Model creation method, model creation program, defect determination device, and defect determination program

Abstract

A model creation method, a model creation program, a defect determination device, and a defect determination program that can accurately determine whether or not a product is defective. [Solution] In a forming device 70, state waveforms Aw and Bw of two types of parts involved in the forming of an alloy product are acquired, and centered waveforms Ac and Bc are calculated by centering these. An interaction waveform ABt is then created by integrating the centered waveforms Ac and Bc, and a judgment model M is created from this interaction waveform ABt using machine learning using a random forest. Here, since the interaction waveform ABt is obtained by integrating the centered waveforms Ac and Bc, state changes of each part due to phenomena that cause defects in the alloy product are reflected in the interaction waveform ABt. By learning the interaction waveform ABt that reflects such state changes, a judgment model M that can accurately detect defects in alloy products can be created.

Description

[Technical Field]

[0001] The present invention relates to a model creation method, a model creation program, a defect determination device, and a defect determination program. [Background technology]

[0002] Patent Document 1 discloses an injection device 1 that uses die casting to form products made of aluminum, magnesium, etc. Specifically, the molten metal poured into an injection sleeve 63 is pushed in by a plunger tip 65 provided on a piston rod 68 connected to an injection piston 5 in a first cylinder chamber 3, whereby the molten metal fills a cavity C formed in molds 61 and 62, and after a certain time has passed, the molds 61 and 62 are opened to remove the product in which the molten metal has solidified. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2004-330267 Summary of the Invention [Problem to be solved by the invention]

[0004] However, even if the product has defects such as blowholes or leaks, it is not possible to confirm whether the product is defective until the solidified product is cooled and removed from the molds 61, 62, or until the product removed from the molds 61, 62 is processed.

[0005] Therefore, the applicant of the present application decided to create a judgment model (not publicly known) that judges whether or not a molded product is defective based on waveforms detected from parts of the injection device 1 involved in molding the product, such as the velocity waveform of the piston rod 68, the pressure waveform inside the cavity C, and the pressure waveform in the space on the piston rod 68 side of the injection piston 5 in the first cylinder chamber 3 during molding of the product.

[0006] However, although judgment models have been created using individual waveforms from these waveforms or using waveforms arranged horizontally, there is a problem in that the accuracy of determining whether or not a product is defective is low.

[0007] The present invention has been made to solve the above-mentioned problems, and aims to provide a model creation method, a model creation program, a defect determination device, and a defect determination program that can accurately determine whether a product is defective. [Means for solving the problem]

[0008] To achieve this objective, the model creation method of the present invention includes a waveform acquisition step for acquiring state waveforms detected at two or more parts of a manufacturing device, the state waveforms being based on the transition of states detected at parts involved in the production of a product; an interaction waveform creation step for creating an interaction waveform by mixing the two or more state waveforms acquired in the waveform acquisition step; and a model creation step for creating a judgment model for determining whether or not the product is defective by machine learning the interaction waveform created in the interaction waveform creation step.

[0009] The model creation program of the present invention is a program that causes a computer to execute a model creation process to create a judgment model for determining whether or not a product is defective, and causes the computer to execute the following steps: a waveform acquisition step of acquiring state waveforms based on the transition of states detected at parts of a manufacturing device involved in the production of the product, the state waveforms being detected at two or more of the parts; an interaction waveform creation step of creating an interaction waveform by mixing the two or more state waveforms acquired in the waveform acquisition step; and a model creation step of creating the judgment model by machine learning the interaction waveform created in the interaction waveform creation step.

[0010] The defect judgment device of the present invention is equipped with an acquisition means for acquiring state waveforms based on the transition of states detected at parts involved in the production of products in a manufacturing device, each of which is detected at two or more of the parts, and a judgment means for judging whether or not the product is defective based on the two or more state waveforms acquired by the acquisition means and a judgment model created by the model creation method described in claim 1.

[0011] Furthermore, the defect judgment program of the present invention is a program that causes a computer to execute a defect judgment process for judging whether or not a product is defective, and causes the computer to execute an acquisition step of acquiring state waveforms detected at two or more parts, the state waveforms being based on the transition of states detected at parts involved in the production of the product in the manufacturing equipment, and a judgment step of judging whether or not the product is defective using the two or more state waveforms acquired in the acquisition step and a judgment model created by the model creation method described in claim 1. [Effects of the Invention]

[0012] According to the model creation method of claim 1, state waveforms based on state transitions detected at parts involved in the manufacture of a product in a manufacturing device, detected at two or more parts, are acquired, two or more of the state waveforms are selected, and an interaction waveform is created by mixing the selected state waveforms. Then, machine learning is performed on the created interaction waveform to create a decision model for determining whether or not a product is defective.

[0013] Here, the interaction waveform is a mixture of state waveforms resulting from state transitions in two or more parts, so the state changes in each part due to the phenomenon that causes the product defect are reflected in the interaction waveform. In particular, when the states of two or more parts change simultaneously due to the phenomenon that causes the defect, these changes are reflected prominently in the interaction waveform. In this way, by learning the interaction waveform that reflects the state changes due to product defects in each part of the manufacturing equipment, it is possible to create a judgment model that can accurately detect product defects.

