Method for predicting quality of ready-mixed concrete, quality prediction program, and device for predicting quality of ready-mixed concrete

Abstract

To decrease variation in prediction accuracy resulting from bias in datasets for machine learning.SOLUTION: A method for predicting quality of ready-mixed concrete includes: an acquisition step for acquiring input information containing information related to an electric power load value of a mixer; a prediction step for predicting quality of ready-mixed concrete, on the basis of a prediction model constructed beforehand so as to output a quality value that indicates the quality of ready-mixed concrete according to an input of input information, and the input information acquired in the acquisition step; and a construction step for constructing a prediction model according to machine learning based on a plurality of datasets for learning. Each of the plurality of datasets for learning includes quality values, and input information associated with the quality values. In the construction step, before constructing the prediction model, the plurality of datasets for learning is corrected to reduce a difference in the number of data per quality value in a data frequency distribution representing a distribution of the number of data per quality value in the plurality of datasets for learning.SELECTED DRAWING: Figure 10

Description

[Technical Field]

[0001] The present disclosure relates to a method for predicting quality of ready-mixed concrete, a quality prediction program, and a quality prediction device for ready-mixed concrete. [Background technology]

[0002] Patent Document 1 discloses a method for predicting the quality of ready-mixed concrete using a prediction model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-124304 Summary of the Invention [Problem to be solved by the invention]

[0004] The present disclosure provides a ready-mixed concrete quality prediction method, a quality prediction program, and a ready-mixed concrete quality prediction device that are useful for reducing variability in prediction accuracy due to bias in machine learning datasets when making quality predictions using a prediction model. [Means for solving the problem]

[0005] [1] A method for predicting the quality of fresh concrete, comprising: an acquisition step of acquiring input information including information related to the power load value of a mixer that produces fresh concrete by mixing concrete materials or that stirs the fresh concrete; a prediction model that is constructed in advance to output a quality value that indicates the quality of the fresh concrete in response to input of the input information; a prediction step of predicting the quality of the fresh concrete based on the input information acquired in the acquisition step; and a construction step of constructing the prediction model by machine learning based on a plurality of training datasets, wherein each of the plurality of training datasets includes the quality value and the input information associated with that quality value, and the construction step includes correcting the plurality of training datasets so as to reduce the difference in the number of data points for each quality value in a data frequency distribution that represents the distribution of the number of data points for each quality value in the plurality of training datasets before constructing the prediction model.

[0006] [2] The method for predicting the quality of ready-mixed concrete described in [1] above, wherein, when correcting the plurality of learning data sets in the construction step, in areas of the data frequency distribution where the number of data is less than a predetermined number, a correction is made to increase the number of data to the predetermined number, and in areas of the data frequency distribution where the number of data is more than the predetermined number, no correction is made.

[0007] [3] The method for predicting the quality of ready-mixed concrete described in [1] above, wherein, when correcting the plurality of learning data sets in the construction step, in areas of the data frequency distribution where the number of data is greater than a predetermined number, a correction is made to reduce the number of data to the predetermined number, and in areas of the data frequency distribution where the number of data is less than the predetermined number, no correction is made.

[0008] [4] The method for predicting the quality of ready-mixed concrete described in [1] above, wherein, when correcting the plurality of learning datasets in the construction step, in areas of the data frequency distribution where the number of data is greater than a predetermined number, a correction is made to reduce the number of data to the predetermined number, and in areas of the data frequency distribution where the number of data is less than the predetermined number, a correction is made to increase the number of data to the predetermined number.

[0009] [5] A method for predicting the quality of ready-mixed concrete according to [1] above, wherein when correcting the plurality of learning data sets in the construction process, correction is performed so that the number of data for each quality value in the data frequency distribution becomes constant.

[0010] [6] The method for predicting the quality of ready-mixed concrete according to any one of [1] to [5] above, wherein the information relating to the power load value includes a statistical quantity obtained from time series data of the power load value when the mixer performs one mixing cycle, and the statistical quantity is at least one value selected from the group consisting of a fluctuation range representing the difference between the maximum value and the minimum value, a decline range representing the difference between the maximum value and the final value, a total value within an arbitrarily set time period, a mean value, a standard deviation, a coefficient of variation, a median, a first quartile, a third quartile, kurtosis, and skewness.

[0011] [7] The method for predicting the quality of ready-mixed concrete according to any one of the above [1] to [6], wherein the quality value is any one of slump, slump flow, and air content.

[0012] [8] A quality prediction program that causes a computer to execute the method for predicting the quality of ready-mixed concrete according to any one of [1] to [7] above.

[0013] [9] A ready-mixed concrete quality prediction device comprising: an input information acquisition unit that acquires input information including information related to the power load value of a mixer that produces ready-mixed concrete by mixing concrete materials or that mixes the ready-mixed concrete; a prediction model that is constructed in advance to output a quality value that indicates the quality of the ready-mixed concrete in response to input of the input information; a prediction calculation unit that predicts the quality of the ready-mixed concrete based on the input information acquired by the input information acquisition unit; and a model construction unit that constructs the prediction model by machine learning based on a plurality of training datasets, wherein each of the plurality of training datasets includes the quality value and input information associated with the quality value, and before constructing the prediction model, the model construction unit corrects the plurality of training datasets so as to reduce the difference in the number of data for each quality value in a data frequency distribution that represents the distribution of the number of data for each quality value in the plurality of training datasets. [Effects of the Invention]

[0014] According to the present disclosure, there are provided a ready-mixed concrete quality prediction method, a quality prediction program, and a ready-mixed concrete quality prediction device that are useful for reducing the variation in prediction accuracy caused by bias in machine learning datasets when making quality predictions using a prediction model. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a schematic diagram showing an example of a ready-mix concrete manufacturing system. [Figure 2] FIG. 2 is a block diagram illustrating an example of a functional configuration of the control device. [Figure 3] FIG. 3 is a diagram illustrating data relating to power load values. [Figure 4] Figures 4(a) and 4(b) are schematic diagrams showing an example of the calculation process using a prediction model, and Figure 4(c) is a schematic diagram showing an example of the calculation process using a neural network. [Figure 5] FIG. 5 is a block diagram illustrating an example of a hardware configuration of the control device. [Figure 6] 6(a) and 6(b) are graphs illustrating the relationship between the distribution of the training dataset and the prediction accuracy. [Figure 7] 7(a) is a diagram showing an example of a processing flow in the learning phase, and FIG. 7(b) is a diagram showing an example of a processing flow in the evaluation phase. [Figure 8] FIG. 8 is a graph illustrating a method for correcting a training data set. [Figure 9] FIG. 9 is a graph illustrating a method for correcting a training data set. [Figure 10] FIG. 10 is a graph illustrating a method for correcting a training data set. DETAILED DESCRIPTION OF THE INVENTION

[0016] An embodiment will be described below with reference to the drawings. In the description, identical elements or elements having identical functions are given the same reference numerals, and duplicated explanations will be omitted. Figure 1 shows a schematic diagram of a ready-mixed concrete manufacturing system equipped with a quality prediction device according to one embodiment.

[0017] The concrete materials used in the manufacturing system 1 include cement, admixtures, coarse aggregate, fine aggregate, water, and admixtures. Examples of coarse aggregate include gravel, crushed stone, slag coarse aggregate, lightweight coarse aggregate, recycled coarse aggregate, recovered aggregate, and coarse aggregates made from a mixture of these. Examples of gravel include mountain gravel, land gravel, river gravel, and sea gravel. Examples of slag coarse aggregate include blast furnace slag aggregate, ferronickel slag aggregate, electric arc furnace oxidizing slag aggregate, and coal gasification slag aggregate. Examples of lightweight coarse aggregate include natural lightweight aggregate, by-product lightweight aggregate, and artificial lightweight aggregate. Examples of coarse aggregate include crushed rock or crushed limestone.

[0018] Examples of fine aggregates include sand, crushed sand, slag fine aggregate, lightweight fine aggregate, recycled fine aggregate, recovered aggregate, and fine aggregates made from mixtures of these. Sand includes mountain sand, land sand, river sand, and sea sand. Slag fine aggregates include blast furnace slag aggregate, ferronickel slag aggregate, copper slag aggregate, electric furnace oxidizing slag aggregate, and coal gasification slag aggregate. Lightweight fine aggregates include natural lightweight aggregate, by-product lightweight aggregate, and artificial lightweight aggregate.

