JP7742621B2 - AI-based method for estimating concrete fluidity - Google Patents

Description

The present invention relates to a method for estimating concrete slump or flumph, i.e., concrete fluidity.

Concrete is produced by mixing ordinary Portland cement, mixing water, fine aggregate, coarse aggregate, and optional admixtures in a batcher plant mixer. The produced concrete is transported by agitator truck and unloaded at the concrete pouring site before being poured. The required strength and properties of concrete vary depending on the concrete pouring site and are specified at the time of ordering. One indicator of concrete's properties is slump. Slump is an indicator of fluidity and is officially measured as follows: Concrete is filled into a 30 cm tall conical container (a slump cone), which is then inverted and gently removed upward. The concrete loses its shape and its top drops. The height of this drop is the slump; a higher value indicates greater fluidity, while a lower value indicates less fluidity. In practice, when mixing concrete in a batcher plant mixer, the slump is estimated using a slump monitor based on the current driving the mixer. Alternatively, the operator visually estimates the slump before shipping.

Patent No. 6680936 Japanese Patent Application Laid-Open No. 2003-177115

Patent Document 1 describes a method for estimating slump using machine learning. In this estimation method, concrete is first photographed inside a mixer while being mixed, and the resulting image data is input into a neural network, which then learns the slump at that time as training data. An image of the concrete whose slump is to be estimated is then input into the trained neural network, which then outputs an estimated slump value.

Patent Document 2 describes a method for estimating slump flow, an indicator of the fluidity of highly fluid concrete, known as high-fluidity concrete. While the testing method for slump flow is similar to that for slump, it is measured by removing a slump cone and measuring the diameter of the remaining concrete when it crumbles into a circular shape. Unlike slump, however, this method is suitable for measuring the fluidity of highly fluid concrete. In the method described in Patent Document 2, the sound generated during mixing of multiple high-fluidity concretes is acquired and spectrally analyzed. This means that the distribution of sound energy, i.e., sound pressure, which varies with frequency, is obtained. The slump flows of these multiple high-fluidity concretes are then compared with the spectrally analyzed sound pressure distribution, and frequencies highly correlated with slump flow are identified using regression analysis or similar techniques, resulting in the determination of an estimation formula, or model, for slump flow. Once the model is determined, the slump flow can be estimated for any high-fluidity concrete. Specifically, the sound is acquired when mixing the high-fluidity concrete. The acquired sound is then spectrally analyzed to obtain the sound pressure at that specific frequency. This is then input into the model, which outputs the slump flow. In other words, the slump flow is estimated.

Conventional slump estimation methods using slump monitors are also excellent because they can estimate slump with practical accuracy. However, while slump can be estimated with a certain degree of accuracy at the time of shipment, the properties of concrete can change during transport in an agitator truck. It is therefore impossible to know the slump at the time of unloading concrete. Therefore, when unloading, it is necessary to inspect the slump, for example, by conducting a slump test every 150 ^m3 . On the other hand, the method described in Patent Document 1 estimates slump from images of concrete. Therefore, by photographing the concrete when unloading it from the agitator truck or immediately before pouring, the slump can be estimated at that time. Around 2018, the inventors tested and verified a method similar to the method described in Patent Document 1, i.e., a method for estimating slump from concrete images. Specifically, multiple concretes with different slumps were mixed, and the concrete images photographed at that time were used as input data. The slump was then used as a training signal to train a neural network, and the trained neural network was used to estimate the slump. More specifically, the images were 8-bit grayscale, 360 pixels wide x 240 pixels high. Concrete was produced at two batcher plants, A and B. Images were taken inside the mixer at each plant, and images taken inside the loading hopper when loading onto an agitator truck. A total of 40,489 images were prepared as training images, and these were used for training. Meanwhile, the target concrete for estimation was a different concrete produced at the same batcher plants, A and B. Images taken inside the mixer and images taken inside the loading hopper when loading onto an agitator truck were used, and 78,376 images were prepared. When slump was estimated from these images, the deviation from the correct slump value was as shown in the table below.