[0014] According to the model creation method of claim 2, in addition to the effects achieved by the model creation method of claim 1, an interaction waveform is created by multiplying two or more state waveforms. In particular, when the states of two or more parts change simultaneously, these changes are multiplied, and these changes are reflected more prominently in the interaction waveform. In this way, by learning an interaction waveform that better reflects state changes caused by product defects in two or more parts of a manufacturing device, there is an effect that it is possible to create a judgment model that can detect product defects with higher accuracy.

[0015] According to the model creation method of claim 3, in addition to the effects of the model creation method of claim 1, for each of two or more state waveforms, centering is performed by subtracting each value in the state waveform by the average value of the state waveforms, and an interaction waveform is created from the centered state waveforms. Since the state waveforms are acquired from different parts of the manufacturing equipment, their amplitudes vary. Therefore, by performing centering by subtracting each value in the state waveform by the average value of the state waveforms, the difference in the influence of the amplitudes between the two or more state waveforms is reduced, and the influence between the two or more state waveforms can be reduced.

[0016] An interaction waveform is created from two or more centered state waveforms in this way, and the interaction waveform is prevented from being significantly influenced by a specific state waveform. This allows the waveform to be equally influenced by each of the two or more state waveforms, and has the effect of creating a judgment model that can more accurately detect product defects based on the interaction waveform.

[0017] According to the model creation method described in claim 4, in addition to the effects achieved by the model creation method described in claim 1, there is also the effect of being able to create a judgment model that can accurately detect product defects by using a random forest using interaction waveforms.

[0018] According to the model creation method of claim 5, in addition to the effect achieved by the model creation method of any one of claims 1 to 4, there is an effect that a judgment model can be created that can judge the presence or absence of defective casting voids in a product based on the injection velocity waveform and vacuum waveform acquired as state waveforms.

[0019] According to the model creation method of claim 6, in addition to the effect achieved by the model creation method of any one of claims 1 to 4, there is an effect that a judgment model can be created that can judge whether or not there is a leakage defect in a product based on the injection velocity waveform and the rod pressure waveform acquired as state waveforms.

[0020] According to the model creation method of claim 7, in addition to the effect achieved by the model creation method of any one of claims 1 to 4, there is an effect that a judgment model can be created that can determine whether or not a product has a leakage defect based on the head pressure waveform and vacuum waveform acquired as state waveforms.

[0021] According to the model creation program recited in claim 8, there is an effect that a decision model similar to that of the model creation method recited in claim 1 can be created.

[0022] According to the defect determination device of claim 9, there is an effect that the presence or absence of a defect in a product can be determined with high accuracy using the determination model created by the model creation method of claim 1.

[0023] According to the defect determination program of claim 10, there is an effect that the presence or absence of a defect in a product can be determined with high accuracy using the determination model created by the model creation method of claim 1. [Brief explanation of the drawings]

[0024] [Figure 1] FIG. 1 is a diagram showing an outline of a PC and a molding device. [Figure 2] 1A and 1B are graphs showing a vacuum waveform, an injection speed waveform, a head pressure waveform, and a rod pressure waveform, respectively. [Figure 3]FIG. 10 is a diagram illustrating a learning method for a determination model. [Figure 4] (a) is a block diagram showing the electrical configuration of a PC, (b) is a diagram showing a schematic diagram of a judgment model table, (c) is a diagram showing a schematic diagram of learning data, and (d) is a diagram showing a schematic diagram of verification data. [Figure 5] 10A is a flowchart of the main processing, and FIG. 10B is a flowchart of the model creation processing. [Figure 6] 10 is a flowchart of a candidate model creation process. [Figure 7] 10 is a flowchart of a model verification process. [Figure 8] 10 is a flowchart of a defect determination process. DETAILED DESCRIPTION OF THE INVENTION

[0025] A preferred embodiment will now be described with reference to the accompanying drawings. An overview of the PC1 of this embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing an overview of the PC1 and a forming device 70. The PC1 is an information processing device that creates a determination model M, which is a learning model that determines whether or not there is a defect in an alloy product formed by the forming device 70.

[0026] The PC 1 is provided with a display device 2 that displays the determination results of the presence or absence of defects in the alloy product, and a mouse 3 and keyboard 4 for inputting instructions from the user. The PC 1 is also connected to a data collection device 20. The data collection device 20 is an information processing device that collects state waveforms of each part acquired by the connected molding device 70 and calculates the collected state waveforms. The PC 1 also creates a determination model M based on the integrated waveform acquired from the data collection device 20 and uses the determination model M to determine the presence or absence of defects in the alloy product. Details of the state waveforms will be described later.

[0027] The molding device 70 is a manufacturing device that forms an alloy product (product) by die casting. The molding device 70 is provided with a mold 71 for forming the alloy product, a vacuum valve 72 and a vacuum pump 73 that reduce the pressure inside the cavity formed by the mold 71, and a vacuum pressure sensor 74 that is connected to the vacuum valve 72 and the vacuum pump 73 and detects the pressure inside the cavity.

