Cement mixture quality prediction system

Abstract

This invention provides a system that can automatically predict the 3D printing performance of a cement mixture during its mixing process. [Solution] The quality prediction system includes a storage unit that records a trained model generated by machine learning based on multiple training data, which associates training image data based on a group of image data captured during the mixing process of a training cement mixture for 3D molding with training performance data relating to an index value of the 3D molding performance of the training cement mixture after the completion of the mixing process; an input receiving unit that receives input of target first input data based on image data of a target cement mixture for 3D molding during the mixing process; and a prediction processing unit that applies target image data corresponding to the training image data identified based on the target first input data to the trained model to generate prediction information relating to an index value of the 3D molding performance of the target cement mixture.

Description

[Technical Field]

[0001] The present invention relates to a quality prediction system for cement mixtures, and more particularly to a quality prediction system for cement mixtures used in three-dimensional molding. [Background technology]

[0002] A construction 3D printer (sometimes abbreviated as "3DP") is a specialized 3D printer used to construct buildings and structures. Compared to conventional construction methods, 3DPs can produce concrete structures more quickly and efficiently, and are expected to improve the efficiency of the construction process and the quality of buildings.

[0003] While many technologies exist in 3DP that automate the lamination mechanism, actually streamlining the construction process requires the development of technologies that can reduce labor in areas other than lamination. From a quality assurance standpoint, the current practice is to follow conventional quality control methods, such as measuring the flow of mortar immediately after mixing, and preparing and testing specimens for strength testing.

[0004] When forming structures using 3D deposition modeling (3DP), it is necessary to stack cement mixtures to the desired height. Therefore, the indicator of how high the cement-based material can be stacked while maintaining its shape (referred to as "self-supporting stability" or "stack stability") is a very important parameter.

[0005] The applicants have previously proposed a method for evaluating the properties of cement-based materials that can easily assess the lamination stability of cement mixtures. (See Patent Document 1 below) [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2023-176290 [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] The technology disclosed in Patent Document 1 is characterized by including a measurement step of measuring the vane shear strength of a cement mixture, a calculation step of calculating the shear stress generated in the laminate, and an evaluation step of comparing the vane shear strength and the shear stress, and evaluating that the cement-based material has excellent lamination stability if the value of the vane shear strength is greater. According to this method, the lamination stability when a cement mixture is used in 3DP can be easily evaluated without using special equipment.

[0008] However, when evaluating properties using the method described in Patent Document 1, it is necessary to determine the vane shear strength of the cement mixture using a general vane testing machine. Therefore, since it is necessary to take a test sample from the cement mixture and measure it, the advantages of automation and autonomy of 3DP may be lost.

[0009] In view of the above problems, the present invention aims to provide a system that can automatically predict the three-dimensional molding performance of a cement mixture for three-dimensional molding while it is being mixed. [Means for solving the problem]

[0010] The cement mixture quality prediction system according to the present invention is A storage unit records a trained model generated by machine learning based on multiple training data sets, which is formed by associating training image data generated based on a group of imaging data taken at different times during the mixing process of a training cement mixture for 3D modeling in a mixing device, with training performance data relating to an index value of the 3D modeling performance of the training cement mixture after the mixing process is completed. An input receiving unit that receives input of target first input data, which is a set of multiple image data itself, or a set of data generated from said image data, of the target cement mixture, which is a cement mixture for 3D modeling being mixed in a mixing device, at different time points, Based on the target first input data received by the input receiving unit, a prediction processing unit identifies target image data corresponding to the training image data and applies the target image data to the trained model to generate prediction information regarding the index value of the three-dimensional molding performance of the target cement mixture. The system is characterized by comprising an output unit that outputs the aforementioned prediction information.

[0011] As an indicator of the three-dimensional molding performance of cement mixtures for three-dimensional molding, an indicator related to self-supporting stability (shape retention) can be cited. In addition, at least one of the following indicators may be included: an indicator related to extrudeability (flow retention) and an indicator related to pumpability.

[0012] Extrudeability refers to the performance required to continuously extrude cement mixtures from a nozzle, and corresponds to the ability to maintain fluidity. It is sometimes also called "fluidity retention." Pumpability refers to the performance required to continuously supply cement mixtures by a pump, and is determined by the extrusion stress and extrusion speed. Self-supporting stability (shape retention) is an indicator of how well a cement mixture can maintain its shape after being laminated.

[0013] For example, increasing extrudeability (flow retention) decreases self-supporting stability (shape retention), so these indicators show a trade-off relationship with one another.

[0014] As a result of our diligent research, we have newly discovered that there is an extremely high correlation between images captured during the mixing of a cement mixture for 3D modeling in a mixing device and an index value of 3D modeling performance. The relationship between the index value of 3D modeling performance and images captured during the mixing of a cement mixture for 3D modeling was completely unknown until now, and this is a surprising fact that we have derived as a result of our diligent research.

[0015] The quality prediction system according to the present invention uses a trained model created by machine learning based on multiple training data sets that associate images of a cement mixture for 3D molding being mixed with index values ​​of the 3D molding performance of the cement mixture corresponding to the images. This trained model is recorded in the memory unit of the quality prediction system.