[0019] Examples of rock types for crushed stone and crushed sand include igneous rocks, sedimentary rocks, metamorphic rocks, quartzite, limestone, domalite, and peridotite. Igneous rocks include granite, diorite, gabbro, porphyrite, diabase, rhyolite, andesite, basalt, and serpentinite. Sedimentary rocks include conglomerate, sandstone, shale, slate, and tuff. Metamorphic rocks include gneiss and schist.

[0020] The manufacturing system 1 loads the manufactured ready-mixed concrete onto a transport vehicle 200. After the ready-mixed concrete has been loaded onto the transport vehicle 200, the transport vehicle 200 transports the ready-mixed concrete to the site where the ready-mixed concrete will be used (for example, a construction site). Examples of the transport vehicle 200 include an agitator vehicle (mixer vehicle) or a dump truck. The manufacturing system 1 may manufacture ready-mixed concrete from concrete materials so as to meet a target quality (required quality) set for each site. For example, an operator of the manufacturing system 1 determines the mix of concrete materials so as to meet the target quality set for each site, and inputs operating instructions to the manufacturing system 1.

[0021] The target quality set for each site is different from the quality of the ready-mixed concrete produced by the production system 1 before it is shipped (at the time of shipment). For example, in the production system 1, ready-mixed concrete is produced by the production system 1, and the quality of the ready-mixed concrete is controlled (inspected, etc.) before it is shipped so that the target quality of the ready-mixed concrete when it is used, set for each site, is met. The time when the ready-mixed concrete is used corresponds to the time when the ready-mixed concrete is received at the site. In order to control the quality of the ready-mixed concrete before shipping, the target quality of the ready-mixed concrete when it is shipped may be determined based on the target quality of the ready-mixed concrete when it is used. The target quality of the ready-mixed concrete when it is shipped may be set according to in-house standards, etc., set for each factory. In setting the target quality of the ready-mixed concrete when it is shipped, at least one of the following information may be taken into consideration: the condition (quality) of the materials used during production, the season (temperature), the type of concrete, the target quality when it is used, and the transportation time. The target quality of the ready-mixed concrete when it is shipped is set, for example, by adding a value determined by the in-house standards to the target quality of the ready-mixed concrete when it is used.

[0022] The manufacturing system 1 includes, for example, a manufacturing apparatus 100 and a control device 10. The manufacturing apparatus 100 is an apparatus that manufactures ready-mixed concrete based on operation instructions from the control device 10. The manufacturing apparatus 100 manufactures ready-mixed concrete by mixing concrete materials. The manufacturing apparatus 100 includes, for example, a material storage area 101, a transport device 104, a storage bottle 111, a measuring bottle 112, a collection hopper 113, a mixer 114, and a loading hopper 115.

[0023] The material storage yard 101 is a place where concrete materials are stored. The material storage yard 101 includes a plurality of silos 102. The plurality of silos 102 are containers that store at least a portion of the concrete materials by material type. The plurality of silos 102 include, for example, a silo 102 that stores coarse aggregate, a silo 102 that stores fine aggregate, and a silo 102 that stores cement.

[0024] The transporting device 104 is a device that transports concrete materials stored in the multiple silos 102 to the storage bins 111. The transporting device 104 includes, for example, a belt conveyor that transports the concrete materials. The transporting device 104 may transport the concrete materials by type at different times. In one example, based on an operation instruction from the control device 10, a specific material from among the various concrete materials is transferred to the transporting device 104 and transported to the storage bins 111.

[0025] The storage bottles 111 temporarily store various types of concrete materials. The various types of concrete materials are transported (conveyed) to the storage bottles 111 from the material storage area 101 by the transport device 104. The storage bottles 111 are configured to individually store various types of concrete materials. Hereinafter, "concrete materials" may be simply referred to as "materials." The various materials stored in the storage bottles 111 are supplied to the measuring bottles 112 as needed.

[0026] The measuring bottle 112 is disposed below the storage bottle 111. The measuring bottle 112 operates based on operational instructions from the control device 10, and individually measures various materials. When the measuring bottle 112 detects the target amount of material instructed by the control device 10, it supplies the material to the collecting hopper 113. When water is supplied to the measuring bottle 112, an admixture may be mixed into the water. The collecting hopper 113 is disposed below the measuring bottle 112. The collecting hopper 113 collects the various materials discharged from the measuring bottle 112 and supplies the collected various materials to the mixer 114. Note that the manufacturing apparatus 100 does not necessarily have to be equipped with the collecting hopper 113, and the various materials may be supplied to the mixer 114 from the measuring bottle 112.

[0027] The mixer 114 is disposed below the collecting hopper 113. The mixer 114 is a device that mixes concrete materials. The mixer 114 produces ready-mixed concrete by mixing (kneading) aggregate, cement, water, admixtures, etc. In other words, the mixer 114 produces ready-mixed concrete by mixing the concrete materials. The ready-mixed concrete is discharged from the bottom of the mixer 114 into the loading hopper 115. The mixer 114 may be a tilting mixer, a horizontal single-shaft mixer, a horizontal twin-shaft mixer, or a pan-type mixer. The mixer 114 includes, for example, two stirring members 114a and a mixer drive unit 114b.

[0028] The agitating members 114a are members that agitate the various materials supplied to the mixer 114. The two agitating members 114a are arranged side by side inside the main body (container portion) of the mixer 114 and are rotatable. Each of the two agitating members 114a includes a rotation shaft that extends horizontally in one direction. The mixer driving unit 114b rotates the rotation shaft of each of the two agitating members 114a based on an operation instruction from the control device 10. The mixer driving unit 114b includes, for example, a driving source such as a motor that applies driving force to the agitating members 114a. An opening and closing port is provided at the bottom of the main body of the mixer 114 for discharging the produced ready-mixed concrete into the loading hopper 115.

[0029] The loading hopper 115 is disposed below the mixer 114 and temporarily stores the ready-mixed concrete. The loading hopper 115 supplies the temporarily stored ready-mixed concrete to the transport vehicle 200.

[0030] The manufacturing apparatus 100 described above is an example of a ready-mixed concrete manufacturing apparatus, and the ready-mixed concrete manufacturing apparatus may be configured in any way as long as it is capable of mixing concrete materials using a mixer and manufacturing ready-mixed concrete.

[0031] In this disclosure, a unit of ready-mixed concrete produced by one mixing in the mixer 114 and loaded onto the transport vehicle 200 is defined as "one batch." The process executed by the manufacturing system 1 for producing one batch of ready-mixed concrete is defined as "batch processing." In one example, one to three batches of ready-mixed concrete are loaded onto one transport vehicle 200. For example, when two batches of ready-mixed concrete are loaded onto one transport vehicle 200, two batch processes according to the same manufacturing conditions are performed at different times (in different orders).

[0032] <Control device (quality prediction device)> The control device 10 is a device that controls the manufacturing equipment 100. The control device 10 is configured with one or more computers. When the control device 10 is configured with multiple computers, these computers are connected to each other so that they can communicate with each other. The control device 10 controls the manufacturing equipment 100 in accordance with set operating conditions. At least some of the operating conditions may be determined by instructions from an operator such as a worker.

[0033] An input device 12 and a monitor 14 may be connected to the control device 10. The input device 12 is a device that inputs information indicating instructions from a worker or the like to the control device 10. The input device 12 may be any device that can input desired information, and may be a keyboard (keypad), an operation panel, or a mouse. The monitor 14 is a device that displays information from the control device 10 to a worker or the like. The monitor 14 may be any device that can display graphics, and may be a liquid crystal display. The input device 12 and the monitor 14 may be integrated, such as a touch panel. The control device 10, the input device 12, and the monitor 14 may be integrated, such as a tablet computer (tablet terminal).

[0034] In addition to controlling the manufacturing apparatus 100, the control device 10 may have a function of predicting the quality of ready-mixed concrete manufactured by the manufacturing apparatus 100. In this case, the control device 10 constitutes a quality prediction device that predicts the quality of ready-mixed concrete (ready-mixed concrete quality prediction device). In the following description, the quality of ready-mixed concrete predicted by the control device 10 is the quality after it has been manufactured by the manufacturing apparatus 100 and before it is shipped to the site (i.e., the quality of the ready-mixed concrete at the time of shipment).