[0028] The molding device 70 is further provided with an injection sleeve 75 that supplies molten metal into the mold 71, a plunger tip 76 that pushes the molten metal inside the injection sleeve 75 into the cavity formed by the mold 71, a rod 77 that operates the plunger tip 76, a stroke sensor 78 that detects the displacement of the rod 77, an injection piston 80 and a booster piston 81 that are provided in an injection cylinder 79 that operates the rod 77, a rod-side pressure sensor 82, and a head-side pressure sensor 83.

[0029] The rod-side pressure sensor 82 is a sensor that detects the pressure of a space formed on the rod 77 side of the injection piston 80 in the injection cylinder 79. The head-side pressure sensor 83 is a sensor that detects the pressure of a space formed between the injection piston 80 and the booster piston 81 in the injection cylinder 79.

[0030] The molten metal poured into the injection sleeve 75 is pushed in by the plunger tip 76, filling the cavity formed in the mold 71, and after a certain time has passed, the mold 71 is opened to remove the solidified molten metal alloy product. In the molding device 70, during molding of such an alloy product, the pressure detected by the vacuum pressure sensor 74, the speed of the rod 77 detected by the stroke sensor 78, the pressure detected by the rod-side pressure sensor 82, and the pressure detected by the head-side pressure sensor 83 are all detected.

[0031] Then, a vacuum waveform, which is a waveform that represents the pressure detected by the vacuum pressure sensor 74 over time, an injection velocity waveform, which is a waveform that represents the velocity of the rod 77 detected by the stroke sensor 78 over time, a rod pressure waveform, which is a waveform that represents the pressure detected by the rod-side pressure sensor 82 over time, and a head pressure waveform, which is a waveform that represents the pressure detected by the head-side pressure sensor 83 over time, are generated and transmitted to the data collecting device 20. These vacuum waveform, injection velocity waveform, head pressure waveform, and rod pressure waveform are collectively referred to as "state waveforms." Figure 2 shows examples of the vacuum waveform, injection velocity waveform, head pressure waveform, and rod pressure waveform.

[0032] Figure 2 shows graph G1 of the vacuum waveform, graph G2 of the injection speed waveform, graph G3 of the head pressure waveform, and graph G4 of the rod pressure waveform. The horizontal axis in Figure 2 represents time, and the vertical axis represents pressure (graphs G1, G3, G4) or speed (graph G2).

[0033] 2 are the periods required for each step in molding the alloy product, specifically, period P1 is the period during which the injection sleeve 75 is filled with molten metal, and period P2 is the period during which the vacuum pump is used to evacuate the gas from the cavity formed in the mold 71. Furthermore, period P3 is the period during which the cavity is filled with molten metal in a short period of time, and period P4 is the period during which the pressure-increasing piston 81 is pushed toward the rod 77 to replenish the molten metal in the cavity by adding the amount of molten metal that has solidified and contracted in the cavity.

[0034] The vacuum waveform, injection speed waveform, head pressure waveform, and rod pressure waveform detected and generated in the molding device 70 in this way are transmitted to the data collection device 20. The data collection device 20 generates an integrated waveform by integrating these vacuum waveform, injection speed waveform, head pressure waveform, and rod pressure waveform into one piece of data, and transmits the generated integrated waveform to the PC1.

[0035] In the PC 1, a determination model M is created by extracting two types of state waveforms from the integrated waveform received from the data collecting device 20 and learning the resulting mixed waveform. Then, the created determination model M and the integrated waveform acquired by the forming device 70 during the forming of the alloy product and received from the data collecting device 20 are used to determine whether or not the formed alloy product is defective. First, a learning method for the determination model M in this embodiment will be described with reference to Fig. 3.

[0036] FIG. 3 is a diagram illustrating a learning method for the determination model M. In learning the determination model M, first, the PC1 receives an integrated waveform from the data collection device 20. At the same time, when the received integrated waveform is detected, information on defects that occurred in the molded alloy product is also input to the PC1. In this embodiment, "blowhole defects" and "leakage defects" are set as defects in the alloy product, but other defects in the alloy product may also be set.

[0037] Furthermore, the information on defects occurring in the alloy product used by the PC 1 may be information on defects corresponding to the integrated waveform that is registered in the database when the integrated waveform is acquired by the data collecting device 20, or may be information on defects input by the user via the mouse 3 or keyboard 4. The integrated waveform acquired in this way and information on defects in the alloy product corresponding to the integrated waveform are stored in association with the PC 1.

[0038] To create the judgment model M, first, an integrated waveform associated with a specific alloy product defect (e.g., a "defective casting cavity") is acquired from the integrated waveforms stored in PC1. For each of the acquired integrated waveforms, two types of state waveforms are extracted from the vacuum waveform, injection velocity waveform, head pressure waveform, and rod pressure waveform contained in the integrated waveform. The extracted waveforms are referred to as the state waveform Aw (e.g., vacuum waveform) and the state waveform Bw (e.g., injection velocity waveform), respectively.