[0016] The training data used in machine learning is data that associates training image data and training performance data. Here, "training image data" is data generated based on a set of image data captured while a cement mixture for 3D modeling is being mixed in a mixing device. Hereafter, the cement mixture for 3D modeling that is the subject of imaging will be referred to as the "training cement mixture." A cement mixture for 3D modeling is typically a mortar made by mixing cement-based materials, fine aggregate, and admixtures.

[0017] In this specification, "imaging data" refers to still images captured at predetermined timings. A video captured continuously over a predetermined period of time is a group of multiple still images captured at different times, and corresponds to a "group of imaging data with different capture times." Similarly, multiple still images obtained by capturing at predetermined time intervals (e.g., every minute) also correspond to a "group of imaging data with different capture times."

[0018] The image data for learning is data based on an imaged data group of the cement kneaded body for learning, and for example, is image data obtained by cutting out a specific region (hereinafter referred to as "reference region") from each piece of imaged data and performing grayscale conversion at a predetermined resolution (hereinafter referred to as "reference resolution"). That is, the image data for learning is a plurality of image data groups obtained from a plurality of pieces of imaged data captured at different times while the same cement kneaded body for learning is being kneaded. More specifically, the image data for learning can be information relating to the luminance distribution of each pixel contained in an image obtained by cutting out the imaged data of the cement kneaded body for learning captured in the kneading apparatus at a predetermined resolution.

[0019] The performance data for learning is data relating to the index value of the three-dimensional shaping performance of the cement kneaded body for learning, which is obtained by a predetermined performance test using a test specimen obtained as a sample after the kneading treatment on the cement kneaded body for learning is completed. As the performance test, a lamination height evaluation test can be used. In addition to this, at least one of a mortar flow test and a vane shear strength measurement test may be used. As the index value, a measured value obtained through the test results can be adopted.

[0020] For the mortar flow test, for example, a method conforming to JIS R 5201 can be used. For the vane shear strength measurement test, for example, a method conforming to the method specified in "In-situ Vane Shear Test Method" of the Japanese Geotechnical Society Standard (JGS1411-012) can be used.

[0021] For the lamination height evaluation test, a method can be used in which the cement kneaded body for learning is extruded to laminate molded bodies of a predetermined shape, typically a columnar body with a hollow rectangular cross section, and the lamination height until the lower part of the laminate collapses is evaluated.

[0022] In other words, the training data Di for the training cement mixture i is prepared by associating training image data dgi, which consists of a set of image data based on imaging data captured at different times during mixing, with training performance data dci, which is an index value of the 3D molding performance of the training cement mixture i. Such a set of training data is prepared for each training cement mixture i (i=0, 1, …).

[0023] In the learning unit, machine learning is performed using training data Di (i=0, 1, …), which is formed by associating training image data dgi with training performance data dci, and a trained model is generated. The obtained trained model is recorded in the memory unit.

[0024] This section describes how to predict the 3D printing performance of a cement mixture used for 3D printing while it is being mixed. In the following, the cement mixture used for 3D printing that is the subject of the prediction will be referred to as the "target cement mixture."

[0025] The target cement mixture is imaged intermittently or continuously by an imaging unit while being mixed in the mixing device. The former corresponds to capturing multiple still images at predetermined time intervals, while the latter corresponds to capturing moving images. In this way, the image data itself, captured at different times while the learning cement mixture is being mixed in the mixing device, or the data generated by applying image processing to these image data, is input into the quality prediction system according to the present invention. Hereinafter, the data group that the input receiving unit of the quality prediction system accepts will be referred to as the "target first input data".

[0026] The quality prediction system has a prediction processing unit. The prediction processing unit is a functional means for performing calculations. First, the prediction processing unit identifies data corresponding to training image data (target image data) from the target first input data input to the quality prediction system. For example, if the target first input data is a group of image data taken at different times, the prediction processing unit extracts a reference region from each of the image data taken at different times and performs image processing to convert them to grayscale at a reference resolution to create image data. The prediction processing unit identifies these multiple created image data as "target image data". If the target first input data has already undergone image processing before being input to the input receiving unit, the prediction processing unit identifies the input target first input data itself as "target image data".

[0027] The prediction processing unit applies the identified "target image data" to a trained model stored in the memory unit to generate prediction information regarding the index values ​​of the 3D molding performance of the target cement mixture. The generated prediction information is output from the output unit.

[0028] As described above, the cement mixture quality prediction system according to the present invention can automatically output predicted index values ​​for the 3D molding performance of the target cement mixture by imaging the mixture during mixing and sending the image data or data processed from the image data to the prediction processing unit. In other words, since predicted values ​​can be output in real time during the mixing process, there is no need to conduct separate tests to obtain index values ​​after mixing is completed, as in the conventional method. The superior accuracy of these prediction results will be described later with reference to the examples.

[0029] The prediction information may be configured as a combination of index values ​​for multiple items. In this case, the cement mixture quality prediction system may include a determination unit that determines whether the quality of the target cement mixture is good or bad when used for three-dimensional molding based on the prediction information, and the output unit may output the determination result from the determination unit.