[0035] The quality of fresh concrete to be predicted by the control device 10 may include one or more qualities of slump, slump flow, and air content. The quality of fresh concrete to be predicted by the control device 10 may include two or more qualities of slump, slump flow, and air content. The quality of fresh concrete to be predicted by the control device 10 may be one quality of slump, slump flow, and air content, or may be two or more qualities of slump, slump flow, and air content.

[0036] The control device 10 is configured to execute at least an acquisition process, a prediction process, and a construction process. The acquisition process is a process of acquiring input information including information related to the power load value of the mixer 114. The prediction process is a process of predicting the quality of the ready-mixed concrete to be predicted based on the input information acquired in the acquisition process and a prediction model that has been constructed in advance by machine learning so as to output a quality value indicating the quality of ready-mixed concrete in response to the input of the input information. The construction process is a process of constructing the prediction model by machine learning based on multiple learning datasets. The construction process is executed before the prediction process and the acquisition process.

[0037] 2 shows an example of functional components (hereinafter referred to as "functional blocks") included in the control device 10. The control device 10 has, for example, as functional blocks, an operation control unit 22, an input information acquisition unit 24, a model construction unit 30, a model holding unit 32, a prediction calculation unit 28, and a display output unit 34. The processing executed by these functional blocks corresponds to the processing executed by the control device 10.

[0038] The operation control unit 22 controls the manufacturing apparatus 100 to manufacture ready-mixed concrete in accordance with predetermined operating conditions. At least some of the operating conditions may be determined by an operator, such as a worker, each time ready-mixed concrete is manufactured. The operation control unit 22 may control the mixer driving unit 114b of the mixer 114 so that the rotation speed of the mixer driving unit 114b follows a target rotation speed defined in the operating conditions. When controlling the mixer driving unit 114b, the operation control unit 22 may adjust the power (e.g., current value) supplied to the mixer driving unit 114b. If the ready-mixed concrete to be manufactured is hard, the power load value tends to be large, and if the ready-mixed concrete to be manufactured is soft, the power load value tends to be small.

[0039] The input information acquisition unit 24 acquires input information for a prediction model for predicting quality. The input information acquired by the input information acquisition unit 24 includes information related to the power load value of the mixer 114. The power load value of the mixer 114 may be a value indicating the power (kW) itself supplied to the mixer 114, or a value indicating the current value (A) supplied to the mixer 114. Alternatively, the power load value of the mixer 114 may be replaced with a value indicating the load hydraulic pressure (MPa). The information related to the power load value of the mixer 114 may be continuous time-series data obtained while the mixer 114 is operating during the processing of one batch, or may be statistical data obtained from the time-series data. The information related to the power load value of the mixer 114 may be a moving average value of the time-series data, or may be statistical data obtained from the moving average value.

[0040] FIG. 3 schematically illustrates time-series data relating to the power load values ​​of the mixer 114 in processing one batch. The time-series data relating to the power load values ​​is, for example, data obtained by repeatedly measuring the power (kW) supplied to the mixer 114 at a predetermined sampling period. The input information acquisition unit 24 may calculate (acquire) as statistical data at least one selected from the group consisting of a fluctuation range (difference between the maximum value and the minimum value), a decline range (difference between the maximum value and the final value), a total value within an arbitrarily set time period, an average value, a standard deviation, a coefficient of variation, a median, a first quartile, a third quartile, kurtosis, and skewness. The input information acquisition unit 24 may calculate at least one of a minimum value and a maximum value as statistical data in addition to at least one selected from the above group. The statistical data acquired by the input information acquisition unit 24 may include statistical values ​​obtained by calculation using two or more power load values ​​included in the time-series data.

[0041] In the graph shown in FIG. 3, "P1" represents the initial value in the time-series data. The initial value P1 is the power load value at the start of mixing in the time-series data. "P2" represents the minimum value in the time-series data. The minimum value P2 is the minimum power load value after the time when the initial value P1 is obtained. Note that the power load value at the start of mixing (initial value P1) may also be the minimum. "P3" represents the maximum value in the time-series data. The maximum value P3 is the maximum power load value after the time when the initial value P1 is obtained. "P4" represents the final value in the time-series data. The final value P4 is the power load value at the time when the operation control unit 22 determines that the condition for terminating mixing by the mixer 114 is met. For example, when a predetermined time has elapsed since the start of driving the agitating member 114a, the operation control unit 22 determines that the above condition is met and stops driving the agitating member 114a.

[0042] The model construction unit 30 constructs a model (hereinafter referred to as "prediction model M") for predicting the quality of ready-mixed concrete. The prediction model M is a model that outputs a quality value indicating the quality of ready-mixed concrete in response to the input of the input information, including information related to the power load value of the mixer 114. The model construction unit 30 constructs the prediction model M through machine learning based on the input information and the correct quality values ​​associated with the input information. The prediction model M may be constructed to output one or more predicted values ​​of slump, slump flow, and air content as quality values ​​indicating the quality of ready-mixed concrete. The prediction model M may be constructed to output a predicted value of the ratio of slump flow to slump (slump flow / slump) as a quality value indicating the quality of ready-mixed concrete, instead of or in addition to the one or more predicted values. In the following description, unless otherwise specified, "quality value" refers to a quality value indicating the quality of ready-mixed concrete, and "input information" refers to input information including information related to the power load value of the mixer 114.

[0043] Machine learning is a technique in which a machine (computer) autonomously finds laws or rules by repeatedly learning based on given information. A prediction model M can be constructed using an algorithm and a data structure. A prediction model M is realized, for example, using a neural network, which is an information processing model that mimics the mechanism of the human brain and nerves. There are no particular limitations on the specific algorithm of machine learning used when constructing a prediction model M. A neural network has, for example, an input layer, one or more intermediate layers, and an output layer. By including one or more intermediate layers, a more complex prediction model M can be constructed, thereby improving prediction accuracy.

[0044] The model construction unit 30 may autonomously construct a prediction model M for predicting the quality of ready-mixed concrete by performing machine learning using data provided as input for machine learning and correct data (correct values ​​of slump, etc.) that are the output of the machine learning. The input for the machine learning is various data sets of input information, including information related to the power load value of the mixer 114. The output of the machine learning is data (numerical values) that indicate the quality of ready-mixed concrete. The model construction unit 30 iteratively learns a model that outputs predicted values ​​of slump, etc., using multiple combinations of data sets of input information and correct values ​​of slump, etc. The stage in which the prediction model M is autonomously constructed corresponds to the learning phase. The learning phase may be performed before the production phase in which ready-mixed concrete is manufactured, or may be performed in the early stages of the production phase.

[0045] In the learning phase, the model construction unit 30 constructs a prediction model M by machine learning based on multiple training datasets (hereinafter referred to as "multiple training datasets TD"). Each of the multiple training datasets TD (each training dataset TD) includes a quality value and input information associated with the quality value. The quality value included in each training dataset TD is correct data, and is obtained by actual measurement by, for example, an operator. Before constructing the prediction model M, the model construction unit 30 corrects the multiple training datasets TD. Then, the model construction unit 30 performs machine learning using the corrected multiple training datasets TD to construct the prediction model M. The correction of the multiple training datasets TD will be described later.

[0046] 4(a) and 4(b) each schematically show a prediction model M constructed by the model construction unit 30. The prediction model M shown in FIG. 4(a) is a model that outputs a single quality prediction value in response to input information. The prediction model M shown in FIG. 4(b) is a model that outputs two or more types of quality prediction values ​​in response to input information. In the present disclosure, machine learning performed so that the prediction model M outputs a single quality prediction value (a prediction value of only one type of quality) is referred to as "single-task learning." Furthermore, machine learning performed so that the prediction model M outputs two or more types of quality prediction values ​​is referred to as "multi-task learning."