This experiment confirmed that slump can be estimated with a certain degree of accuracy. Therefore, slump can be estimated by photographing concrete during unloading. However, there is room for improvement. Specifically, there is a demand for further improvement in the accuracy of the estimated slump, i.e., the accuracy of the estimation. When examining the deviation from the correct slump of the slump estimated from concrete images, the percentage within ±1.5 cm was relatively high at 87-92%, while the percentage within ±0.5 cm was slightly lower at 43-47%. Further improvement in estimation accuracy should make it possible to omit slump inspections during unloading. Furthermore, there are problems during the harsh winter. While widely practiced, particularly in cold regions such as Hokkaido, the aggregate is heated with steam and warm water is used for mixing to prevent the concrete from freezing. This practice causes steam to escape from the concrete, clouding the camera lens and preventing clear images. Or, even if an image is taken from a distance to prevent the lens from fogging up, a clear image cannot be obtained due to steam rising from the concrete. When the inventors conducted experiments in midwinter, they found that the accuracy of estimating slump from images was lower than usual, and there were cases where slump could not be estimated even within a range of ±2.5 cm. In other words, there appear to be problems with methods of estimating slump using only images of concrete.

It is possible to estimate concrete slump by referring to the method described in Patent Document 2. However, the method described in Patent Document 2 is a method for estimating slump flow, which is different from slump, and the concrete in question is high-fluidity concrete, which has significantly different properties from ordinary concrete. As a result, there is no guarantee that it can be applied to estimating the fluidity of ordinary concrete. Furthermore, even if it is used as a reference, the method described in Patent Document 2 requires processing by spectral analysis when estimating slump flow from the sound obtained when mixing concrete. In other words, preprocessing is essential. This means that estimation takes a certain amount of time and the work becomes complicated.

The present invention therefore aims to provide a method for estimating the fluidity of concrete that can estimate slump and slump flow, i.e., the fluidity of concrete, with high accuracy regardless of the season, without the need for complicated work, and that can estimate the fluidity of concrete not only at the batcher plant but also when unloading the concrete.

To achieve the above-mentioned objectives, the present invention is configured to estimate slump or slump flow, i.e., the fluidity of concrete, using a multi-layered neural network. Basic data and fluidity data consisting of slump or slump flow are obtained in advance for multiple concrete samples, and the basic data are provided as input data and the fluidity data as training data to the neural network for learning. Then, when basic data for concrete with unknown fluidity data is input to the trained neural network, the slump or slump flow can be obtained. In this invention, the basic data is time-series acoustic data sampled at a predetermined interval from the sound of concrete being mixed in a mixer or the sound of concrete being poured into a hopper.

That is, the invention described in claim 1 is a method for estimating the fluidity of concrete, which comprises obtaining basic data for a plurality of concretes and fluidity data consisting of slump, providing the basic data as input data and the fluidity data as teacher data to a multi-layered neural network to learn, and inputting the basic data into the neural network that has already learned about concrete whose fluidity data is unknown, thereby outputting fluidity data, characterized in that the basic data is time-series acoustic data on sound pressure obtained by sampling at a predetermined period the sound when mixing concrete in a mixer.
The invention described in claim 2 is configured as a method for estimating the fluidity of concrete, which is the estimation method described in claim 1, characterized in that the basic data includes image data of concrete in addition to the time-series acoustic data.

As described above, the present invention is configured as a concrete fluidity estimation method that obtains basic data and fluidity data consisting of slump flow for multiple concretes, provides the basic data as input data and the fluidity data as training data for a multilayered neural network, and trains the network using the basic data for concretes with unknown fluidity data, outputting fluidity data. In other words, this method estimates slump using AI that learns using so-called training data. In such AI learning, if there is a correlation between the input data and the output data, a nonlinear model of this relationship is automatically constructed within the neural network, making estimation with a certain degree of accuracy possible. However, whether or not high-accuracy estimation can be achieved depends on the selection of input data, i.e., whether input data that is highly correlated with the output data can be selected, and how the selected input data is processed before being provided to the AI. According to the present invention, the basic data is configured as time-series acoustic data of sound pressure sampled at a predetermined interval during concrete mixing in a mixer. As will be described in detail later with reference to an experiment, estimating slump fluidity data using time-series acoustic data during concrete mixing, etc., enables stable and highly accurate estimation. Acquiring time-series acoustic data requires only a microphone and is inexpensive. Furthermore, concrete fluidity data can be accurately estimated not only when concrete is produced in a batcher plant, but also at the concrete pouring site, i.e., when the concrete is being unloaded. This eliminates the need for inspections during unloading, reducing costs. Furthermore, because microphones are not affected by steam, it is guaranteed that concrete fluidity data can be estimated reliably even in the coldest of winters, just as it can be estimated in other seasons. According to another invention, the basic data includes image data of the concrete in addition to the time-series acoustic data. While time-series acoustic data has a high correlation with concrete fluidity data, concrete images also have a relatively high correlation with slump. Therefore, using concrete images in addition to time-series acoustic data as basic data enables even more accurate estimation of concrete fluidity data.