[0030] In this case, the training performance data dci for the 3D printing performance index of the training cement mixture i, which is prepared as training data Di for the training cement mixture i, is prepared according to the number of items that make up the index value. If the layer height is adopted as the index value for 3D printing performance, one data set is prepared as the training performance data dci corresponding to the training cement mixture i. If there are multiple items as index values, the training performance data dci for the 3D printing performance index of the cement mixture i will be a set of training performance data dcij (j=1, …, k), where k is the number of items and j is prepared for each item. For example, if the mortar flow value and vane shear strength listed above are adopted as index values ​​for 3D printing performance in addition to the layer height, this corresponds to the number of items k=3.

[0031] In this case, a set of training image data dgi, consisting of image data based on imaging data captured at different times during mixing, and a set of training performance data dcij (j=1, …, k) relating to the index values ​​of the 3D molding performance of the training cement mixture i are prepared in an associated state. Such a set of training data is prepared for each training cement mixture i (i=0, 1, …). In the learning unit, machine learning is performed using the training data Di (i=0, 1, …), which is formed by associating the training image data dgi and the training performance data dcij, and a trained model is generated. The obtained trained model is recorded in the memory unit. When only the layer height is targeted as item j, k=1.

[0032] As described above, the prediction processing unit identifies data corresponding to the training image data (target image data) from the target first input data input to the quality prediction system. Then, by applying the identified "target image data" to the trained model stored in the memory unit, it generates prediction information that predicts the index values ​​of the 3D molding performance of the target cement mixture for each item j. For example, if only the layer height is assumed as item j, prediction information including information on the layer height is generated as the index value of the 3D molding performance of the target cement mixture. Also, if item j includes all three types: mortar flow value, vane shear strength, and layer height, prediction information including these three types of index values ​​is generated as the index value of the 3D molding performance of the target cement mixture.

[0033] The index values ​​for the 3D molding performance of a cement mixture include those that exhibit trade-offs with each other. The performance required for a cement mixture changes depending on the object to be molded using that mixture. Therefore, by pre-setting the conditions for each index value as the performance required for the target cement mixture, the judgment unit can determine whether the index values ​​of each item included in the prediction information satisfy the conditions, thereby determining the quality of the 3D molding performance of the target cement mixture during mixing.

[0034] The kneading apparatus comprises a kiln for containing the material to be kneaded, a stirring blade for stirring the material, and a drive shaft that provides rotational drive to the stirring blade. The aforementioned target first input data may be a group of image data obtained by imaging the target cement mixture located closer to the inner wall of the kiln than to the center of the kiln, or a group of data generated from said image data.

[0035] As described above, the area where the cement mixture is photographed during mixing is close to the inner wall of the kiln, making it an area where the cement mixture is likely to collide with the inner wall and change shape. This increases the correlation between the index value of the 3D printing performance of the cement mixture and the images of the cement mixture being mixed for 3D printing, allowing for more accurate prediction of the index value of the 3D printing performance of the target cement mixture.

[0036] The kneading apparatus comprises a kiln for containing the material to be kneaded, a stirring blade for stirring the material, and a drive shaft that provides rotational drive to the stirring blade. The aforementioned target first input data may be, at a minimum, a group of image data consisting of images of the target cement mixture located within the region in which the stirring blade or the drive shaft is captured, or a group of data generated from said image data.

[0037] As described above, the area where the cement mixture is photographed during mixing is near the stirring blades or drive shaft, and this area is prone to shape changes due to the movement of the stirring blades. This improves the correlation between the index value of the 3D molding performance of the cement mixture and the images of the cement mixture being mixed for 3D molding, allowing for more accurate prediction of the index value of the 3D molding performance of the target cement mixture.

[0038] The storage unit records the trained model generated by machine learning based on the plurality of training data, which is formed by associating the training image data, the training mix data relating to the training cement mixture mix information, and the training performance data. The input receiving unit is configured to receive, in addition to the target first input data, target second input data which is information regarding the composition of the target cement mixture. The prediction processing unit may also identify target blending data corresponding to the training blending data based on the target second input data received by the input receiving unit, and generate the prediction information by applying the target image data and the target blending data to the trained model.

[0039] In this case, the learning unit may generate the trained model by performing machine learning based on the multiple training data sets, in which the training image data, the training formulation data, and the training performance data are associated. [Effects of the Invention]

[0040] According to the present invention, a system is realized that enables the automatic prediction of the three-dimensional molding performance of a cement mixture for three-dimensional molding while the mixture is being kneaded. [Brief explanation of the drawing]

[0041] [Figure 1] This diagram schematically shows the state of a cement mixture during mixing, which is the target of predicting the index value of 3D molding performance. [Figure 2] These are schematic diagrams and photographs showing images of the target cement mixture 2 during mixing. [Figure 3] This is a schematic block diagram showing the configuration of the first embodiment of the quality prediction system. [Figure 4] This is another block diagram schematically showing the configuration of the first embodiment of the quality prediction system. [Figure 5] This is another block diagram schematically showing the configuration of the first embodiment of the quality prediction system. [Figure 6] This is a block diagram schematically showing the configuration of the second embodiment of the quality prediction system. [Figure 7] This is another block diagram schematically showing the configuration of the second embodiment of the quality prediction system. [Figure 8] This is another block diagram schematically showing the configuration of the second embodiment of the quality prediction system. [Modes for carrying out the invention]

[0042] Hereinafter, embodiments of the cement mixture quality prediction system according to the present invention (hereinafter sometimes abbreviated as "quality prediction system") will be described with reference to the drawings as appropriate. Note that the following drawings are all schematic illustrations.