[0047] The model construction unit 30 may construct the prediction model M by performing single-task learning. In this case, the model construction unit 30 may construct a prediction model M for each type of quality. For example, the model construction unit 30 constructs two or more models from among a prediction model M that outputs a predicted value of slump, a prediction model M that outputs a predicted value of slump flow, and a prediction model M that outputs a predicted value of air volume. The model construction unit 30 may construct the prediction model M by performing multi-task learning. For example, the model construction unit 30 constructs a prediction model M (one prediction model M) that outputs a predicted value for each of two or more qualities of slump, slump flow, and air volume. The model construction unit 30 may construct a prediction model M (one prediction model M) that outputs a ratio of slump flow to slump in addition to one or more qualities of slump, slump flow, and air volume.

[0048] An example of the prediction model M will be described below using simplified formulas for ease of understanding. The prediction model M constructed by the model construction unit 30 can be simply expressed, for example, as in the following formulas (1) and (2).

number

number

[0049] In equation (2), Y represents the output value of quality, and in a prediction model M constructed by single-task learning, it is the output value of a single quality. In a prediction model M constructed by multi-task learning, Y is the output value of one of two or more types of quality, and equations (1) and (2) are calculated for each type of quality. N is an integer of 2 or more and represents the number of input data. x represents various input values ​​included in the input data, and the input data corresponds to the above input information, and the input data includes at least values ​​related to power load values ​​(e.g., 2 or more statistics).

[0050] wi is a weight (coefficient), and b is a bias term (coefficient). f(U) represents the activation function. The activation function can be a linear function (identity function) or a nonlinear function such as a polynomial, absolute value, step function, sigmoid function, hardsigmoid function, logsigmoid function, softmax function, logsoftmax function, softmin function, softplus function, softsign function, tanh function, tanhShrink function, hardtanh function, tanhexp function, ReLU function, ReLU6 function, Leaky-ReLU function, PReLU function, ELU function, SELU function, CELU function, Swith function, Mish function, or ACON function.

[0051] The model construction unit 30 may repeatedly evaluate the error and loss value between Y (predicted value) obtained by Equation (2) and the correct quality value using multiple corrected training datasets TD, and determine the weights wi and bias term b in Equation (1) so as to minimize the error (loss value). The model construction unit 30 may use any type of loss function as a function for evaluating the error between Y (output value from an intermediate model at an intermediate stage in constructing the prediction model M) obtained by Equation (2) and the correct quality value. The role of the loss function is to input the predicted value and the correct value into the loss function and output a loss value L(a) based on the error a between the predicted value and the correct value. The weights wi are then calculated using the loss value L(a). The model construction unit 30 may use, for example, one loss function selected from the group consisting of the Huber loss function (HuberLoss), the mean absolute error (MAE), and the ε-insensitive loss function (ε-insensitiveloss). From the viewpoint of improving prediction accuracy, it is preferable to use a loss function that ensures that the loss value L(a) is always equal to or less than the absolute value of the error a. The mean absolute error or the ε-allowable loss function may be used as a function that ensures that the loss value L(a) is always equal to or less than the absolute value of the error a. Alternatively, by setting δ of the Huber loss function shown in the following formula (3) to 1.0 or less, it is also possible to use a function that ensures that the loss value L(a) is always equal to or less than the absolute value of the error a.

number

[0052] The model construction unit 30 may repeatedly update the weights wi using a gradient method so as to minimize the error and loss value evaluated by the loss function. The model construction unit 30 may use any type of update formula (weight update formula) when updating the weights wi. For example, the model construction unit 30 uses one selected from the group consisting of Adam, AdamBelief, Adamax, AdaBound, Adagrad, AMSGRAD, AMSBound, RMSprop, SgdW, Momentum, and Nesterov as the weight update formula. The weight update formula is also referred to as an optimization algorithm or an optimization method.

[0053] The model construction unit 30 may construct the prediction model M by machine learning using a neural network that does not perform an operation to normalize the input information and an operation to delete some connections in the fully connected layer. The operation to normalize the input information refers to an operation (batch normalization) that normalizes (standardizes) each value included in the input information that is input to the prediction model M so that the average is 0 and the standard deviation is 1.

[0054] The operation of dropping some connections in the fully connected layer means that, when performing machine learning using a neural network, learning is performed while deactivating an arbitrary proportion of nodes (dropout). Figure 4(c) shows a schematic diagram of the calculation process of the prediction model M when some connections in the fully connected layer are dropped (when dropout is performed). Note that the model construction unit 30 may construct the prediction model M by machine learning using a neural network that performs at least one of an operation to normalize input information and an operation to drop some connections in the fully connected layer.

[0055] The model storage unit 32 stores the prediction model M constructed by the model construction unit 30. The prediction model M, which is a trained model, may be transferable between computers. Therefore, the prediction model M constructed in the control device 10 may be used in a manufacturing system other than the manufacturing system 1.

[0056] In the evaluation phase, the prediction calculation unit 28 predicts the quality of the ready-mixed concrete to be predicted based on the input information acquired by the input information acquisition unit 24 and the prediction model M. The prediction calculation unit 28 inputs the acquired input information into the prediction model M and acquires a predicted value output from the prediction model M. The input information acquired in the evaluation phase (input information for evaluation) is information whose quality value is unknown.

[0057] The display output unit 34 outputs the quality value predicted by the prediction calculation unit 28 to the monitor 14. As a result, the predicted quality value is displayed on the monitor 14, and an operator of the manufacturing system 1 or the like can grasp the predicted value of the quality of the ready-mixed concrete that is the prediction target (manufactured ready-mixed concrete).

[0058] As shown in Fig. 5, the control device 10 includes a circuit 50. The circuit 50 includes a processor 51, a memory 52, a storage 53, and an input / output port 54. The storage 53 is configured with one or more non-volatile memory devices such as a flash memory or a hard disk. The storage 53 stores a quality prediction program that causes a computer to execute at least the acquisition process, the prediction process, and the construction process. The storage 53 stores a quality prediction program for configuring each functional block of the control device 10.

[0059] The memory 52 is composed of one or more volatile memory devices such as a random access memory. The memory 52 temporarily stores a quality prediction program loaded from the storage 53. The processor 51 is composed of one or more arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The processor 51 configures each functional block of the control device 10 by executing the quality prediction program loaded into the memory 52. ​​The calculation results by the processor 51 are temporarily stored in the memory 52. ​​The input / output port 54 inputs and outputs information to and from the input device 12, the monitor 14, the mixer 114, etc. in response to a request from the processor 51.

[0060] The timer 55 measures the elapsed time by, for example, counting reference pulses at a fixed interval. The circuit 50 is not necessarily limited to one in which each function is configured by a program. For example, the circuit 50 may have at least some of its functions configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates such dedicated logic circuits. The quality prediction program may be provided by being permanently recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Alternatively, the quality prediction program may be provided via a communication network as a data signal superimposed on a carrier wave.

[0061] [Ready-mix concrete manufacturing method] Next, an example of a method for producing ready-mixed concrete executed in the production system 1 will be described. The method for producing ready-mixed concrete includes a production process and a quality prediction process. The production process is a process for producing ready-mixed concrete. The quality prediction process is a process for predicting the quality of the ready-mixed concrete produced in the production process. The quality prediction process may be executed during a period that overlaps with at least a portion of the period during which the production process is repeatedly executed.

[0062] The manufacturing process includes, for example, a transporting process, a weighing process, a feeding process, a mixing process, a discharging process, and a loading process. In the transporting process, various types of concrete materials are transported to storage bottles 111 by a transporting device 104, and the various materials are individually supplied to the storage bottles 111. In the weighing process, the various materials are individually supplied from the storage bottles 111 to measuring bottles 112, and the various materials are weighed in the measuring bottles 112. In the weighing process, when the measured amount of each material reaches a predetermined set amount, the material is discharged into a collecting hopper 113. In the feeding process, after all types of materials have been collected in the collecting hopper 113, the materials in the collecting hopper 113 are fed (supplied) into a mixer 114.

[0063] In the mixing process, multiple types of concrete materials are mixed in the mixer 114. In the mixing process, the control device 10 may control the mixer driving unit 114b in accordance with predetermined operating conditions. In the mixing process, the power supplied from the control device 10 to the mixer driving unit 114b may be adjusted so that the rotation speed of the mixer driving unit 114b follows a target rotation speed.

