JP7271295B2 - Machine learning method, information processing device, and program - Google Patents

Description

TECHNICAL FIELD Embodiments of the present invention relate to a machine learning method, an information processing apparatus, and a program.

For example, machine learning techniques are applied in the fields of image recognition, speech recognition, and natural language processing. In machine learning, for example, the accuracy of the calculation result is improved by repeatedly comparing the calculation result from the learning data and the label corresponding to the learning data.

JP 2018-1184 A

Generally, as the number of iterations increases, the accuracy of the computation results improves. The usefulness of machine learning increases when the number of iterations to achieve the desired accuracy can be reduced.

A machine learning method according to one embodiment is a machine learning method executed by a computer, comprising: causing a calculation unit to calculate output information from each of a plurality of pieces of input information; determining whether the output information is correct or incorrect based on correct information set for the input information; calculating the output information from each of at least one first input information corresponding to the output information and at least one second input information included in the plurality of input information; and changing the ratio of the first input information and the second input information input to the calculation unit based on the ratio of the output information determined to be correct. And prepare.

FIG. 1 is an exemplary perspective view schematically showing the layered manufacturing system according to the first embodiment. FIG. 2 is an exemplary cross-sectional view schematically showing a part of the laminate manufacturing system and the modeled object of the first embodiment. FIG. 3 is an exemplary block diagram functionally showing the configuration of the laminate manufacturing system of the first embodiment. FIG. 4 is a diagram showing three examples of images of the first embodiment. FIG. 5 is an exemplary diagram schematically showing an example of the configuration of the CNN of the first embodiment. FIG. 6 is an exemplary block diagram functionally showing the configuration of the learning unit of the first embodiment. FIG. 7 is a flow chart showing an example of the machine learning method of the first embodiment. FIG. 8 is a graph showing an example of correct answer rate of output values in the evaluation results of the first evaluation unit and the second evaluation unit according to the first embodiment. FIG. 9 is a graph showing an example of correct answer rates of output values in the evaluation results of the first evaluation section and the second evaluation section as a comparative example. FIG. 10 is an exemplary block diagram functionally showing the configuration of an estimating unit according to the first embodiment. FIG. 11 is a flow chart showing an example of the layered manufacturing monitoring method of the first embodiment. FIG. 12 is a flow chart showing an example of the layered manufacturing monitoring method according to the second embodiment.

(First embodiment)
A first embodiment will be described below with reference to FIGS. 1 to 11. FIG. In this specification, basically, the vertically upward direction is defined as the upward direction, and the vertically downward direction is defined as the downward direction. In addition, in this specification, a component according to the embodiment and description of the component may be described with a plurality of expressions. The components and their descriptions are examples and are not limited by the expressions herein. Components may be identified by names different from those herein. Also, components may be described in terms that differ from those herein.

FIG. 1 is an exemplary perspective view schematically showing a layered manufacturing system 1 according to the first embodiment. The additive manufacturing system 1 may also be referred to as an additive manufacturing monitoring system. The laminate manufacturing system 1 is a system including a so-called laser material deposition three-dimensional printer. Note that the layered manufacturing system 1 is not limited to this example.

An X-axis, a Y-axis and a Z-axis are defined herein as indicated in each drawing. The X-axis, Y-axis and Z-axis are orthogonal to each other. The Z-axis extends, for example, in the vertical direction. The X-axis and Y-axis extend horizontally, for example. Note that the layered manufacturing system 1 may be arranged so that the Z-axis obliquely crosses the vertical direction.

FIG. 2 is an exemplary cross-sectional view schematically showing part of the laminate manufacturing system 1 and the modeled object 3 of the first embodiment. The layered manufacturing system 1 laminates (additionally manufactures) a modeled object 3 having a predetermined shape, for example, by stacking powdery materials M in layers. The modeled object 3 is an example of an object to be laminate-molded.

As shown in FIG. 1 , the layered manufacturing system 1 has a table 11 , a modeling section 12 , an imaging section 13 , a control section 14 and a plurality of signal lines 16 . The control unit 14 may also be called an information processing device or computer. The table 11, the modeling unit 12, the imaging unit 13, and the control unit 14 are arranged, for example, in a chamber C provided inside the housing of the layered manufacturing system 1 or inside a room for modeling. Chamber C is, for example, a space that can be sealed.

The table 11 has a support surface 11a. The support surface 11a is formed substantially flat and faces the positive direction of the Z-axis (the direction indicated by the arrow of the Z-axis, upward). The support surface 11a supports the modeled object 3 that has been laminate-modeled, the in-process product of the modeled object 3, and the base for laminating the material M. In the following description, the modeled object 3 includes the modeled object 3 for which layered manufacturing has been completed, the in-process product of the modeled object 3, and the base. By rotating at least part of the table 11, the table 11 can rotate the modeled object 3 supported on the support surface 11a around the center of rotation parallel to the Z-axis.

The table 11 may move the modeled object 3 in the X-axis direction, the Y-axis direction, and the Z-axis direction. Further, the table 11 may further rotate the modeled object 3 around a rotation center parallel to the Y-axis or a rotation center parallel to the X-axis.

The modeling unit 12 supplies the material M and stacks it on the support surface 11a or a base supported by the support surface 11a. Material M is, for example, powdered titanium. Note that the material M is not limited to this, and may be other materials such as other metals, synthetic resins, and ceramics. The layered manufacturing system 1 may laminately manufacture the modeled object 3 using a plurality of types of materials M. FIG.

The modeling section 12 has a nozzle 21 , a supply device 22 , a moving device 23 and an adjusting device 24 . The nozzle 21 ejects the material M onto the support surface 11a of the table 11 or onto the modeled object 3 on the support surface 11a. Further, as shown in FIG. 2, energy rays E are emitted from the nozzle 21 to the ejected material M and the modeled object 3 on the support surface 11a. The energy ray E is, for example, laser light.

A laser beam as the energy beam E is emitted from the nozzle 21 in parallel with the material M being supplied. Other energy beams E may be emitted from the nozzle 21 in addition to the laser beam. The energy beam E may be any beam that can melt or sinter the material M, such as a laser beam, and may be, for example, an electron beam or an electromagnetic wave in the microwave or ultraviolet region.