[0043] The quality prediction system according to the present invention is a system that predicts an index value of the three-dimensional molding performance of a cement mixture for three-dimensional molding by performing calculation processing while the mixture is being kneaded.

[0044] Figure 1 is a schematic diagram showing the mixing process of the cement mixture to be predicted (referred to as the "target cement mixture"). The target cement mixture 2 shown in Figure 1 is being mixed in a mixing device 4, and the mixing process is being imaged by the imaging unit 3.

[0045] The mixing device 4 illustrated in Figure 1 is a device called a pan-type mixer. The container (oven, tank) that houses the target cement mixture 2, which is the material to be mixed, is cylindrical in shape, and a drive shaft 6 that provides rotational drive is installed in the center of the container, and stirring blades 7 connected to the drive shaft 6 via arms 8 are installed. Multiple stirring blades 7 are installed, and the lengths of each arm 8 are different from each other in order to mix the target cement mixture 2 evenly. The target cement mixture 2 is mixed by rotating the stirring blades 7 inside the oven via each arm 8 through the rotational drive of the drive shaft 6. However, in the present invention, the mixing device 4 for mixing the target cement mixture 2 is not limited to a pan type.

[0046] The imaging unit 3 is mounted in a position that allows it to image the target cement mixture 2 inside the mixing device 4. Specifically, the imaging unit 3 images the imaging area PA inside the mixing device 4. This imaging area PA is the area in which the target cement mixture 2 being mixed is captured. Preferably, the imaging area PA is a location close to the inner wall of the kiln, or a location in which the drive shaft 6 or stirring blade 7 is captured. Both of these are locations where the movement of the target cement mixture 2 during mixing tends to be relatively large. In the former, the target cement mixture 2 is likely to collide with the inner wall of the kiln during mixing, changing its vector, while in the latter, it is likely to change its vector after being lifted by the stirring blade 7 and then dropped.

[0047] Figure 2 is a schematic diagram and photograph showing images of the target cement mixture 2 during mixing. The imaging unit 3 continuously captures images of the imaging area PA in motion over a predetermined time (hereinafter referred to as "imaging time Tz"). The motion image is a group of multiple still images taken at different times, and corresponds to a "group of imaging data at different times". Figure 2 is a schematic diagram showing the imaging data da(t) of the target cement mixture 2 during mixing, captured at a certain time t. In this way, data captured at different times t over the imaging time Tz is obtained as imaging data da(t).

[0048] As described later, target image data dA(t) at time t is generated by applying predetermined image processing to the image data da(t) captured at time t. In other words, target image data dA(t) is obtained, which is a set of data based on image data da(t) captured at different times t over the imaging time Tz.

[0049] Furthermore, it is possible to generate multiple target image data dA(t) at time t from imaging data da(t) captured at time t. For example, when cropping da(t), which is imaging data of the imaging region PA, to a predetermined size, by shifting the cropping position by predetermined pixels, target image data dAq(t) (q=(0,0), (0,1), … (m, n)) is generated for each cropping position q at time t. In this case, the target image data dA(t) at time t can be considered as a group of target image data dAq(t) equal to the number of cropped images.

[0050] In the example shown in Figure 2, a specific region (hereinafter referred to as the "reference region") is extracted from the imaging data da(t), and the target image data dA(t) is obtained by performing a grayscale conversion process at a predetermined resolution (hereinafter referred to as the "reference resolution"). As described above, multiple reference regions q can be set, in which case, the number of target image data dAq(t) equal to the number of set reference regions q can be obtained at the same time t. Based on the multiple target image data dA(t) obtained at different time ts, an index value of the 3D molding performance of the target cement mixture 2 is predicted.

[0051] Figure 3 is a schematic block diagram showing the configuration of the quality prediction system of the first embodiment. The quality prediction system 1 includes an imaging unit 3 and a prediction device 20. As shown in Figure 1, the imaging unit 3 is a mechanism that images the target cement mixture 2 being mixed in the mixing device 4, and a known digital camera or the like can be used. The imaging unit 3 may be a network-type digital camera equipped with communication functions.

[0052] The prediction device 20 includes an input receiving unit 21, a prediction processing unit 22, an output unit 23, and a storage unit 24.

[0053] The input reception unit 21 is a functional unit that receives input of data from the outside of the prediction device 20 into the interior of the prediction device 20. Upon receiving the input of the imaging data da(t) of the target cement kneaded body 2 from the imaging unit 3, the input reception unit 21 performs predetermined pre-processing and then outputs the processed data to the prediction processing unit 22. The "pre-processing" referred to herein means, for example, authentication processing when an external data input is recognized, processing for converting the data format into a mode that allows data exchange inside the prediction device 20, and the like.

[0054] As described above, the imaging data da(t) is taken into the prediction device 20 from the imaging unit 3 as data for each time t over the imaging time Tz. The imaging data da(t) may be input from the imaging unit 3 to the prediction device 20 each time an image is captured, or after the imaging data da(t) is accumulated over a predetermined batch time Tb (<Tz), the accumulated imaging data da(t) for each time t within the batch time Tb may be collectively input to the prediction device 20. Here, the imaging data da(t) corresponds to "target first input data".