[0039] Then, the average value Aav of the state waveform Aw and the average value Bav of the state waveform Bw are calculated. Specifically, the average value Aav is calculated by summing the pressure or velocity values ​​included in the state waveform Aw and dividing that value by the number of pressures or velocities included in the state waveform Aw. Similarly, the average value Bav is calculated by summing the pressure or velocity values ​​included in the state waveform Bw and dividing that value by the number of pressures or velocities included in the state waveform Bw.

[0040] Next, for each of the acquired integrated waveforms, the centered waveform Ac is calculated by subtracting the average value Aav from each pressure or velocity value included in the state waveform Aw. Similarly, the centered waveform Bc is calculated by subtracting the average value Bav from each pressure or velocity value included in the state waveform Bw.

[0041] Then, the centered waveform Ac and the centered waveform Bc are integrated to create a mixed interaction waveform ABt. Specifically, the interaction waveform ABt is created by integrating the values ​​of the centered waveform Ac and the centered waveform Bc at the same time.

[0042] Such an interaction waveform ABt is created using all of the integrated waveforms corresponding to the defects of the above-mentioned predetermined alloy products. Then, the created interaction waveform ABt is subjected to machine learning using a random forest to create a judgment model M, which is a learning model. Note that the machine learning method using the random forest using the interaction waveform ABt is a known method, so a detailed description will be omitted.

[0043] The determination model M using these state waveforms Aw and Bw is created for all combinations of waveforms included in the integrated waveform. That is, six types of determination models M are created: a determination model M in which the state waveform Aw is the vacuum waveform and the state waveform Bw is the injection velocity waveform, a determination model M in which the state waveform Aw is the vacuum waveform and the state waveform Bw is the rod pressure waveform, a determination model M in which the state waveform Aw is the vacuum waveform and the state waveform Bw is the head pressure waveform, a determination model M in which the state waveform Aw is the injection velocity waveform and the state waveform Bw is the rod pressure waveform, a determination model M in which the state waveform Aw is the injection velocity waveform and the state waveform Bw is the head pressure waveform, and a determination model M in which the state waveform Aw is the rod pressure waveform and the state waveform Bw is the head pressure waveform.

[0044] Next, the accuracy rate is obtained when the verification integrated waveform, which is the integrated waveform obtained in advance when the above-mentioned specified defects occur in the alloy product, is input into the six types of created judgment models M. Specifically, multiple verification integrated waveforms are provided for each defect in the alloy product. Of the state waveforms included in each verification integrated waveform, two types of state waveforms used to create the judgment model M are extracted, and the extracted state waveforms are input into the judgment model M.

[0045] Here, the judgment model M is configured to output a result of whether or not a predetermined defect is present when two types of state waveforms are input. Therefore, information on whether or not a predetermined defect is present, obtained by inputting two types of state waveforms extracted from the integrated waveform for verification, is stored in the judgment model M. This type of verification, in which two types of state waveforms extracted from the integrated waveform for verification are input to the judgment model M, is also performed on other integrated waveforms for verification when a predetermined defect occurs. Furthermore, similar verification is performed on other judgment models M.

[0046] The accuracy rate, which is the rate of correct answers as to whether or not a predetermined defect is present, obtained by verifying each judgment model M and each combination of two types of state waveforms, is calculated for each judgment model M and each combination of two types of state waveforms. Then, a judgment model M with an accuracy rate of 80% or more is obtained. As a result, a judgment model M that can accurately determine whether or not the above-mentioned predetermined defect is determined and stored in PC1. Note that the rate for comparing the accuracy rate is set to 80%, but this is not limited thereto, and the accuracy rate may be compared with a rate of 80% or more, or a rate of 80% or less.

[0047] A specific example of a determination model M that determines the presence or absence of a predetermined defect and a combination of two types of state waveforms input to the determination model M is a determination model M that determines the presence or absence of a "defective casting cavity," an "injection speed waveform," and a "vacuum waveform." By inputting the "injection speed waveform" and the "vacuum waveform" from the integrated waveform acquired from the molding device 70 into the determination model M, the presence or absence of a "defective casting cavity" can be determined with high accuracy.

[0048] Another example is a combination of a judgment model M that judges whether or not there is a "leakage defect," and an "injection speed waveform" and a "rod pressure waveform." By inputting the "injection speed waveform" and the "rod pressure waveform" from the integrated waveform acquired from the molding device 70 into such a judgment model M, it is possible to accurately judge whether or not there is a "leakage defect."

[0049] Alternatively, a different determination model M for determining the presence or absence of a "leakage defect" may be combined with the "head pressure waveform" and the "vacuum waveform." By inputting the "head pressure waveform" and the "vacuum waveform" from the integrated waveform acquired from the molding device 70 into such a determination model M, the presence or absence of a "leakage defect" can be determined with high accuracy.

[0050] As described above, in this embodiment, state waveforms Aw and Bw of two types of parts involved in the forming of an alloy product are acquired in the forming device 70, and centered waveforms Ac and Bc are calculated by centering these state waveforms Aw and Bw. Then, an interaction waveform ABt is created by integrating the centered waveform Ac and the centered waveform Bc, and a determination model M is created from the interaction waveform ABt by machine learning using a random forest.