[0064] In the discharging process, after mixing of the concrete materials in the mixer 114 is completed, the ready-mixed concrete is discharged from the mixer 114 into the loading hopper 115. In the loading process, the ready-mixed concrete discharged into the loading hopper 115 is loaded onto the transport vehicle 200.

[0065] The quality prediction process (a method for predicting the quality of ready-mixed concrete) includes a model construction process in a learning phase and a quality evaluation process in an evaluation phase. In the quality prediction process, the model construction process is executed before the quality evaluation process. Below, the model construction process and the quality evaluation process will be explained using an example in which a prediction model M is constructed by single-task learning and is constructed so as to output a predicted value of slump as a quality value.

[0066] (Model building process) First, to facilitate understanding of the correction of the multiple training datasets TD, the issues that arise when the multiple training datasets TD are not corrected will be described with reference to Figures 6(a) and 6(b). To distinguish between the training datasets before and after correction, the multiple training datasets TD before correction will be referred to as "multiple training datasets TD0", and the multiple training datasets TD after correction will be referred to as "multiple training datasets TD1". The multiple training datasets TD collectively refer to the multiple training datasets TD0 and the multiple training datasets TD1.

[0067] The graphs shown in Figures 6(a) and 6(b) show "data frequency distributions" for multiple training datasets TD0. The data frequency distribution represents the distribution of the number of data items for each quality value in the multiple training datasets TD. The number of data items for each quality value represents the number of datasets for each quality value among the multiple training datasets TD. For example, if the number of datasets in which the actual measured value (correct value) of the slump is 20 cm is 120 among the multiple training datasets TD, then the number of data items with a slump of 20 cm in the data frequency distribution will be 120. Note that in the example shown in Figures 6(a) and 6(b), the slump is measured in increments of 0.5 cm.

[0068] When preparing a machine learning dataset by operating the actual manufacturing apparatus 100, a bias in the number of data points for each measured slump value may occur in the data frequency distribution. That is, the data frequency distribution may have regions with a large number of data points and regions with a small number of data points. To confirm the impact of the bias in the number of data points in the data frequency distribution, the inventors used an evaluation dataset to verify the prediction accuracy of a prediction model (hereinafter referred to as "prediction model M0") constructed by machine learning based on multiple training datasets TD0 for each quality value. The prediction model M0 is a model constructed in the same way as the prediction model M, except for the training data used. The evaluation dataset, like the training dataset TD, is a dataset in which input information and measured quality values ​​(slump) are associated.

[0069] Each of Figures 6(a) and 6(b) shows the accuracy rate for each quality value as the verification result. In the verification for each quality value, the predicted value by the prediction model M0 is compared with the correct value in the evaluation dataset. The accuracy rate is the proportion of datasets determined to be correct, where the prediction result by the prediction model M0 is defined as being correct when it is within a range obtained by adding a predetermined tolerance to the correct value in the evaluation dataset. The evaluation dataset is divided according to the correct value of the quality value, and the accuracy rate is calculated for each quality value.

[0070] Figure 6(a) shows the accuracy rate for each quality value when the tolerance is set to ±1.0 cm, and Figure 6(b) shows the accuracy rate for each quality value when the tolerance is set to ±2.0 cm. The verification results shown in Figures 6(a) and 6(b) show that the accuracy rate tends to be lower in areas with fewer data points in the data frequency distribution than in areas with more data points. In order to reduce such differences in accuracy rate, i.e., the variability in prediction accuracy, corrections are made to multiple training datasets TD0.

[0071] 7(a) is a flowchart showing an example of a series of processes executed in the model construction process. This model construction process is executed before the manufacturing process is executed in the manufacturing apparatus 100, or at an early stage after the manufacturing process has started. In this model construction process, for example, ready-mixed concrete actually manufactured in the manufacturing apparatus 100 is used as ready-mixed concrete for learning.

[0072] In the model construction process, step S11 is first executed. In step S11, for example, a worker or the like prepares a plurality of training data sets TD0 as training data for machine learning. Each of the plurality of training data sets TD0 includes input information (training input information) obtained when training ready-mix concrete is produced, and a correct slump value associated with the input information. The plurality of training data sets TD0 may be input to the control device 10 via the input device 12 by a worker or the like.

[0073] The correct value of the slump may be a value obtained by actually measuring the quality of the ready-mixed concrete for training. In one example, after the ready-mixed concrete for training is loaded onto the transport vehicle 200, a portion of the ready-mixed concrete is extracted by a worker or the like. The worker or the like then measures the slump of the extracted ready-mixed concrete, and this measured value is used as the correct value in the multiple training data sets TD0.

[0074] Next, step S12 is executed. In step S12, for example, the model construction unit 30 of the control device 10 generates a plurality of training data sets TD1 (a plurality of training data sets TD after correction) by correcting the plurality of training data sets TD0 prepared in step S11. The model construction unit 30 may perform correction on the plurality of training data sets TD0 in accordance with a predetermined correction procedure, and in the correction procedure, some correction conditions may be determined by an operator or the like.

[0075] The model construction unit 30 corrects the multiple training data sets TD0 so as to reduce the difference in the number of data points for each measured value of the slump (each quality value) in the data frequency distribution. In this case, the difference between the minimum and maximum values ​​of the number of data points in the corrected data frequency distribution becomes smaller than the difference between the minimum and maximum values ​​of the number of data points in the data frequency distribution before correction. For example, the model construction unit 30 corrects the multiple training data sets TD0 so that the number of data points for each measured value of the slump becomes constant in the corrected data frequency distribution. In this case, the number of data points becomes the same for all measured values ​​of the slump in the corrected data frequency distribution. Below, a specific example of a correction method for reducing the bias in the number of data points in the data frequency distribution will be described.

[0076] <Correction method (i): Oversampling> When correcting the multiple training datasets TD, the model construction unit 30 performs a correction to increase the number of data to the predetermined number AN in regions of the data frequency distribution where the number of data is less than the predetermined number AN, and does not perform a correction in regions of the data frequency distribution where the number of data is greater than the predetermined number AN. This correction method is referred to as "oversampling" in the present disclosure. The predetermined number AN is determined in advance when the model construction unit 30 performs the correction and may be specified, for example, by a worker such as an operator. Instead of being specified by a worker, the model construction unit 30 may autonomously set the predetermined number AN from the data frequency distribution related to the multiple training datasets TD0. The predetermined number AN is set to a value between the minimum and maximum number of data in the data frequency distribution.

[0077] FIG. 8 shows the data frequency distribution for multiple training datasets TD0 and the data frequency distribution after correction by oversampling. "No adjustment" indicates that the data is before correction. In the example shown in FIG. 8, the predetermined number AN is set to 75. For actual slump measurements with fewer than 75 data sets, correction is performed so that the number of training datasets is 75. For actual slump measurements with 75 or more data sets, no correction is performed. By performing oversampling, the minimum number of data sets in the data frequency distribution for multiple training datasets TD1 becomes 75 (predetermined number AN). This reduces the difference between the maximum and minimum number of data sets before and after correction.

[0078] When increasing the training dataset at a certain quality value, the model construction unit 30 may randomly select a training dataset included before correction and duplicate the data of the training dataset to create a new training dataset. Instead of duplicating the data, the model construction unit 30 may use various data augmentation techniques to create a new training dataset. Examples of such data augmentation techniques include Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), Borderline SMOTE, and Safe-level SMOTE.

[0079] <Correction method (ii): Undersampling> When correcting the multiple training data sets TD, the model construction unit 30 performs a correction to reduce the number of data to the predetermined number AN in areas of the data frequency distribution where the number of data is greater than a predetermined number AN, and does not perform a correction in areas of the data frequency distribution where the number of data is less than the predetermined number AN. This correction method is referred to as "Undersampling" in the present disclosure. The predetermined number AN may be set in the same way as in the correction using Oversampling.

[0080] FIG. 9 shows the data frequency distribution for multiple training datasets TD0 and the data frequency distribution after correction by undersampling. In the example shown in FIG. 9, the predetermined number AN is set to 75. For actual slump measurements with more than 75 data points, the number of training datasets is corrected to 75. For actual slump measurements with 75 or fewer data points, no correction is performed. By performing undersampling, the maximum number of data points in the data frequency distribution for multiple training datasets TD1 becomes 75 (predetermined number AN). This reduces the difference between the maximum and minimum number of data points before and after correction.