[0055] The prediction processing unit 22 generates target image data dA(t) based on sequentially fed imaging data da(t). The prediction processing unit 22 then applies this target image data dA(t) to a trained model M1 recorded in the storage unit 24 to generate prediction information do regarding the index value of the 3D molding performance of the target cement mixture 2, and outputs it via the output unit 23. The prediction processing unit 22 is a processing means that performs calculations using the trained model M1 and outputs the results, and is composed of, for example, a CPU or MPU. The storage unit 24 is a storage medium or storage area in which predetermined information including the trained model M1 is stored, and is typically a non-volatile memory such as flash memory, a hard disk, or cloud storage. The output unit 23 is an output interface, and is a functional means that performs signal processing for sending information to, for example, a display means. Specifically, the output unit 23 performs processing to output the prediction information do from the prediction processing unit 22 to a display means such as a monitor located at a remote location via a wireless or wired telecommunication line. Furthermore, if the computer equipped with the arithmetic processing means corresponding to the prediction processing unit 22 is also equipped with a monitor, the output unit 23 may output the prediction information do to the monitor.

[0056] Next, the trained model M1 will be described. Figure 4 is another block diagram schematically showing the configuration of the quality prediction system of the first embodiment. The quality prediction system 1 comprises a training data input unit 41 and a learning device 10. The learning device 10 comprises an input receiving unit 11, a learning unit 12, and a storage unit 13.

[0057] The input receiving unit 11 is a functional means for receiving data input from outside the learning device 10 into the learning device 10. Specifically, the input receiving unit 11 receives teacher data Di (i=0, 1, …) from the learning data input unit 41.

[0058] The training data Di is data that associates training image data dgi with training performance data dci. These data are described below.

[0059] The training image data (dgi) is data generated based on a set of image data captured while a training cement mixture i for 3D modeling was being mixed in a mixing device. In the following, for convenience of explanation, the image data captured at time t of the training cement mixture i being mixed in the mixing device will be denoted as "dfi(t)". ​​Multiple sets of image data (dfi) are obtained by continuously or intermittently capturing the training cement mixture i during the mixing process over the imaging period Ti.

[0060] Here, the mixing device (hereinafter referred to as "mixing device 4i") used when the learning cement mixture i is imaged may be the same mixing device 4 used to mix the target cement mixture 2, or it may be a different mixing device from mixing device 4 that achieves substantially the same mixing conditions as mixing device 4. For example, mixing device 4i may be a mixing device of the same type as mixing device 4, or it may be a mixing device that has substantially the same size and dimensional ratio as mixing device 4.

[0061] Furthermore, the imaging data of the learning cement mixture i during the mixing process may be captured by an imaging unit installed in a predetermined orientation. Typically, the imaging data of the learning cement mixture i can be obtained by imaging the learning cement mixture i during the mixing process using an imaging unit adjusted to capture the same imaging area PA as the imaging unit 3 described above (see Figure 1).

[0062] The training image data dgi is a set of data obtained by performing a predetermined image processing on each of the image data sets dfi, which are captured while the training cement mixture i is being mixed in a mixing device. For example, the training image data dgi(t) at a certain time t is data obtained by extracting a reference region from the image data dfi(t) and converting it to grayscale at a reference resolution. Similar image processing is performed on the image data dfi(t) at each time t over the imaging period Ti, thereby preparing multiple sets of training image data dgi. The reference region can also be set at multiple locations that are adjacent or close to each other.

[0063] The learning performance data DCI is data relating to the index values ​​of the three-dimensional molding performance of the learning cement mixture i, obtained by performing a predetermined performance test using a test specimen obtained as a sample after the mixing process of the learning cement mixture i is completed in the mixing device 4i. The learning performance data DCI can be data based on the value of the stacking height until the lower part of the stack collapses, after extruding the learning cement mixture i to stack a predetermined shape, typically a hollow rectangular cross-section columnar body.

[0064] Another example of learning performance data (dci) is to perform a mortar flow test in accordance with JIS R 5201 using the learning cement mixture i after the mixing process, and obtain data based on the resulting flow values.

[0065] Another example of DCI performance data for learning is data based on hardness values ​​obtained by conducting tests in accordance with the rubber hardness measurement method compliant with JIS K 6253.

[0066] The DCI performance data used for training can be based on a single metric value (typically the stack height value) or on the metric values ​​of multiple items.

[0067] In other words, the training data Di for the training cement mixture i, which is input from the training data input unit 41 to the training device 10, is prepared by associating training image data dgi, which consists of a group of image data based on imaging data dfi(t) captured at different time t during mixing, with training performance data dci, which relates to an index value of the three-dimensional molding performance of the training cement mixture i. Such a group of training data is prepared for each training cement mixture i (i=0, 1, …).

[0068] The learning unit 12 performs machine learning using training data Di (i=0, 1, …), which is formed by associating training image data dgi and training performance data dci, and generates a trained model M1. The generated trained model M1 is recorded in the storage unit 13. Here, "machine learning" refers to learning by a machine (especially a computer) alone, without human thought. The learning unit 12 is a means of computation. The storage unit 13 is a storage medium or storage area.