1 is a schematic diagram of a concrete fluidity data estimation system that implements a concrete fluidity data estimation method according to an embodiment of the present invention. FIG. 1 is a schematic diagram showing a neural network that realizes a method for estimating fluidity data of concrete according to an embodiment of the present invention. FIG. 10 is a schematic diagram of a concrete fluidity data estimation system that implements a concrete fluidity data estimation method according to a second embodiment of the present invention. FIG. 10 is a schematic diagram showing a neural network that realizes a method for estimating fluidity data of concrete according to a second embodiment of the present invention.

An embodiment of the present invention will now be described. The concrete fluidity data estimation method according to this embodiment is a method for estimating concrete slump or slump flow. It can be implemented in a conventional batcher plant 1, as shown in Figure 1, and at any concrete pouring site 2. Although only a portion of the batcher plant 1 is shown in the figure, it is composed of a cement bin for storing and dispensing ordinary Portland cement, an aggregate bin for storing and dispensing aggregate, multiple measuring tanks for measuring the materials dispensed from these bins, a measuring tank for dispensing mixing water and admixtures, and a mixer 4 for mixing the materials dispensed from these bins to produce concrete. The mixer 4 is equipped with a slump monitor 7 that measures the torque required to drive the mixer, i.e., the current consumed, to estimate the slump as the concrete is produced. This allows the slump to be obtained with sufficient accuracy for practical use. The concrete mixed in the mixer 4 is loaded onto an agitator truck 6 via a loading hopper 5. The loaded concrete is transported by an agitator truck 6 and unloaded at any concrete pouring site 2. For example, a concrete pump truck 8 is prepared at the concrete pouring site 2, and the concrete is unloaded into the hopper of this concrete pump truck 8. The concrete is poured using the concrete pump truck 8.

The concrete fluidity estimation method according to this embodiment estimates slump, an indicator of concrete fluidity, when concrete is produced in a batcher plant 1 or when concrete is unloaded at a concrete pouring site 2. The concrete fluidity estimation system 3 according to this embodiment comprises a computer 10, multiple microphones 11, 12, and 13, and a slump monitor 7. The first and second microphones 11 and 12 are installed in the batcher plant 1 and are connected to the computer 10 via a network, along with the slump monitor 7. The first and second microphones 11 and 12 are installed near the mixer 4 and the loading hopper 5, respectively. They are configured to pick up the sounds of concrete being mixed and concrete being poured into the loading hopper 5, respectively. The third microphone 13 is portable and is appropriately connected to the computer 10 via a wireless network connection or the like. The third microphone 13 is used to pick up the sounds of concrete being unloaded from the agitator truck 6. In the concrete fluidity estimation system 3 according to this embodiment, the computer 10 is trained using a neural network, which will be described below, and the fluidity of the concrete is estimated using the trained neural network. The estimated slump value obtained from the slump monitor 7 is only needed during the training phase, and once training is complete, the slump monitor 7 may be disconnected from the computer 10.

The computer 10 includes a multi-layered neural network 14, as shown in Figure 2. The neural network 14 is composed of an input layer 15 consisting of multiple neurons, an output layer 16 consisting of a predetermined number of neurons, and multiple intermediate layers 17 consisting of multiple neurons. However, in this embodiment, there is only one output layer 16. The neurons in each layer have an activation function such as a sigmoid function and a predetermined bias. They add the bias to input data, process it using the activation function, and output the data. Neurons in a given layer are connected to neurons in the adjacent layer by synapses, and the synapses consist of output weights, i.e., coefficients by which the output is multiplied. Therefore, the outputs of multiple neurons in a given layer are multiplied by synapses, i.e., coefficients, added together, and input to neurons in the next layer. In other words, when input data is given, it is processed by neurons in the input layer 15 and transmitted to the intermediate layer 17. This data is then processed by neurons in the intermediate layer 17 and transmitted to the output layer 16, where it is finally processed by neurons in the output layer 16 to obtain output data. Meanwhile, in the output layer 16, training data, which is the output data to be output, is provided and learning is performed using so-called backpropagation. In other words, the synaptic weights, i.e., coefficients, and the bias of each neuron are corrected. When specific input data is provided to a neural network 14 that has undergone sufficient learning, output data appropriate for that input data is output. So-called deep learning is a neural network that employs several techniques to dramatically improve the efficiency of learning using backpropagation, and the neural network 14 in this embodiment is also designed to learn with high efficiency.