The modeling unit 12 heats the base and the discharged material M with the energy beam E to form a melted region (bead) 3a. The nozzle 21 irradiates the material M with the energy beam E to melt or sinter the material M in the melting region 3a, thereby gathering the material M. As shown in FIG. Thus, the melting region 3a can include not only the supplied material M, but also a part of the base and the modeled object 3 irradiated with the energy rays E. As shown in FIG. Further, the melted region 3a may be not only a completely melted material M but also partially melted materials M joined together.

By solidifying the melted region 3a, a layer 3b is formed on the base and the modeled object 3 as a set of layers or thin films of the material M. As shown in FIG. It should be noted that the material M may be laminated in a granular state by being cooled by heat transfer to the aggregate of the material M to form a granular aggregate (layer).

The modeling unit 12 may perform the annealing process by irradiating the set of materials M with energy rays E from the nozzle 21 . The assembly of material M is re-melted or re-sintered by energy beam E and solidified to form layer 3b.

The modeling unit 12 laminate-models the modeled object 3 by repeatedly stacking the layers 3b. In this way, the nozzle 21 of the modeling unit 12 is supported by the support surface 11a by irradiating the energy beam E to melt or sinter the material M to shape the layer 3b, and by repeating the shaping of the layer 3b. The modeled object 3 is laminate-molded.

The nozzle 21 has a nozzle head 31 . A tip 31a of the nozzle head 31 faces the object 3 with a gap therebetween. The nozzle head 31 is provided with an exit path 32 , a discharge path 33 and a supply path 34 . The output path 32, the discharge path 33, and the supply path 34 are opened at the tip 31a, for example.

The exit path 32 is a hole having a substantially circular cross section. The energy ray E passes through the emission path 32 and is emitted to the outside of the nozzle head 31 . The discharge path 33 is a hole having a substantially annular cross section and is provided so as to surround the output path 32 . A carrier gas and material M are discharged outside the nozzle head 31 through the discharge passage 33 . The supply path 34 is a hole having a substantially annular cross section and is provided so as to surround the discharge path 33 . A shielding gas G is discharged to the outside of the nozzle head 31 through the supply path 34 . Shield gas G is, for example, an inert gas such as nitrogen or argon.

As shown in FIG. 1 , the supply device 22 has an optical device 41 , a material supply device 42 and a gas supply device 43 . The optical device 41 has, for example, a light source and an optical system. The light source has an oscillating element, and emits laser light as energy rays E by oscillation of the oscillating element. The light source can change the output (power) of the energy ray E to be emitted.

The light source causes the emitted energy ray E to enter the optical system. The energy ray E enters the nozzle 21 through the optical system. The optical system can change the focal diameter of the energy beam E. FIG. The optical device 41 supplies the energy ray E to the emission path 32 of the nozzle 21 and causes the energy ray E to be emitted from the emission path 32 .

The nozzle 21 heats the ejected material M by irradiating the energy beam E, thereby forming the layer 3b of the material M and performing an annealing treatment. Further, the nozzle 21 can remove an unnecessary portion of the modeled object 3 by irradiating the energy beam E. FIG.

The material supply device 42 has a material supply section 42a and a tank 42b. The tank 42b contains the material M. The material supply device 42 may have a plurality of tanks 42b containing different kinds of materials M from each other.

The material supply unit 42a supplies the material M in the tank 42b to the nozzle 21 through the supply pipe 21a. The material supply part 42a supplies the material M to the nozzle 21 by carrier gas, for example. A carrier gas is, for example, an inert gas such as nitrogen or argon.

The material supply part 42a supplies the carrier gas and the material M to the ejection path 33 of the nozzle head 31 via the supply pipe 21a. Thereby, the nozzle 21 ejects the carrier gas and the material M from the ejection path 33 . The material supply unit 42a can change the amount of the material M ejected from the nozzle 21 per unit time and the speed of the material M ejected.

The material supply unit 42a has, for example, a tank containing a carrier gas, a compressor for causing the carrier gas in the tank to flow to the supply pipe 21a, and a device for supplying the material M in the tank 42b to the flow of the carrier gas. Note that the material supply unit 42a may supply the material M to the nozzle 21 by other means.

The gas supply device 43 has a gas supply portion 43a and a tank 43b. The tank 43b accommodates the shielding gas G. The gas supply unit 43a supplies the shielding gas G in the tank 43b to the nozzle 21 through the supply pipe 21a. The supply tube 21a includes a tube through which the carrier gas and material M passes, a tube through which the shielding gas G passes, and a cable through which the energy line E passes. The gas supply unit 43a can change the amount of the shielding gas G discharged from the nozzle 21 per unit time and the speed of the shielding gas G discharged.

The moving device 23 moves and rotates the nozzle 21 . For example, the moving device 23 can move the nozzle 21 parallel in the X-axis direction, the Y-axis direction, and the Z-axis direction, and rotate it around the center of rotation parallel to the X-axis.

The moving device 23 can move the nozzle 21 relative to the support surface 11 a to change the orientation of the nozzle 21 . The moving device 23 can change the moving speed of the nozzle 21 with respect to the support surface 11a. The table 11 may move and rotate to move the nozzle 21 relative to the support surface 11a, thereby changing the orientation of the nozzle 21 with respect to the support surface 11a.

The adjuster 24 adjusts the oxygen concentration in chamber C. For example, the adjusting device 24 adjusts the oxygen concentration around the nozzle 21 and the modeled object 3 by exhausting air from the chamber C or supplying an inert gas to the chamber C.

The imaging unit 13 has an imaging device 13a. The imaging device 13 a is, for example, a video camera capable of capturing moving images and transferring the moving images to the control unit 14 . Note that the imaging device 13a may be another imaging device such as a still camera capable of capturing still images. The imaging device 13a photographs the nozzle 21 and the modeled object 3 . The imaging device 13 a may be movable following the nozzle 21 .

The control unit 14 is electrically connected to the table 11, the modeling unit 12, and the imaging device 13a via signal lines 16. FIG. The control unit 14 may be, for example, a control unit provided integrally with the modeling unit 12 or a computer provided separately from the modeling unit 12 .

The control unit 14 includes, for example, a control device such as a CPU (Central Processing Unit) 14a, a ROM (Read Only Memory) 14b, a RAM (Random Access Memory) 14c, an external storage device 14d, an output device 14e, It has an input device 14f and has a hardware configuration using a normal computer. The CPU 14a, ROM 14b, RAM 14c, external storage device 14d, output device 14e, and input device 14f are connected to each other via a bus or an interface.