[0069] The training data Di used for learning is prepared for each training cement mixture i. The number of samples in the training data Di (i.e., the number of i) is preferably 50 or more, more preferably 1,000 or more, and particularly preferably 10,000 or more, from the viewpoint of improving judgment accuracy. The number of training iterations is preferably 50,000 or more, more preferably 100,000 or more, and particularly preferably 200,000 or more, from the viewpoint of improving judgment accuracy.

[0070] Examples of machine learning methods include neural networks, linear regression, decision trees, support vector regression, ensemble methods, support vector machines, discriminant analysis, simple Bayesian methods, and nearest neighbor methods. Only one of these methods may be used, or two or more may be combined.

[0071] Among these methods, machine learning using neural networks is preferred from the viewpoint of being able to predict quality with higher accuracy. A hierarchical neural network having one or more hidden layers between the input layer and the output layer is particularly preferred from the viewpoint of being able to predict quality with higher accuracy.

[0072] Examples of neural networks include convolutional neural networks (CNNs), such as 3D convolutional neural networks (3DCNNs), deep neural networks (DNNs), recurrent neural networks (RNNs), and long-short-term memory (LSTM) neural networks (an improved version of a recurrent neural network using LSTMs).

[0073] Among these, convolutional neural networks (neural networks having convolutional layers, pooling layers, etc. as hidden layers) are more preferred because they have excellent performance in the field of image recognition. Convolutional neural networks can detect features from image data and create predictive models that can perform classification or regression using these features. In a convolutional neural network, the number of layers consisting of combinations of convolutional layers and pooling layers is preferably two or more, and more preferably three or more, from the viewpoint of being able to make predictions with higher accuracy.

[0074] Furthermore, tools for performing machine learning include, for example, "TensorFlow®," a software library developed by Google, and "IBM Watson®," a system developed by IBM.

[0075] Thus, the trained model M1 created by machine learning using the training data Di (i=0, 1, …) by the learning unit 12 is recorded in the memory unit 24 in Figure 3. Note that the memory unit 24 in Figure 3 and the memory unit 13 in Figure 4 may refer to the same memory area, or they may refer to different memory areas.

[0076] The prediction processing unit 22 applies the target image data dA(t) to a pre-trained model M1 to generate prediction information do regarding the index value of the 3D molding performance of the target cement mixture 2. This prediction information do is output via the output unit 23, allowing the index value of the 3D molding performance of the target cement mixture 2 to be recognized in real time during the mixing process of the target cement mixture 2.

[0077] As shown in Figure 5, the quality prediction system 1 may include a determination unit 25. This determination unit 25 is a functional means that determines whether the target cement mixture 2 is good or bad based on prediction information do regarding the index values ​​of the three-dimensional molding performance of the target cement mixture 2, which is generated by the prediction processing unit 22. The determination unit 25 records information regarding preferred numerical ranges for the index values ​​of the three-dimensional molding performance required for the target cement mixture 2, which have been set in advance. Hereinafter, this information will be referred to as "good product condition information" for convenience. More specifically, when multiple items are set as index values ​​for three-dimensional molding performance, the determination unit 25 records information regarding the preferred numerical range for each item as good product condition information. When the determination unit 25 receives prediction information do from the prediction processing unit 22, it determines whether each index value described in prediction information do is included in the preferred numerical range based on the recorded good product condition information, and outputs the determination result dp from the output unit 23.

[0078] This allows the quality judgment result of the target cement mixture 2 to be recognized in real time during the mixing process of the target cement mixture 2. The output unit 23 outputs not only the judgment result dp of the target cement mixture 2, but also predictive information do regarding the index value of the 3D molding performance of the target cement mixture 2. Therefore, even if the judgment result dp indicates an undesirable result, it is possible to recognize which performance is particularly undesirable based on the predictive information do. Thus, countermeasures can be taken, such as temporarily stopping the mixing process of the target cement mixture 2 and adjusting the mix conditions.

[0079] Figure 6 is a schematic block diagram showing the configuration of a second embodiment of the quality prediction system. The quality prediction system 1 shown in Figure 5 differs from the quality prediction system 1 shown in Figure 3 in that it includes a mix information input unit 5. In this embodiment, the input receiving unit 21 receives the input of imaging data da(t) of the target cement mixture 2 from the imaging unit 3 and the input of data dβ related to the mix information of the target cement mixture 2 from the mix information input unit 5. This data dβ corresponds to the "target second input data".

[0080] Furthermore, the quality prediction system 1 shown in Figure 6 stores a trained model M2 in the memory unit 24. This trained model M2 is a model generated by machine learning based on multiple training data Di, which are associated with training image data dgi, training performance data dci, and training formulation data dbi.

[0081] Information regarding the composition of the target cement mixture 2 includes the unit water content (kg / m³). 3 ), unit cement content (kg / m 3 ), Unit amount of fine aggregate (kg / m 3 ), unit volume mass (kg / m³) 3 ), fine aggregate ratio (%), unit admixture amount (kg / m 3 This includes one or more pieces of information belonging to the group consisting of ), and the water-cement ratio (%).

[0082] The prediction processing unit 22 identifies target formulation data corresponding to the learning formulation data, based on target second input data dβ input from the formulation information input unit 5 (hereinafter, for convenience, referred to as "target formulation data db"). Specifically, the target formulation data db is identified by extracting information related to the items listed in the learning formulation data dbi from the information described in the target second input data dβ. Note that the target second input data dβ itself may be used as the target formulation data db.

