation using the test data is also small. The results when early termination was performed in this section are shown in FIG. 26 and FIG. 27. The results of comparing the particle size distribution measured by sieving from the training data and test data and the particle size distribution evaluated by image analysis before and after correction are shown in Figures 28 and 29. Comparing the particle size distribution after correction with the particle size distribution by sieving, it can be confirmed that high reproducibility has been obtained, although some errors remain.

[0032] The configurations shown in the tables of Figures 17 and 18 are complex and have a large amount of input data, so calculations take time. If the selection is made while paying attention to the lack of correlation between the input data, it is also possible to use a NN as shown in Figure 30 instead of a CNN. The input data selected for evaluation with a NN is shown in the table of Figure 31. The network configuration is shown in the table of Figure 32. The NN also implements AdaGrad, just like the CNN. As for the optimization algorithm, learning was performed using a stochastic descent method with mini-batch learning with a batch number of 5. As evaluation results, the transition of the mean square error and the comparison results of the particle size distribution at the end of learning are shown in Figures 33 and 34. Depending on the configuration of the convolution layer and the characteristics of the NN, the data to be corrected can be divided into those that can be corrected and those that cannot. If you know the method or evaluation method that can confirm the effect of correction after processing, it is also effective to use a CNN or NN with a different network configuration to evaluate the effect of correction and select from multiple results, including the results before correction.

[0033] In this embodiment, the raw soil sampled from the soil to be measured was in a stationary state, but if particle size distribution data can be obtained by image analysis, it is possible to target flowing soil particles, etc., and it is also possible to apply this to powders other than soil. A digital camera was used to perform the image analysis, but there is no problem in using a high-speed camera or a high-resolution camera, etc. It is also considered acceptable to make corrections using different measurement methods, such as sieve measurements and laser diffraction scattering methods. It is also effective as a pre-processing step for evaluation when adjustments or controls need to be made during handling based on particle size distribution data, such as changing sieves or adjusting the amount of chemicals added. [Explanation of symbols]

[0034] 1. Imaging device 2 Arithmetic unit 2a Image analysis function section 2b Preprocessing and learning model function part A1, A2, An image data B1.B2,...Bn Particle size distribution measurement data C1, C2, Cn Particle size distribution estimation data

Claims

1. A method for constructing a learning model for measuring particle size distribution by image analysis of soil particles for correcting errors contained in the particle size distribution obtained by image analysis, comprising inputting a large number of particle size distributions obtained by image analysis of soil particles in image data obtained by photographing soil used for learning, a particle size distribution obtained by sieving the soil used for learning, and distribution data linking the particle sizes obtained by image analysis and the particle sizes obtained by sieving, and performing machine learning; The learning input data for the learning model includes the particle size of the raw soil obtained by sieving the raw soil collected from the soil into arbitrary classes, The grain size of the original soil obtained by photographing the same original soil and analyzing the images of the soil particles in the same classification, and A particle size distribution A of the treated soil obtained by image analysis of the treated soil classified into a predetermined particle size range; A particle size ratio distribution B, which represents the ratio of particle sizes of the treated soil obtained by image analysis to particle sizes of the raw soil obtained by sieving for each class; A particle size ratio distribution C representing the ratio of particle sizes of the treated soil by image analysis to particle sizes of the original soil by class by image analysis; A DFT transformation value F obtained by converting the particle size of the treated soil obtained by image analysis into the intensity and phase of frequency components using a Fourier transform; DFT evaluation values ​​G and H representing the ratio of the DFT conversion value F to a DFT conversion value obtained from the particle size of the raw soil by sieving or the particle size of the raw soil by image analysis; The raw soil is sieved into arbitrary classes, and the element particle size distribution is obtained from the single particle size classified soil within each class by sieving and image analysis. A method for constructing a learning model for measuring particle-size distribution by image analysis of soil particles, comprising: composition coefficients D and E that represent a composition ratio of a combination of element particle-size distributions from a relationship between the element particle-size distribution and a DFT transformation value obtained from the particle size obtained by sieving the raw soil or the particle size obtained by image analysis of the raw soil.

2. 2. The method for constructing a learning model for measuring particle size distribution by image analysis of soil particles according to claim 1, wherein the particle size ratio distributions B and C are calculated by the following formula: ##EQU00015## ##EQU00016##

3. 2. The method for constructing a learning model for measuring particle size distribution by image analysis of soil particles according to claim 1, wherein the DFT transformation value F is calculated by the following formula: ##EQU00017##

4. 2. The method for constructing a learning model for measuring particle size distribution by image analysis of soil particles according to claim 1, wherein the DFT evaluation values ​​G and H are calculated by the following formula: [0018]

5. 2. The method for constructing a learning model for measuring particle size distribution by image analysis of soil particles according to claim 1, wherein the configuration coefficients D and E are determined based on the following formula: [0019]

Images