The control unit 14 controls each unit of the laminate manufacturing system 1 by the CPU 14a executing a program incorporated in the ROM 14b or the external storage device 14d. For example, the control unit 14 controls the table 11 , the nozzle 21 of the modeling unit 12 , the moving device 23 , the adjusting device 24 , the optical device 41 , the material supplying device 42 , and the gas supplying device 43 .

The ROM 14b stores programs and data necessary for executing the programs. The RAM 14c functions as a work area during program execution. The external storage device 14d is, for example, a device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive) that can store, change, and delete data. The output device 14e is, for example, a display or a speaker. The input device 14f is, for example, a keyboard or mouse.

FIG. 3 is an exemplary block diagram functionally showing the configuration of the laminate manufacturing system 1 of the first embodiment. The control unit 14 realizes each unit shown in FIG. 3 by reading and executing a program stored in the ROM 14b or the external storage device 14d by the CPU 14a, for example. As shown in FIG. 3, the control unit 14 includes, for example, a storage unit 61, a modeling control unit 62, an image processing unit 63, a learning unit 64, an estimation unit 65, and an output control unit 66.

For example, the CPU 14 a functions as a shaping control section 62 , an image processing section 63 , a learning section 64 , an estimation section 65 and an output control section 66 . Also, the RAM 14 c and the external storage device 14 d function as the storage unit 61 .

The storage unit 61 stores various information including an NC program 71 , a plurality of data sets 72 , a convolutional neural network (CNN) 73 and an incorrect answer list 74 . CNN 73 may also be referred to as a computing unit or learning result.

The NC program 71 is information for modeling a plurality of layers 3b and a modeled object 3 including the layers 3b. The NC program 71 includes, for example, the shape of the object 3 and the layer 3b, the movement path of the nozzle 21, and the parameter information of the layered manufacturing at each position of the movement path. The parameters of the layered manufacturing are, for example, the moving speed of the nozzle 21, the direction of the nozzle 21, the amount of the discharged material M per unit time, the speed of the discharged material M, the output of the irradiated energy ray E, and the supply including the flow rate of the shielding gas G applied. The NC program 71 may further contain other information.

Based on the NC program 71, the modeling control unit 62 controls the modeling unit 12 including the moving device 23, the adjusting device 24, the optical device 41, the material supply device 42, and the gas supply device 43 to form a plurality of layers 3b (modeling Item 3) is laminate-molded. Further, the modeling control unit 62 can change the parameters of the layered modeling by the modeling unit 12 and stop the modeling unit 12 .

The image processing unit 63 acquires a moving image captured by the imaging device 13a and converts the moving image into a plurality of still images (images). For example, when the imaging device 13a captures a moving image at 30 fps, the image processing unit 63 can generate 30 still images per second from the moving image.

The image generated by the image processing unit 63 shows the nozzle 21 and the modeled object 3 . The image generated by the image processing unit 63 can also be called a first image, a second image, or input information. The image processing unit 63 functions as the imaging unit 13 that generates an image together with the imaging device 13a. Note that the imaging unit 13 is not limited to this example, and for example, the imaging device 13a alone may generate an image.

The learning unit 64 performs machine learning based on the data set 72 stored in the storage unit 61, and generates a CNN 73 as a learning result. Based on the image generated by the imaging unit 13 and the CNN 73, the estimation unit 65 estimates whether or not sparks are reflected in the image.

The output control unit 66 controls the output device 14e. For example, the output control unit 66 causes the output device 14e to display the result of machine learning by the learning unit 64 and the result of estimation by the estimation unit 65. FIG.

Machine learning by the learning unit 64 will be described in detail below. For learning by the learning unit 64, a plurality of data sets 72, the CNN 73, and the incorrect answer list 74 stored in the storage unit 61 are used. For example, the storage unit 61 stores hundreds to tens of thousands of data sets 72 .

Data sets 72 each include an image P and a label L; The image P may also be referred to as the first image or input information. Label L may also be referred to as spark information or correct answer information. That is, the storage unit 61 stores a plurality of images P and a plurality of labels L corresponding to the images P. FIG.

FIG. 4 shows three examples of images P according to the first embodiment. As shown in FIG. 4, the image P shows the nozzles 21 . The image P may be captured by the imaging unit 13 or may be captured by another device.

The plurality of images P includes a non-shaped image P1 on the left side in FIG. 4, a normal shaped image P2 on the center, and an abnormal shaped image P3 on the right side. Note that these names are for convenience, and the nozzles 21 shown in the normal modeled image P2 are not necessarily performing normal layered modeling, and the nozzles 21 shown in the abnormally modeled image P3 are It doesn't necessarily mean that you are doing abnormal additive manufacturing.

The non-modeling image P1 shows the nozzles 21 that are not subjected to layered modeling. The nozzle 21 stops discharging the material M and emitting the energy ray E. As shown in FIG. The nozzles 21 that are performing layered manufacturing are shown in the normal modeling image P2 and the abnormal modeling image P3. The nozzle 21 ejects the material M and emits the energy ray E. As shown in FIG. It should be noted that the nozzles 21 shown in the normal shaped image P2 and the abnormal shaped image P3 may emit the energy beam E while stopping the ejection of the material M.

A spark is reflected in the abnormally shaped image P3. For example, when titanium, which is the material M, is oxidized, sparks are generated as reflected in the abnormally shaped image P3. If the material M oxidizes, there is a risk that the modeled object 3 will become brittle. On the other hand, in the normal modeled image P2, the energy rays E and the molten or sintered material M emit light, but no large sparks are generated.

The label L is set in advance for the image P according to the state (feature) of the image P. A label L is a matrix expressed as [l ₀ , l ₁ , l ₂ ]. Different labels L are set for the non-printed image P1, the normal printed image P2, and the abnormal printed image P3.

For example, the label L corresponding to the non-printed image P1 is set to [1,0,0]. A label L corresponding to the normal shaped image P2 is set to [0, 1, 0]. A label L corresponding to the abnormally shaped image P3 is set to [0, 0, 1]. As described above, _l2 of the label L indicates whether or not the spark is reflected in the image P. Note that the label L is not limited to this example.

FIG. 5 is an exemplary diagram schematically showing an example of the configuration of the CNN 73 of the first embodiment. In the example of FIG. 5, the CNN 73 is composed of a total of 11 layers combining five types of layers: convolution layer, activation function, pooling, full connection, and softmax function.