[0083] The prediction processing unit 22 generates target image data dA(t) based on the sequentially fed imaging data da(t), and applies this target image data dA(t) and the target formulation data db to the learned model M2 recorded in the storage unit 24, thereby generating prediction information do related to the index value of the three-dimensional shaping performance of the target cement kneaded body 2, and outputs it via the output unit 23.

[0084] Next, the learned model M2 will be described. Figure 7 is another block diagram schematically showing the configuration of the quality prediction system of the second embodiment. The present embodiment differs from the first embodiment in that the teacher data Di is different. That is, in the present embodiment, the teacher data Di is data in which learning image data dgi, learning formulation data dbi, and learning performance data dci are associated with each other.

[0085] The learning image data dgi and the learning performance data dci have been described above in the first embodiment, so descriptions thereof will be omitted. The learning formulation data dbi is data related to the formulation information of the learning cement kneaded body i, and specifically, includes unit water content (kg / m 3 ), unit cement content (kg / m 3 ), unit fine aggregate content (kg / m 3 ), unit volume mass (kg / m 3 ), fine aggregate ratio (%), unit admixture content (kg / m 3 ), and one or two or more types of information belonging to the group consisting of water-cement ratio (%).

[0086] The learning unit 12 performs machine learning using training data Di (i=0, 1, …), which is formed by associating training image data dgi, training formulation data dbi, and training performance data dci, and generates a trained model M2. The generated trained model M2 is recorded in the storage unit 13. The training data Di used for training is prepared for each training cement mixture i (i=0, 1, …).

[0087] In other words, in the trained model M2 of this embodiment, the relationship between the index value of 3D fabrication performance and not only the information regarding the captured image of the training cement mixture i during the mixing process, but also the formulation information of the training cement mixture i, has been learned. Therefore, the prediction processing unit 22 can generate prediction information do regarding the index value of 3D fabrication performance by applying the target image data dA(t) and the target formulation data db to this trained model M2 recorded in the storage unit 24, while also considering the formulation information of the target cement mixture 2.

[0088] In this embodiment as well, similar to the first embodiment, the quality prediction system 1 may include a determination unit 25 that determines whether the target cement mixture 2 is good or bad based on the prediction information do relating to the index value of the three-dimensional molding performance of the target cement mixture 2 generated by the prediction processing unit 22 (see Figure 8).

[0089] [Alternative Embodiment] The following describes another embodiment of the cement mixture quality prediction system according to the present invention.

[0090] (1) Image processing may be performed on the imaging data da(t) inside the imaging unit 3, or between the imaging unit 3 and the prediction device 20, to generate target image data dA(t), and this target image data dA(t) may be input to the prediction device 20.

[0091] (2) The imaging data da(t) may be used as the target image data dA(t) without being extracted. [Examples]

[0092] The present invention will be described below with reference to examples, but is not limited to the examples described.

[0093] (Combination conditions) As a learning cement mixture i, a sample was prepared by mixing water, ordinary Portland cement, fine aggregate, and a high-performance water-reducing agent.

[0094] (Mixing process) In the mixing process, the above-mentioned material premix powder was placed into a pan-shaped mixing device 4, schematically shown in Figure 1, and mixed for 2 minutes to produce multiple samples of cement mixture for 3D molding.

[0095] During the mixing process, images of the cement mixture were acquired using a digital camera installed above the mixing device 4. The mixing image resolution was set to 1,920px x 1,080px, the frame rate to 30FPS, and the shooting area was limited to the area where the stirring blades 7 were partially visible.

[0096] (Test conditions) For each sample, a cement mixture was taken after the mixing process was completed, and the layer height was measured. This value constitutes an index value for the three-dimensional molding performance of the cement mixture.

[0097] The stacking height was defined as the height at which the lower part of the stacked structure collapsed after extruding the cement mixture into a predetermined shape. Specifically, the cement mixture was stacked to form a hollow rectangular cylindrical body with outer dimensions of 100 mm x 100 mm x target stacking height of 400 mm, a layer thickness of 10 mm, and a layer width of 16 mm, and the stacking height at which collapse began was measured.

[0098] (Machine Learning) In creating the trained model M1, the mixing images of the cement mixture were correlated with the index values ​​of the 3D fabrication performance for each of the above items for each sample. Specifically, still images were obtained from 9 seconds of video footage captured by a digital camera, with a frame rate of 30 frames per second. Furthermore, from each still image, a total of 24 image data files of 256px × 256px (120mm × 120mm) were extracted from positions shifted by 5 pixels in the vertical and horizontal directions (up and down and left and right), resulting in image data where each extracted position was different.

[0099] In other words, a total of 25 image data frames were extracted from a single still image. This resulted in 3,750 (30 × 5 × 25) image data frames per sample, for a total of 33,750 (3,750 × 9) image data frames. These were then converted to grayscale using 256 gradations, where a brightness value of 0 is black and a brightness value of 255 is white. In other words, information about the brightness distribution of each pixel was obtained.

[0100] Of the 33,750 image data points obtained, 27,000 were used as training image data (dgi). The remaining 6,750 image data points were used as test data. This test data corresponds to the target image data (dA(t)).