In this embodiment, the input data provided to the neural network 14 is basic data related to concrete, and the output data is an estimated value of concrete fluidity. Specifically, the basic data is time-series acoustic data obtained by predetermined processing from the sounds of mixing, pouring, and unloading concrete, and slump is estimated as concrete fluidity. The time-series acoustic data is generated by the computer 10 as follows: Sound collected by any of the first to third microphones 11, 12, and 13 is sampled at a predetermined fixed interval over a predetermined period. For example, sampling is performed every 100 ms for 30 seconds. The sound pressure at the sampling timings is then arranged in time series, and this is used as the time-series acoustic data. When 30 seconds of sound are sampled every 100 ms, the time-series acoustic data is composed of a data string of 300 frequencies. In this case, the input layer 15 of the neural network 14 needs to have 300 neurons.

The neural network 14 in this embodiment is configured as a machine learning system that learns using so-called training data, and must first be trained. The input data and output data required for training, i.e., training data, are prepared as follows: Concrete is produced in the batcher plant 1. Time-series acoustic data is obtained from sounds collected by the first and second microphones 11 and 12, and an estimated slump value, i.e., an estimated value of concrete fluidity data, is obtained using the slump monitor 7. Multiple concretes with different slumps are produced, and time-series acoustic data and estimated slump values are obtained in the same way. Once many pairs of time-series acoustic data and estimated slump values have been obtained in this way, the former are provided as input data and the latter as training data to the neural network 14 for repeated training. Once sufficient training has occurred, the preparation stage is complete.

The trained neural network 14 can be used to estimate the slump of concrete with unknown slump. Sounds related to the concrete are collected by any of the first to third microphones 11, 12, and 13 and sent to the computer 10. Time-series acoustic data is obtained in the computer 10 and provided as input data to the neural network 14. The neural network 14 outputs the estimated slump value as output data. The trained neural network 14 can estimate the slump with high accuracy from the sounds of concrete mixing, pouring into a hopper, and unloading. Therefore, highly reliable slump values can be obtained not only at the batcher plant 1 but also at any concrete pouring site 2.

An experiment was conducted to confirm that the concrete fluidity estimation system 3 according to this embodiment can estimate slump with sufficiently high accuracy.
Experimental method: In a batcher plant 1, multiple lots of concrete with different slumps were produced, as shown in Table 2. Estimated slumps were obtained using a slump monitor 7, and the results were compiled for each slump. For example, 38 lots of concrete with a slump of 18.5 were produced. For each lot, the sounds of concrete being mixed were recorded using a first microphone 11 and saved. 10% of the recorded sounds were extracted to generate time-series acoustic data, which were then provided to a neural network 14 for training along with the estimated slump values. Next, the trained neural network 14, trained in this way, was used to estimate the slump. The time-series acoustic data used as input data was generated from the recorded data not used for training, i.e., from 90% of the recorded data.
Experimental results: When the trained neural network 14 was used to estimate the slump, the results shown in Table 3 were obtained.
Observations: It was confirmed that, when estimation was performed using the concrete fluidity estimation system 3 according to this embodiment, the percentage of slump that could be estimated with an error of ±0.5 cm exceeded 80%. Furthermore, it was confirmed that 45.8% of the slump could be estimated with no estimation error, that is, with an error of ±0.0 cm. It was confirmed that slump can be estimated with sufficient accuracy for practical purposes.

Figure 3 shows a concrete fluidity estimation system 3' according to a second embodiment of the present invention. The concrete fluidity estimation system 3' according to the second embodiment includes a computer 10, first, second, and third microphones 11, 12, and 13, a slump monitor 7, and first, second, and third cameras 21, 22, and 23. The first and second cameras 21 and 22 are connected to the computer 10, and the third camera 23 is connected as appropriate via a wireless network or the like. In the concrete fluidity estimation system 3' according to the second embodiment, the first, second, and third cameras 21, 22, and 23 are configured to capture images of the concrete during production, pouring, and unloading, respectively. The concrete fluidity estimation system 3' according to the second embodiment is configured to provide time-series acoustic data and concrete images as basic data related to the concrete, and estimate the slump based on these. Therefore, as shown in Figure 4, the neural network 14' used for slump estimation has neurons for concrete images in addition to neurons for time-series acoustic data in the input layer 15. The concrete fluidity estimation system 3' according to the second embodiment is configured to estimate slump by inputting concrete images in addition to time-series acoustic data, allowing for even more accurate slump estimation.