[0081] When reducing the number of training data sets at a certain quality value, the model construction unit 30 may use any method to reduce (exclude) the training data sets. For example, the model construction unit 30 may exclude, from the training data, one or more training data sets that are randomly selected from the multiple training data sets included before correction.

[0082] <Correction method (iii): Oversampling + Undersampling> The model construction unit 30 performs a correction that combines the above-mentioned oversampling and undersampling. When correcting multiple training data sets TD, in areas where the number of data in the data frequency distribution is greater than a predetermined number AN, a correction is performed to reduce the number of data to the predetermined number AN, and in areas where the number of data in the data frequency distribution is less than the predetermined number AN, a correction is performed to increase the number of data to the predetermined number AN. The predetermined number AN may be set in the same way as in the correction using oversampling.

[0083] FIG. 10 shows the data frequency distributions for multiple training datasets TD0 and the data frequency distributions after correction using a correction method that combines oversampling and undersampling. In the example shown in FIG. 10, the predetermined number AN is set to 200. For actual slump measurements with fewer than 200 data points, the number of training datasets is corrected to 200. For actual slump measurements with more than 200 data points, the number of training datasets is also corrected to 200. By performing correction using a combination of oversampling and undersampling, the minimum and maximum numbers of data in the data frequency distributions for multiple training datasets TD1 are both 200 (predetermined number AN). This reduces the difference between the maximum and minimum numbers of data before and after correction, and furthermore, the number of data points for each actual slump measurement is constant.

[0084] 7(a), after execution of step S12, step S13 is executed. In step S13, for example, the model construction unit 30 constructs a prediction model M by performing machine learning using the multiple training datasets TD1 (multiple training datasets after correction) generated in step S12. The model construction unit 30 may construct the prediction model M by machine learning using a neural network.

[0085] Next, step S14 is executed. In step S14, for example, the model holding unit 32 stores the prediction model M constructed in step S13. This completes the model construction process.

[0086] (Quality evaluation process) 7(b) is a flowchart showing an example of a series of processes executed in the quality evaluation process. This quality evaluation process is executed, for example, during a period overlapping at least a part of the period during which the manufacturing process is executed by the manufacturing apparatus 100.

[0087] In the quality evaluation process, first, the control device 10 executes step S21. In step S21, for example, the control device 10 waits until the evaluation timing arrives, which is the timing for evaluating the quality of the ready-mixed concrete to be evaluated. The ready-mixed concrete to be evaluated is also the ready-mixed concrete to be manufactured by the manufacturing apparatus 100. The evaluation timing may be predetermined to a certain time period in a day, or may be predetermined to the timing of executing a certain number of batch processes in a day. The evaluation timing may also be the timing when an instruction to perform the evaluation is received from a worker such as an operator.

[0088] Next, the control device 10 executes step S22. In step S22, for example, the input information acquisition unit 24 acquires input information including information related to the power load value when the ready-mixed concrete to be evaluated was produced. The input information acquired in step S22 is input information for evaluation in which the slump is unknown.

[0089] Next, the control device 10 executes step S23. In step S23, for example, the prediction calculation unit 28 predicts quality information of the ready-mixed concrete to be evaluated based on the input information acquired in step S22 and the prediction model M held in the model holding unit 32. In one example, the prediction calculation unit 28 inputs the input information acquired in step S22 into the prediction model M and acquires a predicted value of slump output from the prediction model M.

[0090] Next, the control device 10 executes step S24. In step S24, for example, the display output unit 34 displays the predicted value of the step acquired in step S23 on the monitor 14. This allows a worker such as an operator to check the predicted value of quality.

[0091] This completes the quality evaluation process. The control device 10 may execute the series of processes of steps S21 to S24 each time one batch of ready-mixed concrete is produced (for each batch process). The control device 10 may execute the series of processes of steps S21 to S24 each time multiple batches of ready-mixed concrete are produced (for each multiple batch processes). When another quality value such as slump flow is predicted instead of or in addition to slump, the series of processes of steps S11 to S14 and the series of processes of steps S21 to S24 may be executed for the other quality value.

[0092] [Variations] The series of processes shown in Figures 7(a) and 7(b) are examples and can be modified as appropriate. In the series of processes, one step and the next step may be executed in parallel, or some steps may be executed in an order different from that of the example described above. Steps different from those of the example described above may be executed instead of or in addition to at least some of the steps of the series of processes described above.

[0093] The input information acquired by the control device 10 (input information acquisition unit 24) and used as an input to the prediction model M may include at least one of the initial value, minimum value, maximum value, and final value in the time-series data related to the power load value. The input information used as an input to the prediction model M may also include information related to mixing by the mixer 114 (information other than the power load value). Examples of the information related to mixing include the amount of concrete mixed for one batch, the mixing time, the time at which the power load value reaches its maximum value, and the time from when the power load value reaches its maximum value until the fresh concrete is discharged from the mixer 114. The input information used as an input to the prediction model M may also include the nominal strength, information indicating the type of cement, information indicating the type of admixture, information indicating the type of fine aggregate, information indicating the type of coarse aggregate, and information indicating the amount of additive added.

[0094] The input information to be input to the prediction model M may include information relating to images of ready-mixed concrete during or immediately after production in the mixer 114, or images of ready-mixed concrete after it has been discharged from the mixer 114. The input information to be input to the prediction model M may include at least one of the target quality of the ready-mixed concrete when it is used and the target quality of the ready-mixed concrete when it is shipped. The input information to be input to the prediction model M needs to include at least information relating to the power load value of the mixer, and may include other types of information instead of or in addition to the various types of information exemplified above.

[0095] The model construction unit 30 may correct the multiple training data sets TD using a method other than the above-mentioned oversampling, undersampling, and a combination of oversampling and undersampling. The model construction unit 30 may perform a correction to align the number of data to a predetermined number by increasing (or reducing) the training data set for each of all quality values ​​in the data frequency distribution. The model construction unit 30 may increase the training data set in a region in the data frequency distribution where the number of data is less than a first predetermined number so that the number of data becomes the first predetermined number, and may reduce the training data set in a region in the data frequency distribution where the number of data is more than a second predetermined number that is greater than the first predetermined number so that the number of data becomes the second predetermined number.

[0096] In addition to the prediction of the quality of ready-mixed concrete by the control device 10, the quality of ready-mixed concrete may be measured periodically (for example, several times a day) in the manufacturing apparatus 100. In this case, the prediction model M may be updated based on the actual measurement value of the quality of the ready-mixed concrete and the input information when the actual measurement value was obtained. In the prediction step, the quality of the ready-mixed concrete may be predicted using the updated prediction model M. Note that even when the updated prediction model M is used, the step of predicting the quality of the ready-mixed concrete is still performed based on the prediction model M and the input information acquired in the acquisition step.

[0097] The manufacturing system 1 may include, as a functional block, a quality prediction device (ready-mixed concrete quality prediction device) having at least an input information acquisition unit 24, a model construction unit 30, a model storage unit 32, and a prediction calculation unit 28, separate from the control device 10. The computer constituting the quality prediction device may be communicably connected to the control device 10. Hereinafter, the quality prediction device having at least the input information acquisition unit 24, the model construction unit 30, the model storage unit 32, and the prediction calculation unit 28 will be simply referred to as the "quality prediction device."

[0098] In the above example, the control device 10, functioning as a quality prediction device, predicts the quality of ready-mixed concrete after it has been produced by the production apparatus 100 and before it is shipped to the construction site. The timing at which the quality prediction device predicts the quality is not limited to this example. The quality prediction device may also predict the quality of ready-mixed concrete while it is being transported to the construction site where the ready-mixed concrete will be used (including upon arrival at the construction site). For example, the transport vehicle 200 is provided with a mixer 204 that mixes the ready-mixed concrete (see FIG. 1), and the input information acquisition unit 24 of the quality prediction device acquires information related to the power load value of the mixer 204 as at least part of the input information. Note that the mixer 204 in the transport vehicle 200, such as an agitator vehicle, is also referred to as a drum. As described above, at least part of the input information may be information related to the production of ready-mixed concrete by mixing concrete materials, or the power load value of the mixer (114, 204) that mixes the ready-mixed concrete after production.