In the example of FIG. 5, CNN 73 includes two convolutional layers and three fully connected layers. The convolutional layer has matrices [FW ₁ , FB ₁ ], [FW ₂ , FB ₂ ]. A fully connected layer has matrices [W ₁ , B ₁ ], [W ₂ , B ₂ ], [W ₃ , B ₃ ]. The matrices [FW ₁ , FB ₁ ], [FW ₂ , FB ₂ ], [W ₁ , B ₁ ], [W ₂ , B ₂ ], [W ₃ , B ₃ ] are hereinafter referred to as network models.

Forward propagation is an operation from input to output. The input is the image P, for example. The output is the result (output value) Y calculated by the CNN 73 . The forward propagation output value Y is a matrix expressed as [y ₀ , y ₁ , y ₂ ]. That is, the CNN 73 calculates the output value Y from the image P by forward propagation. The output value Y may also be referred to as output information.

Backpropagation is an operation from output to input. The difference between the output value Y calculated through the softmax function and the label L is the gradient value. By propagating the gradient value from the output to the input, the gradient value of each layer of the CNN 73 is calculated.

The above CNN 73 is an example, and the CNN 73 may have other configurations. Moreover, the laminate manufacturing system 1 may have another neural network different from the CNN 73 and another learning model of deep learning (machine learning).

FIG. 6 is an exemplary block diagram functionally showing the configuration of the learning unit 64 of the first embodiment. As shown in FIG. 6, the learning unit 64 includes a selection unit 81, a forward propagation unit 82, a back propagation unit 83, an update unit 84, a first evaluation unit 85, a second evaluation unit 86, and an evaluation output unit 87 . The forward propagation unit 82 and the backward propagation unit 83 may also be referred to as calculation units. Update unit 84 may also be referred to as a change unit. The first evaluator 85 may also be referred to as a first determiner. The second evaluator 86 may also be referred to as a second determiner or determiner. The evaluation output section 87 may also be referred to as an output section.

The selection unit 81 selects a predetermined number of data sets 72 to be used for machine learning from a plurality of data sets 72 . The forward propagation unit 82 receives the image P and carries out forward propagation of the CNN 73 . The back-propagation unit 83 receives the output value Y and performs back-propagation of the CNN 73 . The update unit 84 updates the CNN 73 based on the calculation results of the forward propagation unit 82 and the backward propagation unit 83 .

The first evaluation unit 85 evaluates the CNN 73 based on evaluation data 72 a included in the data set 72 . The second evaluation unit 86 evaluates the CNN 73 based on the learning data 72b included in the data set 72. FIG. The evaluation output section 87 outputs evaluation results of the first evaluation section 85 and the second evaluation section 86 .

As described above, the multiple data sets 72 include multiple evaluation data 72a and multiple learning data 72b. For example, the evaluation data 72a are tens to tens of thousands of data sets 72 among the plurality of data sets 72, and the learning data 72b are the remaining data sets 72. FIG.

The evaluation data 72a and the learning data 72b have images P and labels L, respectively. The image P of the evaluation data 72a can also be called an evaluation image. The image P of the learning data 72b can also be called a learning image. The image P of the learning data 72b and the image P of the evaluation data 72a are included in a plurality of images P and are different from each other.

FIG. 7 is a flow chart showing an example of the machine learning method of the first embodiment. Note that the machine learning method described below with reference to FIG. 7 is merely an example, and the learning unit 64 may perform machine learning by other methods.

In machine learning by the learning unit 64, initialization is first performed (S11). For example, the updating unit 84 generates an unupdated CNN 73 . Note that another part of the learning unit 64 may generate the unupdated CNN 73 .

For example, the size of the network model (matrix) in CNN 73 is determined from the image P and the number of nodes in the hidden layer. Therefore, the size of the matrix is specified programmatically. The numbers in the matrix are generated randomly, for example.

The update unit 84 stores the generated non-updated CNN 73 in the storage unit 61, for example. In the initialization, the learning unit 64 may read out the unupdated CNN 73 stored in advance in the storage unit 61 .

Next, the selection unit 81 selects and acquires one learning data 72b from the plurality of learning data 72b stored in the storage unit 61 (S12). The selection unit 81, for example, randomly selects one learning data 72b from a plurality of learning data 72b.

Next, the forward propagation unit 82 causes the CNN 73 to calculate the output value Y by forward propagation from the selected image P of the learning data 72b (S13). Next, the back propagation unit 83 calculates a gradient value based on the difference between the output value Y and the label L (S14).

Next, the updating unit 84 updates (changes) the CNN 73 stored in the storage unit 61 by the gradient descent method based on the gradient values (S15). In other words, the update unit 84 changes the CNN 73 based on the output value Y calculated by the forward propagation unit 82 from the image P of the learning data 72b. Note that the update unit 84 may change the CNN 73 by other methods. Further, updating of the CNN 73 by the updating unit 84 can also be expressed as generation of a new CNN 73 by the updating unit 84 .

The above S12-S15 can be called a learning process. Next, the update unit 84 determines whether or not the learning process has been performed using a predetermined number of learning data 72b (S16). If the number of learned learning data 72b has not reached the predetermined number (S16: No), the process returns to S12, and repeats S12 to S16.

In S16, when the number of learned learning data 72b has reached a predetermined number (S16: Yes), the first evaluation unit 85, from the randomly selected image P of the evaluation data 72a, CNN73 An output value Y is calculated by forward propagation (S17). Then, the first evaluation unit 85 determines that the output value Y is correct based on the output value Y calculated from the image P of the evaluation data 72a and the label L of the evaluation data 72a corresponding to the output value Y. or incorrect (S18).

Output values Y y ₀ , y ₁ , y ₂ correspond to labels L l ₀ , l ₁ , l ₂ . Therefore, _y2 calculated by the learned CNN 73 from the abnormally shaped image P3 is a value close to 1 between 0 and 1. On the other hand, _y2 calculated by the learned CNN 73 from the non-printed image P1 or the normal printed image P2 is a value close to 0 between 0 and 1. That is, _y2 is linked to the probability that the spark is reflected in the image P. Hereinafter, _y2 is referred to as the spark confirmation intensity. The spark confirmation intensity y ₂ may be indicated, for example, from 0 to 1 as described above, or may be indicated from 0 to 100%. The CNN 73 calculates the output value Y including the spark confirmation intensity _y2 from the image P by forward propagation.