The concrete fluidity estimation system 3 according to this embodiment can be modified in various ways. For example, slump flow may be used as the fluidity of the concrete to be estimated. Slump flow can be obtained by actual measurement or other means and used as training data in place of slump, along with time-series acoustic data, to train the neural network 14. By providing time-series acoustic data, such as the mixing sound of concrete with unknown slump flow, to the trained neural network 14, the slump flow can be estimated. The slump provided as training data can also be modified. In this embodiment, the slump estimated by the slump monitor 7 is used. However, the slump may also be obtained experimentally using a slump cone and provided as training data. Slump may also be obtained by other means. While the time-series acoustic data is obtained from sounds collected by the first through third microphones 11, 12, and 13, it is also possible to install a microphone on the agitator truck 6 and obtain the sound of the drum rotating to agitate the concrete during transport. Any sound associated with concrete processing can be used. Furthermore, the method of processing the time-series acoustic data can also be modified. In the description of this embodiment, the time-series acoustic data is generated as frequency data for 30 seconds with a sampling period of 100 ms. Other periods and sampling periods may also be used. Furthermore, if only short-term audio data is available, this may be copied multiple times to generate a longer period, from which time-series acoustic data may be generated. While the slump estimation system 3 uses one computer 10, multiple computers may be used. For example, if the neural network 14 has been fully trained, a small computer with low processing power, such as a laptop computer, can be equipped with a replica of the neural network 14 to estimate slump. A laptop computer equipped with the trained neural network 14 and a third microphone 13 can be paired and carried to any concrete pouring site 2, enabling slump estimation without a wireless network connection.

The concrete fluidity estimation system 3' according to the second embodiment can also be modified. In the second embodiment, time-series acoustic data and image data are provided as input data to a single neural network 14'. However, for example, two neural networks could be prepared, one provided with time-series acoustic data and the other with image data, and each could estimate concrete fluidity data. These could then be combined using another AI to adopt one of the concrete fluidity data, or the average could be taken as the estimated concrete fluidity data. Furthermore, when using two neural networks, one neural network could be provided with time-series acoustic data to estimate concrete fluidity data. The obtained concrete fluidity data could be input, along with image data, to the other neural network to further estimate concrete fluidity data. In the concrete fluidity system 3' according to the second embodiment, additional basic data could be added. That is, data correlated with concrete fluidity could be added as input data. Examples of additional basic data include concrete mix information, mixing time, and mixing temperature, further improving the accuracy of fluidity estimation.

The concrete fluidity estimation system 3 according to this embodiment can be modified in other ways. In this embodiment, the data output by the neural network 14 is fluidity data, that is, only slump or slump flow. However, by adding neurons to the output layer, it is possible to predict other indicators related to concrete performance, such as the air content of concrete, compressive strength at 28 days, and mix proportions, as well as the proper operation of concrete manufacturing equipment. Naturally, these data are also required as training data, but this system is highly valuable as it can estimate many indicators related to concrete performance.

1. Batcher plant
2. Concrete pouring site 3. Concrete fluidity estimation system
4 Mixer 5 Loading hopper
6 Agitator truck 7 Slump monitor
8 Concrete pump truck 10 Calculator
11 First microphone 12 Second microphone
13 Third microphone 14 Neural network
15 Input layer 16 Output layer
17 Intermediate layer 21 First camera
22 Second camera 23 Third camera

Claims (2)

A method for estimating the fluidity of concrete, comprising the steps of: obtaining basic data for a plurality of concretes and fluidity data consisting of slump; providing the basic data as input data and the fluidity data as training data to a multi-layered neural network for learning; and inputting the basic data for concrete with unknown fluidity data into the trained neural network to output the fluidity data,
A method for estimating the fluidity of concrete, characterized in that the basic data is time-series acoustic data on sound pressure sampled at a predetermined period from the sound produced when mixing concrete in a mixer. The estimation method for estimating concrete fluidity according to claim 1, wherein the basic data includes image data of the concrete in addition to the time-series acoustic data.