[0099] In the above example, ready-mixed concrete is manufactured at a location (manufacturing apparatus 100) separate from the construction site. The location where ready-mixed concrete is manufactured is not limited to this example. Concrete materials (for example, concrete materials excluding water) may be transported to the construction site by a transport vehicle, and ready-mixed concrete may be manufactured at the construction site. At that time, water may be added to the concrete materials transported by the transport vehicle using a mixer provided on the transport vehicle, and the materials may be mixed to manufacture ready-mixed concrete. When ready-mixed concrete is manufactured at the construction site, the quality prediction device may construct a prediction model M at the construction site and predict the quality of the ready-mixed concrete. The input information acquisition unit 24 of the quality prediction device may acquire, as at least part of the input information, information related to the power load value of the mixer provided on the transport vehicle when ready-mixed concrete is manufactured at the construction site.

[0100] When comparing the magnitude of two numerical values ​​within a computer, either of the two criteria "greater than or equal to" and "greater than" may be used, or either of the two criteria "less than or equal to" and "under." The selection of such criteria does not change the technical significance of the process of comparing the magnitude of two numerical values. In one example of the various examples described above, at least some of the matters described in the other examples may be combined.

[0101] [Verification of prediction accuracy using a prediction model] Next, we will explain the results of verifying the slump prediction accuracy of a prediction model M constructed after correcting multiple training datasets TD using multiple evaluation datasets for which correct values ​​are known. In the evaluation dataset, as in the multiple training datasets TD, a set of various input values ​​included in the input information is associated with the correct value (actual measured value) of the slump. Statistics obtained from time-series data related to power load values ​​are used as part of the input information in the multiple training datasets TD and the evaluation dataset.

[0102] (Comparative Example) To verify the effect of correction on the multiple training datasets TD, machine learning was performed using multiple training datasets TD0 (2765 training datasets), which are the multiple training datasets TD before correction, to construct a prediction model for comparison. The multiple training datasets TD0 have a data frequency distribution indicated by "No adjustment" in Figures 8 to 10.

[0103] (Verification example 1) In Verification Example 1, a prediction model M was constructed after performing correction on multiple training datasets TD0 (2765 training datasets) by oversampling. The predetermined number AN was set to 75. The data frequency distribution after correction is shown in Figure 8.

[0104] (Verification example 2) In Verification Example 2, correction was performed on multiple training datasets TD0 (2765 training datasets) using Undersampling, and then a prediction model M was constructed. The predetermined number AN was set to 75. The data frequency distribution after correction is shown in Figure 9.

[0105] (Verification example 3) In Verification Example 3, a prediction model M was constructed after performing correction on multiple training datasets TD0 (2765 training datasets) using a combination of undersampling and oversampling. The predetermined number AN was set to 200. The data frequency distribution after correction is shown in FIG. 10.

[0106] (Verification results) The following absolute difference average, root mean square error (RMSE), and accuracy rate were calculated as evaluation indicators. The smaller the absolute difference average and RMSE, the smaller the error and the higher the prediction accuracy. The higher the accuracy rate, the smaller the error and the higher the prediction accuracy. Average absolute difference: The absolute value of the difference between the correct value and the value predicted by the prediction model is calculated for each evaluation dataset, and the sum of these absolute values ​​is divided by the number of evaluation datasets. RMSE: The square of the error between the correct value and the value predicted by the prediction model is calculated for each evaluation dataset, and the sum of these squared errors is divided by the number of evaluation datasets, and then the square root is calculated. Accuracy rate: The percentage of datasets for which the predictions made by the predictive model are within a given tolerance, relative to the total number of evaluation datasets.

[0107] Seven hundred eighty-nine evaluation datasets were prepared, and the three evaluation indices described above were calculated for each of the Comparative Example and Verification Examples 1 to 3. The 789 evaluation datasets were prepared so that the range of the measured slump values ​​was roughly the same as the range in the multiple training datasets TD0 (2,765 training datasets). To confirm the influence of correction, the range of the measured slump values ​​was set to three ranges: "full range," "18 cm or less," and "more than 18 cm," and three evaluation indices were calculated. The evaluation datasets were divided into the "18 cm or less" and "more than 18 cm" ranges according to the measured slump values ​​(correct values) in the evaluation datasets. The tolerances were set to ±0.5 cm, ±1.0 cm, ±1.5 cm, ±2.0 cm, and ±2.5 cm.

[0108] The calculation results of the index for the entire range of measured slump values ​​from 10.5 cm to 25.0 cm are shown in Table 1 below. Note that in the data frequency distribution for multiple learning datasets TD0, the range of measured slump values ​​is 10 cm to 25.5 cm, but since the evaluation dataset does not include a dataset with a measured slump value of 10 cm or 25.5 cm, the entire range (the range to be evaluated) is 10.5 cm to 25.0 cm. [Table 1]

[0109] The calculation results of the index for the range where the measured slump value is 18 cm or less are shown in Table 2 below. The range where the measured slump value is 18 cm or less includes 74 evaluation datasets out of 789 evaluation datasets. [Table 2]

[0110] The calculation results of the index for the range where the actual slump value is over 18 cm (18.5 cm or more) are shown in Table 3 below. The range where the actual slump value is over 18 cm includes 715 of the 789 evaluation datasets. [Table 3]

[0111] The verification results shown in Tables 2 and 3 were calculated by roughly dividing the entire range of measured slump values ​​in half and calculating the indices for each half. The results shown in Tables 2 and 3 indicate that correcting the training data to reduce the difference between each slump before constructing the prediction model M tends to reduce the difference in the evaluation indices between the range of 18 cm or less and the range of over 18 cm. For example, in the comparative example, the difference in RMSE between the range of 18 cm or less and the range of over 18 cm is 1.19. In contrast, the corresponding difference in Verification Example 1 is 0.63, the corresponding difference in Verification Example 2 is 1.00, and the corresponding difference in Verification Example 3 is 0.50. In other words, it can be seen that the difference between the ranges in RMSE is reduced in Verification Examples 1 to 3 compared to the comparative example.

[0112] Summary of this disclosure The above-described method for predicting the quality of ready-mixed concrete includes: an acquisition step of acquiring input information including information related to the power load value of a mixer (114, 204) that produces ready-mixed concrete by mixing concrete materials or mixes the ready-mixed concrete; a prediction step of predicting the quality of the ready-mixed concrete based on the input information acquired in the acquisition step by using a prediction model (M) constructed in advance to output a quality value indicating the quality of the ready-mixed concrete in response to input of the input information; and a construction step of constructing the prediction model (M) by machine learning based on multiple training datasets (TD). Each of the multiple training datasets (TD) includes a quality value and input information associated with the quality value. The construction step includes correcting the multiple training datasets (TD) before constructing the prediction model (M) to reduce differences in the number of data points for each quality value in a data frequency distribution that represents the distribution of the number of data points for each quality value in the multiple training datasets (TD).

[0113] It has been found that if there is a bias in the number of data points per quality value in a machine learning dataset used to build a model for predicting the quality of ready-mixed concrete, the accuracy will be relatively low in areas of quality values ​​where the number of data points is small. In the above method, multiple training datasets (TD) are corrected to reduce the difference in the number of data points per quality value, and then a prediction model (M) is built. This reduces the difference in the accuracy of predictions made by the prediction model (M) between areas of quality values ​​with a small number of data points and other areas. Therefore, this method is useful for reducing the variation in prediction accuracy caused by bias in machine learning datasets.

[0114] In the quality prediction method for ready-mixed concrete described above, when correcting multiple training datasets (TD) in the construction process, correction may be performed to increase the number of data to the predetermined number (AN) in areas of the data frequency distribution where the number of data is less than a predetermined number (AN), and correction may not be performed in areas of the data frequency distribution where the number of data is more than the predetermined number (AN). In this case, the difference in the number of data between areas in the data frequency distribution before correction where the number of data is less than the predetermined number (AN) and areas where the number of data is more than the predetermined number (AN) is reduced by the correction. Therefore, the difference in prediction accuracy by the prediction model (M) can be reduced between areas where the number of data is less than the predetermined number (AN) and areas where the number of data is more than the predetermined number (AN).