The first evaluation unit 85 determines whether the output value Y is correct or incorrect by comparing the maximum index of the output value Y and the label L, for example. For example, the label L[1,0,0] has l ₀ =1, l ₁ =0, l ₂ =0, so the maximum index is "0" (l ₀ ). Similarly, the maximum index of the output value Y[0.1, 0.1, 0.8] is y ₀ =0.1, y ₁ =0.1, y ₂ =0.8, so "2 '(y ₂ ). Since the maximum index between the label L and the output value Y is different between "0" and "2", the first evaluation unit 85 determines the output value Y as an incorrect answer.

When comparing the label L[1,0,0] and the output value Y[0.8,0.1,0.1], the maximum index of the label L is "0" (l ₀ ), and the output value Y is The maximum index is '0' (y ₀ ). Since the label L and the output value Y have the same maximum index, the first evaluation unit 85 determines that the output value Y is correct.

Note that the first evaluation unit 85 may determine whether the output value Y is correct or incorrect by another method. For example, when _l2 of label L is 0, the first evaluation unit 85 determines that the answer is incorrect if spark confirmation strength _y2 is 0.5 or more, and determines that spark confirmation strength _y2 is greater than 0.5. If it is below, the correct answer is also good. A threshold (for example, 0.5) for judging these correct or incorrect answers can be set arbitrarily.

Next, the first evaluation unit 85 determines whether or not the CNN 73 has been evaluated using a predetermined number of evaluation data 72a (S19). If the number of pieces of evaluation data 72a used for evaluation has not reached the predetermined number (S19: No), the process returns to S17 and repeats S17 to S19.

In S19, when the number of evaluation data 72a used for evaluation has reached a predetermined number (S19: Yes), the evaluation output section 87 outputs the evaluation result of the first evaluation section 85 (S20). For example, the evaluation output unit 87 outputs the correct answer rate of the output value Y calculated by the CNN 73 from the evaluation data 72a.

Next, the second evaluation unit 86 calculates the output value Y from the randomly selected image P of the learning data 72b by forward propagation of the CNN 73 (S21). Then, the second evaluation unit 86 determines that the output value Y is correct based on the output value Y calculated from the image P of the learning data 72b and the label L of the learning data 72b corresponding to the output value Y. or incorrect (S22). As with the first evaluation unit 85, the second evaluation unit 86 determines whether the output value Y is correct or incorrect by comparing the maximum index of the output value Y and the label L, for example.

In S22, if the output value Y is determined to be incorrect (S22: No), the second evaluation unit 86 adds the data set 72 corresponding to the output value Y determined to be incorrect to the incorrect answer list 74. Register (S23). For example, the second evaluation unit 86 registers the number (ID) of the data set 72 corresponding to the output value Y determined to be incorrect in the incorrect answer list 74 .

When the output value Y is determined to be correct in S22 (S22: Yes), or when S23 is completed, the second evaluation unit 86 evaluates the CNN 73 using a predetermined number of learning data 72b. is determined (S24). If the number of learning data 72b used for evaluation has not reached the predetermined number (S24: No), the process returns to S21 and repeats S21 to S24.

In S24, when the number of learning data 72b used for evaluation has reached a predetermined number (S24: Yes), the evaluation output section 87 outputs the evaluation result of the second evaluation section 86 (S25). For example, the evaluation output unit 87 outputs the correct answer rate of the output value Y calculated by the CNN 73 from the learning data 72b.

The evaluation output unit 87 outputs the evaluation result of the first evaluation unit 85 and the evaluation result of the second evaluation unit 86 to the output control unit 66, for example. The output control unit 66 causes the output device 14e to output a graph based on the evaluation result of the first evaluation unit 85 and the evaluation result of the second evaluation unit 86, for example.

Next, the selection unit 81 determines whether or not there is an abnormality in the evaluation result of the first evaluation unit 85 and the evaluation result of the second evaluation unit 86 (S26). When there is no abnormality in the evaluation result (S26: No), the selection unit 81 determines whether or not the percentage of correct answers for the output value Y has converged (S27).

For example, the selection unit 81 selects whether at least one of the percentage of correct answers in the evaluation results of the first evaluation unit 85 and the percentage of correct answers in the evaluation results of the second evaluation unit 86 changes over a predetermined number of times in a predetermined range. If it is within, it is determined that the percentage of correct answers for the output value Y has converged. In other words, the selection unit 81 determines that the correct answer rate of the output value Y has converged when the evaluation results of the first evaluation unit 85 and the second evaluation unit 86 do not improve over a predetermined number of times. Not limited to this example, for example, the selection unit 81 selects the output value Y It may be determined that the percentage of correct answers has converged.

If the percentage of correct answers for the output value Y has not converged (S27: No), the process returns to S12, and S12 to S26 are executed (repeatedly) again. In S12 at the time of repetition, the selection unit 81 selects the learning data 72b so that, for example, a predetermined number of the learning data 72b used in the learning steps of S12 to S15 includes incorrect answer data 72bf at a predetermined ratio. select.

The incorrect answer data 72bf is learning data 72b registered in the incorrect answer list 74 corresponding to the output value Y judged to be incorrect by the second evaluation unit 86 . The learning data 72b used in the learning process includes incorrect answer data 72bf selected from the incorrect answer list 74 and learning data 72b randomly selected from a plurality of learning data 72b.

In S13 at the time of repetition, the forward propagation unit 82 causes the CNN 73 to calculate the output value Y from each of the image P of the incorrect answer data 72bf and the randomly selected image P of the learning data 72b. That is, the selection unit 81 selects the image P of the incorrect answer data 72bf and the randomly selected image P of the learning data 72b as the images P to be input to the CNN 73 .

The image P of the incorrect answer data 72bf is included in the plurality of images P stored in the storage unit 61, and corresponds to the output value Y judged to be incorrect. The image P of the incorrect answer data 72bf can also be referred to as first input information.

The randomly selected image P of the learning data 72b is included in the plurality of images P stored in the storage unit 61, and is randomly selected from the plurality of images P. The randomly selected image P of the learning data 72b can also be referred to as second input information.

FIG. 8 is a graph showing an example of the correct answer rate of the output value Y in the evaluation results of the first evaluation section 85 and the second evaluation section 86 of the first embodiment. As a comparative example, FIG. 9 shows the output values of the evaluation results of the first evaluation unit 85 and the second evaluation unit 86 when all of the predetermined number of learning data 72b used in the learning process are randomly selected. It is a graph which shows an example of the correct answer rate of Y. FIG.