[0051] Here, the interaction waveform ABt is obtained by integrating the centered waveforms Ac and Bc, so that the state changes of each part due to the phenomenon that causes the defect in the alloy product are reflected in the interaction waveform ABt. In particular, when the state of two or more parts changes simultaneously due to the phenomenon that causes the defect, those changes are reflected prominently in the interaction waveform ABt. In this way, by learning the interaction waveform ABt that reflects the state changes due to the defect in the alloy product at each part of the forming device 70, a judgment model M that can accurately detect defects in alloy products can be created.

[0052] Furthermore, centered waveforms Ac and Bc are used to generate the interaction waveform ABt. Here, the state waveforms Aw and Bw are waveforms acquired from different regions, and therefore the basic magnitude of their amplitudes (gain) also differs depending on the region. Therefore, if the state waveforms Aw and Bw are used as they are to generate the interaction waveform ABt, the state waveform with the larger basic magnitude of amplitude will have too much influence in the generated interaction waveform ABt, and conversely, the influence of the other state waveform on the interaction waveform ABt will be too small.

[0053] Therefore, by using centered waveforms Ac and Bc, which are obtained by subtracting the state waveforms Aw and Bw by their respective average values ​​Aav and Bav, to create the interaction waveform ABt, the difference in the influence of the amplitudes between the centered waveform Ac and the centered waveform Bct becomes smaller, and the created interaction waveform ABt can be a waveform that is equally influenced by the state waveforms Aw and Bw. Based on such an interaction waveform ABt, a judgment model M can be created that can more accurately detect defects in alloy products.

[0054] Next, the electrical configuration of PC 1 will be described with reference to Fig. 4. Fig. 4(a) is a block diagram showing the electrical configuration of PC 1. PC 1 has a CPU 10, a hard disk drive 11 (hereinafter abbreviated as "HDD 11"), and a RAM 12, which are each connected to an input / output port 14 via a bus line 13. The input / output port 14 is further connected to the above-mentioned display device 2, mouse 3, keyboard 4, and communication device 15.

[0055] The CPU 10 is a computing device that controls each unit connected by a bus line 13. The HDD 11 is a rewritable nonvolatile storage device that stores programs executed by the CPU 10, fixed value data, etc., and stores a control program 11a, a determination model table 11b, learning data 11c, and verification data 11d. When the control program 11a is executed by the CPU 10, the main processing of FIG. 5(a) is executed. The determination model table 11b, learning data 11c, and verification data 11d will be described with reference to FIGS. 4(b) to 4(d).

[0056] Fig. 4(b) is a diagram schematically showing the judgment model table 11b. As shown in Fig. 4(b), the judgment model table 11b stores the judgment model M created by the method of Fig. 3, in association with the type of detection defect and the designated waveform ("designated waveform 1" and "designated waveform 2" in the figure). The detectable defects store the types of defects in alloy products that the corresponding judgment model M can detect with high accuracy.

[0057] The designated waveform stores the type of state waveform to be input (designated) to the corresponding judgment model M when the corresponding judgment model M determines whether or not there is a detectable defect. In this embodiment, a state waveform of the same type as the state waveform used to create the judgment model M is input to the judgment model M when the judgment model M determines whether or not there is a detectable defect, so two types of designated waveforms (designated waveforms 1 and 2) are also stored in the judgment model table 11b.

[0058] A judgment model M stored in the judgment model table 11b is specified by the user, and an integrated waveform during molding of an alloy product is received from the data collecting device 20. A specified waveform of the judgment model M specified by the user is acquired from the judgment model table 11b, and a state waveform of the same type as the acquired specified waveform is extracted from the integrated waveform received from the data collecting device 20. Then, the extracted state waveform is input into the judgment model M specified by the user, thereby determining whether or not there is a detectable defect in the alloy product.

[0059] Next, the training data 11c will be described. FIG. 4(c) is a diagram schematically illustrating the training data 11c. The training data 11c stores training integrated waveforms, which are integrated waveforms acquired from the forming device 70 and used to create the determination model M. As shown in FIG. 4(c), the training data 11c stores, for each type of defect in the alloy product, an integrated waveform received from the data collection device 20 when the defect was confirmed during forming by the forming device 70. The determination model M is created using the type of defect in the alloy product and the training integrated waveform stored in the training data 11c using the method described above in FIG. 3.

[0060] Next, the verification data 11d will be described. FIG. 4(d) is a diagram schematically illustrating the verification data 11d. The verification data 11d stores a verification integrated waveform, which is an integrated waveform acquired from the created molding device 70 and is used to verify the determination model M described above in FIG. 3. As shown in FIG. 4(d), the verification data 11d stores a verification integrated waveform for each type of defect in the alloy product. The determination model M is verified using the type of defect in the alloy product and the verification integrated waveform stored in the verification data 11d using the method described above in FIG. 3. Note that the verification integrated waveform may be the integrated waveform received from the data collecting device 20 as is, or may be a waveform obtained by arbitrarily modifying the integrated waveform received from the data collecting device 20 by the user.