[0115] In the quality prediction method for ready-mixed concrete described above, when correcting multiple training datasets (TD) in the construction process, correction may be performed to reduce the number of data to the predetermined number (AN) in areas of the data frequency distribution where the number of data is greater than a predetermined number (AN), and correction may not be performed in areas of the data frequency distribution where the number of data is less than the predetermined number (AN). In this case, the difference in the number of data between areas where the number of data is greater than the predetermined number (AN) and areas where the number of data is less than the predetermined number (AN) in the data frequency distribution before correction is reduced by the correction. Therefore, the difference in prediction accuracy by the prediction model (M) can be reduced between areas where the number of data is greater than the predetermined number (AN) and areas where the number of data is less than the predetermined number (AN).

[0116] In the quality prediction method for ready-mixed concrete described above, when correcting multiple training datasets (TD) in the construction process, a correction may be performed to reduce the number of data to the predetermined number (AN) in regions of the data frequency distribution where the number of data is greater than the predetermined number (AN), or a correction may be performed to increase the number of data to the predetermined number (AN) in regions of the data frequency distribution where the number of data is less than the predetermined number (AN). In this case, the difference in the number of data between regions where the number of data is greater than the predetermined number (AN) and regions where the number of data is less than the predetermined number (AN) in the data frequency distribution before correction is further reduced by the correction. Therefore, the difference in prediction accuracy by the prediction model (M) can be further reduced between regions where the number of data is greater than the predetermined number (AN) and regions where the number of data is less than the predetermined number (AN).

[0117] In the quality prediction method for ready-mixed concrete described above, when correcting multiple training datasets (TD) in the construction process, the correction may be performed so that the number of data for each quality value is constant in the data frequency distribution. In this case, there is no bias in the number of data for each quality value in the multiple training datasets (TD1) used for machine learning. This makes it possible to avoid differences in prediction accuracy due to bias in the number of data for each quality value.

[0118] In the above-described method for predicting the quality of ready-mixed concrete, the information related to the power load value may include statistics obtained from the time-series data of the power load value when the mixer (114, 204) performs one mixing cycle. The statistics may be at least one value selected from the group consisting of a fluctuation range representing the difference between the maximum and minimum values, a decline range representing the difference between the maximum and final value, a total value within an arbitrarily set time period, a mean value, a standard deviation, a coefficient of variation, a median, a first quartile, a third quartile, kurtosis, and skewness. In this case, using the statistics related to the power load value can reduce the influence of disturbances that may be contained in each value in the time-series data of the power load value. This is therefore useful for improving prediction accuracy.

[0119] In the above-described method for predicting the quality of ready-mixed concrete, the quality value may be any one of slump, slump flow, and air content. The quality of ready-mixed concrete is often managed by at least one of slump, slump flow, and air content. In the above-described method, any one of these quality values ​​is predicted by the prediction model (M). Furthermore, since prediction accuracy is maintained over a wide range of quality values, predictions by the prediction model (M) can be used to manage the production of many types of ready-mixed concrete. Therefore, this method is useful for simplifying the work of managing the quality of the produced ready-mixed concrete.

[0120] The quality prediction method program described above is a program that causes a computer to execute the above-mentioned ready-mix concrete quality prediction method. This quality prediction program can execute the above-mentioned quality prediction method, and is therefore useful for reducing variations in prediction accuracy due to bias in machine learning datasets.

[0121] The above-described ready-mixed concrete quality prediction device (10) includes an input information acquisition unit (24) that acquires input information including information related to the power load value of a mixer (114, 204) that produces ready-mixed concrete by mixing concrete materials or that mixes the ready-mixed concrete; a prediction model (M) that is pre-constructed to output a quality value indicating the quality of the ready-mixed concrete in response to input of the input information; a prediction calculation unit (28) that predicts the quality of the ready-mixed concrete based on the input information acquired by the input information acquisition unit (24); and a model construction unit (30) that constructs the prediction model (M) by machine learning based on multiple training data sets (TD). Each of the multiple training data sets (TD) includes a quality value and input information associated with the quality value. Before constructing the prediction model (M), the model construction unit (30) corrects the multiple training data sets (TD) to reduce differences in the number of data points for each quality value in a data frequency distribution that represents the distribution of the number of data points for each quality value in the multiple training data sets (TD). Similar to the above-described quality prediction method, this quality prediction device (10) is useful for reducing variations in prediction accuracy due to bias in the data set used for machine learning. [Explanation of symbols]

[0122] 1... manufacturing system, 10... control device, 24... input information acquisition unit, 28... prediction calculation unit, 30... model construction unit, AN... predetermined number, M... prediction model, 100... manufacturing apparatus, 114... mixer, 114a... stirring member, 114b... mixer drive unit, 204... mixer.

Claims

1. an acquisition step of acquiring input information including information related to a power load value of a mixer that produces ready-mixed concrete by mixing concrete materials or mixes the ready-mixed concrete; a prediction step of predicting the quality of the ready-mixed concrete based on a prediction model constructed in advance to output a quality value indicating the quality of the ready-mixed concrete in response to input of the input information and the input information acquired in the acquisition step; A construction process of constructing the predictive model by machine learning based on a plurality of learning datasets, each of the plurality of learning datasets includes the quality value and the input information associated with the quality value; the construction step includes, before constructing the prediction model, correcting the plurality of training data sets so as to reduce a difference in the number of data sets for each quality value in a data frequency distribution representing a distribution of the number of data sets for each quality value in the plurality of training data sets. Methods for predicting the quality of ready-mix concrete.

2. When correcting the plurality of training data sets in the construction step, In an area in the data frequency distribution where the number of data is less than a predetermined number, a correction is performed to increase the number of data to the predetermined number; No correction is performed in an area in the data frequency distribution where the number of data items is greater than the predetermined number. The method for predicting the quality of ready-mixed concrete according to claim 1.

3. When correcting the plurality of training data sets in the construction step, In an area in the data frequency distribution where the number of data is greater than a predetermined number, a correction is performed to reduce the number of data to the predetermined number; No correction is performed in an area in the data frequency distribution where the number of data is less than the predetermined number. The method for predicting the quality of ready-mixed concrete according to claim 1.

4. When correcting the plurality of training data sets in the construction step, In an area in the data frequency distribution where the number of data is greater than a predetermined number, a correction is performed to reduce the number of data to the predetermined number; In an area in the data frequency distribution where the number of data is less than the predetermined number, a correction is performed to increase the number of data to the predetermined number. The method for predicting the quality of ready-mixed concrete according to claim 1.

5. When correcting the plurality of learning data sets in the construction step, the correction is performed so that the number of data for each quality value in the data frequency distribution becomes constant. The method for predicting the quality of ready-mixed concrete according to claim 1.

6. the information relating to the power load value includes a statistic obtained from time series data of the power load value when the mixer performs one mixing operation, The statistical quantity is at least one value selected from the group consisting of a fluctuation range representing the difference between a maximum value and a minimum value, a decline range representing the difference between a maximum value and a final value, a total value within an arbitrarily set time period, an average value, a standard deviation, a coefficient of variation, a median, a first quartile, a third quartile, kurtosis, and skewness. The method for predicting the quality of ready-mixed concrete according to any one of claims 1 to 5.

7. The quality value is any one of slump, slump flow, and air content. The method for predicting the quality of ready-mixed concrete according to any one of claims 1 to 5.

8. A quality prediction program that causes a computer to execute the ready-mixed concrete quality prediction method according to any one of claims 1 to 5.

9. an input information acquisition unit that acquires input information including information related to a power load value of a mixer that produces ready-mixed concrete by mixing concrete materials or mixes the ready-mixed concrete; a prediction calculation unit that predicts the quality of the ready-mixed concrete based on a prediction model that is constructed in advance to output a quality value that indicates the quality of the ready-mixed concrete in response to input of the input information, and the input information acquired by the input information acquisition unit; a model construction unit that constructs the prediction model by machine learning based on a plurality of learning datasets; each of the plurality of learning datasets includes the quality value and input information associated with the quality value; the model construction unit corrects the plurality of training data sets so as to reduce a difference in the number of data items for each quality value in a data frequency distribution representing a distribution of the number of data items for each quality value in the plurality of training data sets before constructing the prediction model; Ready-mix concrete quality prediction device.

Images