Graphs such as those in FIGS. 8 and 9 are generated by the evaluation output unit 87, for example. 8 and 9, the correct answer rate of the output value Y in the first evaluation unit 85 using the evaluation data 72a and the correct answer rate of the output value Y in the second evaluation unit 86 using the learning data 72b , is shown. 8 and 9, the vertical axis indicates the percentage of correct answers for the output value Y, and the horizontal axis indicates the number of repetitions of S12 to S27.

As shown in FIGS. 8 and 9, the incorrect answer data 72bf are included at a predetermined rate in the predetermined number of learning data 72b used in the learning process, so that the percentage of correct answers for the output value Y can be easily increased. For example, in the present embodiment, the correct answer rate of the output value Y in the evaluation result of the first evaluation unit 85 converges at 98% or more earlier than in the comparative example. The ratio or number of incorrect answer data 72bf included in the predetermined number of learning data 72b used in the learning process can be set arbitrarily.

In S26 of FIG. 7, if there is an abnormality in the evaluation result (S26: Yes), the selection unit 81 selects the incorrect answer data 72bf included in the predetermined number of learning data 72b used in the learning steps of S12 to S15. is changed (S28). In other words, the selection unit 81 selects an image P of the incorrect answer data 72bf and an image of the randomly selected learning data 72b, which are input to the CNN 73, based on the ratio of the output values Y determined to be correct (percentage of correct answers). Change the proportion of P.

For example, when the proportion of incorrect data 72bf is high, overlearning occurs. In this case, as the number of iterations increases, the percentage of correct answers decreases. The selection unit 81 reduces the ratio of the incorrect answer data 72bf when such an abnormality occurs in the evaluation result (percentage of correct answers).

On the other hand, when the proportion of the incorrect answer data 72bf is small, it is difficult for the correct answer rate to converge even if the number of iterations increases. The selection unit 81 increases the ratio of the incorrect answer data 72bf when such an abnormality in the evaluation result (percentage of correct answers) occurs. The ratio or number of incorrect answer data 72bf to be increased or decreased can be set arbitrarily.

When the selection unit 81 changes the ratio of the incorrect answer data 72bf in S28, the process returns to S12, and S12 to S28 are repeated. In S27, if the correct answer rate of the output value Y has converged (S27: Yes), the machine learning by the learning unit 64 is completed. By the above, CNN73 is produced|generated as a learning result of machine learning.

In the machine learning described above, the first evaluation unit 85 and the second evaluation unit 86 evaluated the correct answer rate of the output value Y. However, the first evaluation unit 85 and the second evaluation unit 86 are not limited to this example, and may evaluate average loss, for example. The loss can be calculated based on the output value Y and the label L using cross entropy.

FIG. 10 is an exemplary block diagram functionally showing the configuration of the estimation unit 65 of the first embodiment. As shown in FIG. 10 , the estimator 65 includes an intensity calculator 91 and a determiner 92 .

The intensity calculation unit 91 acquires the image from the image processing unit 63 and acquires the CNN 73 from the storage unit 61 . The CNN 73 acquired by the intensity calculator 91 is a learning result generated by machine learning by the learning unit 64 described above. The intensity calculator 91 causes the CNN 73 to calculate the output value Y from the image acquired from the image processor 63 by forward propagation.

Based on the output value Y, the determination unit 92 determines whether or not the image corresponding to the output value Y includes sparks. In this embodiment, the determination unit 92 determines from the image whether the layered manufacturing is stopped, whether the layered manufacturing is normally performed, or whether the layered manufacturing is abnormal.

FIG. 11 is a flow chart showing an example of the layered manufacturing monitoring method of the first embodiment. The monitoring of layered manufacturing by the layered manufacturing system 1 will be described in detail below. Note that the lamination manufacturing monitoring method described below with reference to FIG. 11 is merely an example, and the lamination manufacturing system 1 may perform lamination manufacturing monitoring by other methods.

As shown in FIG. 11, first, the modeling unit 12 starts lamination modeling based on the NC program 71 (S101). It should be noted that the start of lamination manufacturing in S101 is the start of execution of the NC program 71 . Therefore, the nozzle 21 does not have to discharge the material M and emit the energy beam E.

Next, the intensity calculator 91 acquires an image showing the nozzle 21 from the image processor 63 (S102). Next, the intensity calculator 91 calculates the output value Y from the image by forward propagation of the CNN 73 (S103). In other words, the intensity calculator 91 uses the CNN 73 to calculate the spark confirmation intensity _y2 from the image.

Next, the determination unit 92 determines whether or not _y0 is the largest among the output values Y[ _y0 , _y1 , _y2 ] (S104). In other words, the determination unit 92 determines whether the image corresponding to the output value Y is similar to the non-formed image P1. When _y0 is the maximum (S104: Yes), the determination unit 92 determines that the nozzle 21 is not performing layered manufacturing (S105).

Next, the determination unit 92 determines whether or not the layered manufacturing is completed (S106). If the layered manufacturing has not been completed (S106: No), the process returns to S102, and the intensity calculation unit 91 acquires the next image from the image processing unit 63.

In S104, when _y0 is not the maximum (S104: No), the determination unit 92 determines whether _y1 is the maximum among the output values Y[ _y0 , _y1 , _y2 ] (S107). . In other words, the determination unit 92 determines whether the image corresponding to the output value Y is similar to the normal molded image P2. If _y1 is the maximum (S107: Yes), the determination unit 92 determines that the nozzle 21 is performing normal layered manufacturing (S108), and proceeds to S106.

In S107, when _y1 is not the maximum (S107: No), the spark detection intensity _y2 is the maximum among the output values Y[ _y0 , _y1 , _y2 ]. Therefore, it is highly possible that the image corresponding to the output value Y is similar to the abnormally shaped image P3 and sparks are generated.

When the spark confirmation intensity _y2 is the maximum, the determination unit 92 determines that an abnormality has occurred in the layered manufacturing (S109). In this case, the modeling control unit 62 changes the parameters of the layered modeling (S110), and proceeds to S106.

For example, the modeling control unit ₆₂ determines the flow rate of the inert gas supplied to the chamber C, the flow rate of the shielding gas G discharged from the nozzle 21, and the energy ray E emitted from the nozzle 21 based on the spark confirmation intensity y2. At least one of the output, the speed of the nozzle 21, and the amount of material M ejected from the nozzle 21 is changed to suppress spark generation and oxidation of the material M. The amount of parameter change by the modeling control unit 62 may be constant, or may be linked to the spark confirmation strength. When the speed of the nozzle 21 and the amount of the material M to be discharged are changed, the parameters of the layered manufacturing are adjusted so that the height of the melted region 3a is substantially constant.