[0061] Returning to FIG. 4(a), the RAM 12 is a memory for rewritably storing various work data, flags, etc. when the CPU 10 executes the control program 11a. The communication device 15 is a device for communicating with external devices. In this embodiment, a data collecting device 20 is connected to the communication device 15, and the integrated waveform collected and created by the data collecting device 20 is received by the communication device 15.

[0062] Next, the processing executed by the CPU 10 of the PC 1 will be described with reference to Figures 5 to 8. Figure 5(a) is a flowchart of the main processing. The main processing is processing executed by the CPU 10 when an instruction to execute the control program 11a is input.

[0063] The main process first checks whether a command to create a determination model M has been input from the user via the mouse 3 or keyboard 4 (S1). If the input of a command to create a determination model M is confirmed in the process of S1 (S1: Yes), a model creation process (S2) is executed. The model creation process will be described with reference to Fig. 5(b) and Figs. 6 and 7.

[0064] Fig. 5(b) is a flowchart of the model creation process. The model creation process first acquires the type of defect in the alloy product that is the target of the judgment model M (S10). The type of defect in the alloy product acquired by the process of S10 may be acquired from defect information registered in the database when the integrated waveform is acquired in the data collection device 20, as described above in Fig. 3, or may be acquired by inputting it via the mouse 3 or keyboard 4 by the user. Hereinafter, the type of defect in the alloy product acquired by the process of S10 will be referred to as the "target defect type".

[0065] After the process of S10, a candidate model creation process (S11) is executed, and a model verification process (S12) is executed using the determination model M created in the candidate model creation process. The candidate model creation process and the model verification process will be described with reference to FIGS.

[0066] 6 is a flowchart of the candidate model creation process. The candidate model creation process is a process for creating first to sixth models, which are six types of judgment models M based on a target defect type and a combination of state waveforms. The candidate model creation process first acquires all of the learning integrated waveforms corresponding to the target defect type from the learning data 11c (S20). For example, if the target defect type is a "defective blowhole," all of the integrated waveforms Lw1, Lw2, Lw3, ... corresponding to the "defective blowhole" are acquired from the learning data 11c as the integrated learning waveforms.

[0067] After the process of S20, the vacuum waveform, injection velocity waveform, head pressure waveform, and rod pressure waveform are extracted from each of the acquired integrated waveforms (S21). After the process of S21, each state waveform acquired in the process of S20 is centered by the method described above in FIG. 3 (S22).

[0068] After the process of S22, an interaction waveform is created by multiplying the injection speed waveform and vacuum waveform centered in the process of S22 (S23). Specifically, of the injection speed waveform and vacuum waveform centered in the process of S22, those extracted from the same integrated waveform obtained in the process of S20 are obtained, and an interaction waveform is created by multiplying the obtained centered injection speed waveform and vacuum waveform by the method described above in Figure 3.

[0069] As a result, the number of interaction waveforms between the centered injection velocity waveform and the vacuum waveform is equal to the number of integrated waveforms obtained in the process of S20. Similar processes are also performed in the following processes of S25, S27, S29, S31, and S33.

[0070] After the process of S22, the created centered waveform is used to create a determination model M using the random forest method described above in Fig. 3 (S24). Hereinafter, the determination model M created from the interaction waveform between the injection velocity waveform and the vacuum waveform in this way will be referred to as the "first model."

[0071] After the process of S24, an interaction waveform is created by multiplying the centered injection velocity waveform and the head pressure waveform together in the process of S22 (S25). As a result, interaction waveforms of the centered injection velocity waveform and the head pressure waveform are created for the number of integrated waveforms obtained in the process of S20. After the process of S25, a determination model M is created using the created interaction waveform using the random forest technique described above in FIG. 3 (S26). Hereinafter, the determination model M created from the interaction waveform between the injection velocity waveform and the head pressure waveform in this way will be referred to as the "second model."

[0072] After the process of S26, an interaction waveform is created by multiplying the injection velocity waveform and rod pressure waveform centered in the process of S22 together (S27). As in the process of S23, interaction waveforms are created by the centered injection velocity waveform and rod pressure waveform, the number of which is equal to the number of integrated waveforms obtained in the process of S20, from the injection velocity waveform and rod pressure waveform centered in the process of S22. After the process of S27, a determination model M is created using the random forest method described above in FIG. 3 from the created interaction waveform (S28). Hereinafter, the determination model M created from the interaction waveform between the injection velocity waveform and the rod pressure waveform in this way will be referred to as the "third model."

[0073] After the process of S28, an interaction waveform is created by multiplying the centered vacuum waveform and the head pressure waveform in the process of S22 (S29). As a result, interaction waveforms of the centered vacuum waveform and the head pressure waveform are created for the number of integrated waveforms obtained in the process of S20. After the process of S29, a determination model M is created using the created interaction waveform using the random forest technique described above in Figure 3 (S30). Hereinafter, the determination model M created from the interaction waveform between the vacuum waveform and the head pressure waveform in this way will be referred to as the "fourth model."