[0101] The values ​​related to the test results performed on the cement mixture, as described above, correspond to the training performance data DCI. Specifically, the stacking height was used as the training performance data DCI. Then, based on the training image data DGI and the training performance data DCI, machine learning was performed using a CNN, a type of convolutional neural network. In detail, machine learning was performed using a 5-layer CNN with TensorFlow®. During machine learning, image features were extracted from the pooling layer. The number of training iterations was set to 10,000, and the number of image data (training image data DGI) input in each training iteration was 50 (randomly selected).

[0102] (Verification method) The test data (target image data dA(t)) was applied to the generated trained model M1. In all examples, the zero-pour mortar flow test was performed in accordance with JIS R 5201 and the zero-pour mortar flow was measured. Generally, if the zero-pour flow is large, the shape cannot be maintained and lamination becomes difficult, so the zero-pour flow is used as a guideline for the lamination height.

[0103] The results are shown in Table 1. Note that in Table 1, the stacking height is "predicted" using the trained model M1 for all examples.

[0104] [Table 1]

[0105] Generally, zero flow rate is used as an indicator of the stacking height, but as shown in Table 1, the stacking height differs even with the same zero flow rate (Examples 2 and 4), making it difficult to predict the stacking height from the zero flow rate.

[0106] Table 1 shows that the pre-trained model M1 can predict the layer height, one of the indicators of the 3D molding performance of cement mixtures, with high accuracy. However, in all examples, the zero-pour flow values ​​are similar to each other, but there can be significant differences in the measured layer height values. This suggests that there is not necessarily a high correlation between the flow value and the layer height depending on the cement mixture composition, or in other words, it is difficult to estimate the layer height with high accuracy using flow value information.

[0107] Regarding the stacking height, categories can be set for each range, and quality can be judged based on which category each item belongs to. Table 2 shows the results of quality judgment applied to three different samples using the trained model M1. Evaluation "A" indicates the best product, and the quality decreases in the order of "B" and "C".

[0108] [Table 2] [Explanation of Symbols]

[0109] 1: Quality prediction system 2: Target cement mixture 3: Imaging Unit 4: Mixing device 5: Formulation Information Input Section 6: Drive shaft 7: Agitator blade 8: Arm 10: Learning device 11: Input Reception Section 12: Learning Department 13: Storage section 20: Prediction device 21: Input Reception Section 22: Prediction Processing Unit 23: Output section 24: Storage section 25: Judgment section 41: Learning data input section Di: Training data M1: Pre-trained model M2: Pre-trained model PA: Imaging area da: Imaging data (target first input data) dbi: Training formula data dci: Training performance data do: Predictive information dp: Judgment result dβ: Formula data (Target second input data)

Claims

1. A storage unit records a trained model generated by machine learning based on multiple training data sets, which is formed by associating training image data generated based on a group of image data taken at different times during the mixing process of a training cement mixture for 3D modeling in a mixing device, with training performance data relating to an index value of the 3D modeling performance of the training cement mixture after the mixing process is completed. An input receiving unit that receives input of target first input data, which is a set of multiple image data itself, or a set of data generated from said image data, of the target cement mixture, which is a cement mixture for 3D modeling being mixed in a mixing device, at different time points, Based on the target first input data received by the input receiving unit, a prediction processing unit identifies target image data corresponding to the training image data and applies the target image data to the trained model to generate prediction information regarding the index value of the three-dimensional molding performance of the target cement mixture. A system for predicting the quality of a cement mixture, characterized by comprising an output unit that outputs the aforementioned prediction information.

2. The cement mixture quality prediction system according to claim 1, characterized in that the index value of the three-dimensional molding performance is a value corresponding to the stacking height until the learning cement mixture collapses when stacked.

3. The kneading apparatus comprises a kiln for containing the material to be kneaded, a stirring blade for stirring the material, and a drive shaft that provides rotational drive to the stirring blade. The cement mixture quality prediction system according to claim 1 or 2, characterized in that the target first input data is a group of image data obtained by imaging the target cement mixture located closer to the inner wall of the kiln than to the center of the kiln, or a group of data generated from said image data.

4. The kneading apparatus comprises a kiln for containing the material to be kneaded, a stirring blade for stirring the material, and a drive shaft that provides rotational drive to the stirring blade. The cement mixture quality prediction system according to claim 1 or 2, characterized in that the target first input data is a group of image data obtained by imaging the target cement mixture located within an area in which the stirring blade or the drive shaft is reflected, or a group of data generated from said image data.

5. The cement mixture quality prediction system according to claim 1 or 2, characterized in that the learning image data is information regarding the brightness distribution of each pixel contained in an image obtained by cropping imaging data captured in the mixing apparatus at a predetermined resolution.

6. A cement mixture quality prediction system according to claim 1 or 2, characterized in that it comprises a learning unit that generates a trained model by performing machine learning based on a plurality of training data in which the training image data and the training performance data are associated.

7. The aforementioned forecast information is composed of a combination of indicator values ​​for multiple items, The system includes a determination unit that determines the quality of the target cement mixture when used for three-dimensional molding, based on the aforementioned prediction information. The cement mixture quality prediction system according to claim 1 or 2, characterized in that the output unit outputs the determination result made by the determination unit.

Images