In S106, when the layered manufacturing is completed (S106: Yes), the monitoring of the layered manufacturing by the layered manufacturing system 1 is also completed. As described above, the layered manufacturing system 1 monitors the generation of sparks in layered manufacturing. When sparks occur in lamination molding, the modeling control unit 62 changes the parameters of lamination molding to suppress the generation of sparks.

In the lamination manufacturing monitoring described above, if _y1 is not the maximum in S107, it is determined that an abnormality has occurred in the lamination manufacturing. However, not limited to this example, for example, when the spark confirmation intensity _y2 becomes maximum over a predetermined number of times, or when the spark confirmation intensity _y2 accumulated over a predetermined number of times exceeds the threshold, It may be determined that an abnormality has occurred in the layered manufacturing.

The determination unit 92 may output the determination result to the output control unit 66 in association with the coordinates of the nozzle 21, for example. Based on the determination result, a record (log) of the state of lamination manufacturing at each coordinate of the nozzle 21 can be obtained.

In the laminate manufacturing system 1 according to the first embodiment described above, the learning unit 64 discharges the material M of the modeled object 3, and irradiates the material M with the energy beam E to melt or sinter it. Machine learning is performed based on a plurality of images P showing the nozzles 21 and labels L indicating whether or not sparks are shown in the plurality of images P to generate the CNN 73 . Based on the image captured by the imaging unit 13 and the CNN 73, the estimation unit 65 estimates whether or not sparks are reflected in the image. In the melting or sintering of material M such as titanium, if sparks are generated, material M may be oxidized. Since the monitoring of spark generation can be automated based on the estimation result of the estimation unit 65, weakening of the modeled object 3 due to the oxidized material M is suppressed.

For example, if a light intensity sensor monitors the generation of sparks instead of a computing unit such as CNN 73 that is updated by machine learning, there is a risk that light emission in normal lamination molding will be erroneously recognized as generation of sparks. In the present embodiment, the accuracy of monitoring the occurrence of sparks is improved by using a calculation unit such as the CNN 73 that is updated by machine learning.

The material M in this embodiment contains titanium. As described above, in the melting or sintering of material M such as titanium, if sparks are generated, material M may be oxidized. Therefore, since it is possible to automate the monitoring of spark generation based on the estimation result of the estimation unit 65, weakening of the modeled object 3 due to oxidized titanium is suppressed.

The estimating unit 65 uses the CNN 73 to calculate, from the image captured by the imaging unit 13, the spark confirmation strength _y2 linked to the probability that sparks are reflected in the image. Since the estimation unit 65 calculates the numerical value (spark confirmation strength y ₂ ), various controls can be performed based on the numerical value.

The modeling control unit 62 changes the parameters of the layered modeling by the modeling unit ₁₂ based on the spark verification intensity y2. For example, the modeling control unit ₆₂ determines the flow rate of the inert gas supplied to the chamber C where the layered modeling is performed, the flow rate of the shielding gas G discharged from the nozzle 21, and the At least one of the output of the energy beam E, the speed of the nozzle 21, and the amount of the material M ejected from the nozzle 21 is controlled to suppress generation of sparks and oxidation of the material M. As a result, weakening of the modeled object 3 due to the oxidized material M is suppressed.

The first evaluation unit 85 determines whether the output value Y is correct or incorrect based on the label L and the output value Y calculated by the forward propagation unit 82 from the image P of the evaluation data 72a. The second evaluation unit 86 determines whether the output value Y is correct or incorrect based on the label L and the output value Y calculated by the forward propagation unit 82 from the image P of the learning data 72b different from the image P of the evaluation data 72a. Determine if correct. The evaluation output section 87 outputs the evaluation result of the first evaluation section 85 and the evaluation result of the second evaluation section 86 . Thereby, based on the determination result of the second evaluation unit 86, the accuracy of the CNN 73 generated by machine learning can be evaluated. Furthermore, based on the determination result of the first evaluation unit 85, the accuracy of the CNN 73 targeting the image P of the evaluation data 72a that has not been used for machine learning can be evaluated. Therefore, the accuracy of monitoring the generation of sparks is improved, and weakening of the modeled object 3 due to the oxidized material M is suppressed.

The CNN 73 extracts the output value Y from each of the image P of the incorrect answer data 72bf corresponding to the output value Y that is included in the plurality of images P and determined to be incorrect, and the image P included in the plurality of images P. calculate. Then, the CNN 73 is changed based on the output value Y. That is, the CNN 73 performs machine learning using an image P that has been mistaken at least once. As a result, the accuracy of the output value Y calculated by the CNN 73 can be easily improved, and the time (number of iterations) required for machine learning can be shortened.

In a plurality of images P, the ratio of non-printed images P1, normal printed images P2, and abnormal printed images P3 may be uneven. In the present embodiment, the CNN 73 performs machine learning using an image P that has been mistaken at least once. A similar learning effect can be obtained when the ratio is uniform.

The CNN 73 calculates the output value Y from each of the image P of the correct proportion of the incorrect answer data 72bf and the image P randomly selected from the plurality of images P. Thereby, the bias of the image P input to the CNN 73 is reduced. Therefore, the accuracy of the output value Y calculated by the CNN 73 can be easily improved, and the time (number of iterations) required for machine learning can be shortened.

Based on the ratio of the output values Y determined to be correct, the ratio of the image P of the incorrect answer data 72bf input to the CNN 73 is changed. For example, the ratio of the image P of the incorrect answer data 72bf is reduced when overlearning occurs, and the ratio of the image P of the incorrect answer data 72bf is increased when the convergence of the ratio of the output values Y determined to be correct is slow. be As a result, the accuracy of the output value Y calculated by the CNN 73 can be easily improved, and the time (the number of iterations) required for machine learning can be shortened.

The program executed by the control unit 14 of the present embodiment is a file in an installable format or an executable format, and can be stored on a computer such as a CD-ROM, a flexible disk (FD), a CD-R, a DVD (Digital Versatile Disk), etc. It is recorded on a readable recording medium and provided.

Further, the program executed by the control unit 14 of this embodiment may be stored in a computer connected to a network such as the Internet, and may be provided by being downloaded via the network. Further, the program executed by the control unit 14 of this embodiment may be configured to be provided or distributed via a network such as the Internet.

Further, the program of the present embodiment may be configured so as to be pre-installed in a ROM or the like and provided.