[0074] After the processing of S30, an interaction waveform is created by multiplying the centered vacuum waveform and the rod pressure waveform in the processing of S22 (S31). As a result, interaction waveforms of the centered vacuum waveform and the rod pressure waveform are created for the number of integrated waveforms obtained in the processing of S20. After the processing of S31, a determination model M is created using the created interaction waveform using the random forest technique described above in Figure 3 (S32). Hereinafter, the determination model M created from the interaction waveform between the vacuum waveform and the rod pressure waveform in this way will be referred to as the "fifth model."

[0075] After the processing of S32, an interaction waveform is created by multiplying the head pressure waveform centered in the processing of S22 by the rod pressure waveform (S33). As a result, interaction waveforms of the centered head pressure waveforms and the rod pressure waveforms are created for the number of integrated waveforms obtained in the processing of S20. After the processing of S33, a determination model M is created using the created interaction waveforms using the random forest technique described above in FIG. 3 (S34). Hereinafter, the determination model M created from the interaction waveform between the head pressure waveform and the rod pressure waveform in this way will be referred to as the "sixth model." After the processing of S34, the candidate model creation process ends.

[0076] Next, the model verification process will be described with reference to Fig. 7. Fig. 7 is a flowchart of the model verification process. The model verification process is a process in which the verification integrated waveform of the verification data 11d corresponding to the target defect type is input to each of the first to sixth models created in the above model creation process, and the accuracy rate of each is calculated.

[0077] In the model verification process, first, an integrated verification waveform corresponding to the target defect type is obtained from the verification data 11d (S40). For example, if the target defect type is a "leakage defect," integrated waveforms Tw10, Tw11, ... are obtained as the integrated verification waveforms from the verification data 11d. After the process of S40, a vacuum waveform, an injection velocity waveform, a head pressure waveform, and a rod pressure waveform are extracted from each of the obtained integrated verification waveforms (S41).

[0078] After the process of S41, the injection speed waveform and vacuum waveform extracted for each acquired integrated verification waveform are input into a first model to determine whether or not there is a defect for each acquired integrated verification waveform (S42). After the process of S42, the accuracy rate for determining whether or not there is a defect based on the first model determined by the process of S42 is calculated (S43). Specifically, the accuracy rate for determining whether or not there is a defect for each acquired integrated verification waveform determined by the process of S41 is calculated by dividing the number of correct answers by the number of acquired integrated verification waveforms. Similar processes are also performed in the following processes of S45, S47, S49, S51, and S53.

[0079] After the processing of S43, the injection speed waveform and head pressure waveform are input into the second model for each of the acquired integrated waveforms for verification, and the presence or absence of defects is determined for each of the acquired integrated waveforms for verification (S44), and the accuracy rate of the second model for determining the presence or absence of defects by the processing of S44 is calculated (S45).

[0080] After the processing of S45, the injection velocity waveform and rod pressure waveform are input into the third model for each acquired integrated waveform for verification, and the presence or absence of defects is determined for each acquired integrated waveform for verification (S46), and the accuracy rate regarding the presence or absence of defects according to the third model processed in S46 is calculated (S47).

[0081] After processing S47, the vacuum waveform and head pressure waveform for each acquired integrated verification waveform are input into the fourth model, and the presence or absence of a defect is determined for each acquired integrated verification waveform (S48), and the accuracy rate for the presence or absence of a defect based on the fourth model processed in S48 is calculated (S49).

[0082] After processing S49, the vacuum waveform and rod pressure waveform for each acquired integrated waveform for verification are input into the fifth model, and the presence or absence of a defect is determined for each acquired integrated waveform for verification (S50), and the accuracy rate for the presence or absence of a defect according to the fifth model as a result of processing S50 is calculated (S51).

[0083] After the process of S51, the head pressure waveform and rod pressure waveform for each acquired integrated waveform for verification are input to the fifth model, the presence or absence of a defect is determined for each acquired integrated waveform for verification (S52), and the accuracy rate for the presence or absence of a defect according to the sixth model obtained by the process of S52 is calculated (S53). After the process of S53, the model verification process ends.

[0084] Returning to FIG. 5(b), after the model verification process in S12, it is confirmed whether any of the first to sixth models has an accuracy rate of 80% or more among the accuracy rates calculated in the model verification process (S13). If it is confirmed in the process of S13 that any of the first to sixth models has an accuracy rate of 80% or more (S13: Yes), the first to sixth models with an accuracy rate of 80% or more, the target defect type, and the type of state waveform used to create the first to sixth models are each saved in the judgment model table 11b (S14). At this time, the target defect type is saved as a detectable defect, and the type of state waveform used to create the first to sixth models is saved as a designated waveform in the judgment model table 11b.

[0085] As a result, a combination of a determination model M that can accurately determine the presence or absence of a detectable defect (with a correct answer rate of 80% or more) and a designated waveform corresponding to the determination model M is stored in the determination model table 11b.

[0086] On the other hand, if it is confirmed in the processing of S13 that there are no cases with a correct answer rate of 80% or more (S13: No), the parameters used to create the judgment model M in the random forest are adjusted (S15) in the processing of S24, S26, S28, S30, S32, and S34 in Figure 6, and the processing from S11 onwards is repeated.