The program executed by the control unit 14 of this embodiment has a module configuration including the above-described units (the modeling control unit 62, the image processing unit 63, the learning unit 64, the estimation unit 65, and the output control unit 66). As actual hardware, a CPU (processor) reads a program from the storage medium and executes it, thereby loading each of the above sections onto the main storage device, forming a modeling control section 62, an image processing section 63, a learning section 64, an estimation A part 65 and an output control part 66 are generated on the main storage device.

(Second embodiment)
A second embodiment will be described below with reference to FIG. In the following description of the embodiments, constituent elements having functions similar to those already explained may be assigned the same reference numerals as the constituent elements already explained, and further explanation may be omitted. In addition, a plurality of components with the same reference numerals may not all have common functions and properties, and may have different functions and properties according to each embodiment.

FIG. 12 is a flow chart showing an example of the layered manufacturing monitoring method according to the second embodiment. As shown in FIG. 12, in the second embodiment, when S109 is executed, the modeling control unit 62 stops the layered manufacturing (S201) instead of changing the parameters of the layered manufacturing, and the layered manufacturing system 1 monitoring of additive manufacturing is also completed. That is, the modeling control unit 62 stops the modeling unit 12 based on the spark confirmation intensity _y2 .

In the layered manufacturing system 1 of the second embodiment described above, the modeling control unit 62 can stop the modeling unit 12 based on the spark confirmation strength _y2 . That is, when the material M is likely to be oxidized, the modeling control unit 62 can stop the layered modeling, and the brittleness of the modeled object 3 due to the oxidized material M is suppressed.

By outputting the determination result from the determination unit 92 to the output control unit 66, it is possible to easily identify the location where the material M may be oxidized from the record (log) of the determination result. Therefore, it is possible to remove the part where the material M may have been oxidized and restart the layered manufacturing.

In at least a plurality of embodiments described above, the control unit 14 includes the storage unit 61 , the learning unit 64 and the estimation unit 65 . However, for example, at least one of the storage unit 61 , the learning unit 64 , and the estimation unit 65 may be provided in a device or system different from the laminate manufacturing system 1 .

Further, in the above-described multiple embodiments, the machine learning method by the learning unit 64 is applied to the image recognition technique for estimating whether or not the image P contains sparks. However, the machine learning method by the learning unit 64 described above may be applied to other techniques such as other image recognition techniques, voice recognition techniques, and natural language processing techniques.

While several embodiments of the invention have been described, these embodiments have been presented by way of example and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and modifications can be made without departing from the scope of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are included in the scope of the invention described in the claims and equivalents thereof.

DESCRIPTION OF SYMBOLS 1... Layered manufacturing system, 3... Modeled object, 13... Imaging part, 14... Control part, 14a... CPU, 14b... ROM, 14c... RAM, 14d... External storage device, 14e... Output device, 14f... Input device, 21 Nozzle 61 Storage unit 62 Modeling control unit 63 Image processing unit 64 Learning unit 65 Estimation unit 66 Output control unit 72 Data set 72a Evaluation data 72b Learning Data for 72bf Incorrect answer data 73 CNN 74 Incorrect answer list 81 Selection unit 82 Forward propagation unit 83 Back propagation unit 84 Update unit 85 First evaluation unit 86 Second evaluation section 87 Evaluation output section 91 Intensity calculation section 92 Determination section M Material E Energy ray G Shield gas P Image P1 Non-formed image P2 Normal modeled image, P3... Abnormal modeled image, L... Label, Y... Output value.

Claims (8)

A computer implemented machine learning method comprising:
causing a calculation unit to calculate output information from each of a plurality of pieces of input information;
Determining whether the output information is correct or incorrect based on the output information and correct information set for the plurality of pieces of input information;
at least one piece of first input information included in the plurality of pieces of input information and corresponding to the output information determined to be incorrect, and at least one piece of second information included in the plurality of pieces of input information; calculating the output information from each of the input information;
changing the computing unit based on the output information;
changing the ratio of the first input information and the second input information input to the calculation unit based on the ratio of the output information determined to be correct;
A machine learning method comprising 2. The machine learning method of claim 1, wherein said second input information is randomly selected from said plurality of input information. 3. The machine learning method according to claim 1 , wherein said computing unit is a neural network. 4. The machine learning method according to any one of claims 1 to 3 , wherein said calculation unit is changed based on a gradient value obtained based on said output information and said correct answer information. causing the calculation unit to calculate the output information from the plurality of pieces of input information; determining whether the output information is correct or incorrect; and providing the calculation unit with the first input information and the second input. repeating calculating the output information from each of the information and changing the computing unit based on the output information;
further comprising
When the number of iterations exceeds a first threshold and the ratio of the output information determined to be correct among the output information in the iterations decreases to exceed a second threshold, reducing the proportion of the first input information that
2. The machine learning method of claim 1. causing the calculation unit to calculate the output information from the plurality of pieces of input information; determining whether the output information is correct or incorrect; and providing the calculation unit with the first input information and the second input. repeating calculating the output information from each of the information and changing the computing unit based on the output information;
further comprising
When the number of iterations exceeds a third threshold and the ratio of the output information determined to be correct among the output information in the iterations fluctuates by exceeding a fourth threshold, increasing the proportion of the first input information that
2. The machine learning method of claim 1. a calculation unit that calculates output information from each of a plurality of pieces of input information;
a determination unit that determines whether the output information is correct or incorrect based on the output information and correct information set for the plurality of pieces of input information;
As the plurality of pieces of input information to be input to the computing unit, at least one piece of first input information included in the plurality of pieces of input information and corresponding to the output information determined to be incorrect, and the plurality of pieces of input information. a selection unit that selects at least one second input information included in the
a changing unit that changes the computing unit based on the output information;
and
The selection unit changes the ratio of the first input information and the second input information input to the calculation unit based on the ratio of the output information determined to be correct.
Information processing equipment. to the computer,
A calculation function for calculating output information from each of a plurality of input information;
a judgment function for judging whether the output information is correct or incorrect based on the output information and correct information set for the plurality of pieces of input information;
at least one first input information included in the plurality of input information and corresponding to the output information determined to be incorrect as the plurality of input information to be input to the arithmetic function; and the plurality of input information a selection function that selects at least one second input information included in
a change function that changes the arithmetic function based on the output information;
a ratio changing function for changing the ratio of the first input information and the second input information input to the arithmetic function based on the ratio of the output information determined to be correct;
program to make it happen.

Images