[0087] Note that the parameters for creating the determination model M in the random forest are known parameters, and therefore detailed description thereof will be omitted. The parameters may be adjusted manually by the user, or may be automatically adjusted depending on, for example, the accuracy rate.

[0088] After the process of S14, the model creation process ends.

[0089] Returning to Fig. 5(a), in the process of S1, if the input of an instruction to create a judgment model M is not confirmed (S1: No), or after the model creation process of S2, it is confirmed whether an instruction to judge defects in alloy products has been input from the user via the mouse 3 or keyboard 4 (S3). In the process of S3, if the input of an instruction to judge defects in alloy products is confirmed (S3: Yes), the defect judgment process of S4 is executed. Here, the defect judgment process will be described with reference to Fig. 8.

[0090] 8 is a flowchart of the defect determination process. The defect determination process is a process for determining whether or not an alloy product is defective using the determination model M created in the model creation process described above and each state waveform detected during molding of the alloy product in the molding device 70, and displaying the result.

[0091] The defect determination process begins by displaying all combinations of the determination model M and the detectable defects of the determination model M, which are stored in the determination model table 11b, on the display device 2 (S60). After checking the combinations of the determination model M and the detectable defects displayed on the display device 2 by the process of S60, the user specifies the determination model M to be used for defect determination via the mouse 3 or keyboard 4.

[0092] After the process of S60, the judgment model M to be used for judging defects specified by the user is obtained from the judgment model table 11b (S61). After the process of S61, the integrated waveform detected during the molding of the alloy product is received from the data collection device 20 via the communication device 15 (S62).

[0093] After the process of S62, a state waveform of the same type as the designated waveform of the judgment model M acquired in the process of S61 is extracted from the integrated waveform received in the process of S62 (S63). Specifically, the designated waveform corresponding to the judgment model M acquired in the process of S61 is acquired from the judgment model table 11b, and a state waveform of the same type as the designated waveform is extracted from the integrated waveform received in the process of S62.

[0094] After the process of S63, the state waveform extracted by the process of S63 is input into the judgment model M acquired by the process of S61 to judge whether or not there is a detectable defect (S64), and the judgment result is displayed on the display device 2 (S65). One example of the display of the judgment result is to display a message saying "Defect Present" if it is judged that there is a detectable defect, and "No Defect" if it is judged that there is no detectable defect, but other displays may also be used. After the process of S65, the defect judgment process ends.

[0095] Returning to Fig. 5(a), if the input of an instruction to judge defects in the alloy product is not confirmed in the process of S3 (S3: No), or after the defect judgment process of S4, the processes from S1 onwards are repeated.

[0096] The present invention has been described above based on an embodiment, but the present invention is not limited to the above-described embodiment, and it can be easily inferred that various improvements and modifications are possible within the scope of the invention.

[0097] In the above embodiment, the interaction waveform is created by multiplying the centered waveform by two state waveforms, but this is not limiting. For example, the centered waveform may be the sum of two state waveforms, the centered waveform may be the division of two state waveforms, or the centered waveform may be a mixture of two state waveforms using other methods.

[0098] In the above embodiment, the interaction waveform is created using a centered waveform, but the waveform used for the interaction waveform is not limited to this. For example, the centered waveform may be created using the state waveform (before centering).

[0099] In the above embodiment, the centered waveform is calculated by subtracting the mean value of the state waveform from the state waveform, but this is not limiting. For example, the centered waveform may be calculated by subtracting the median value of the state waveform from the state waveform, or the maximum value of the state waveform from the state waveform. Alternatively, the centered waveform may be calculated by subtracting a fixed value set in advance by the user from the state waveform.

[0100] In the above embodiment, the determination model M is created by a random forest, but the method for creating the determination model M is not limited to this. For example, the determination model M may be created by other methods such as a decision tree, logistic regression (including regularization), naive Bayes, principal component analysis, k-nearest neighbor algorithm, k-means, FFNN (Feed Forward Neural Network), RNN (Recurrent Neural Network), GAN (Generative Adversarial Network), CNN (Convolutional neural network), MT method (Maharanobis-Taguchi System), or support vector machine, or the determination model M may be created by combining a plurality of the above methods.

[0101] In the above embodiment, two types of state waveforms are used when creating the judgment model M, but this is not limited thereto, and the judgment model M may be created using two or more types of state waveforms. Also, in the above embodiment, the state waveform for creating the judgment model M and the designated waveform input to the judgment model M when determining the presence or absence of a detectable defect are the same type of waveforms, but this is not limited thereto. For example, different state waveforms may be used as the state waveform for creating the judgment model M and the designated waveform input to the judgment model M.

[0102] 5(b) is stored in the determination model table 11b. However, this is not limited to this. For example, regardless of the accuracy rate, the determination model M with the highest accuracy rate may be stored in the determination model table 11b, or only the determination model M with both the accuracy rate of 80% or more and the highest accuracy rate may be stored in the determination model table 11b.

[0103] In the above embodiment, the state waveforms are four types: vacuum waveform, injection velocity waveform, head pressure waveform, and rod pressure waveform, but the state waveforms are not limited to these. For example, waveforms detected from the parts of the molding device 70 involved in molding the alloy product other than the vacuum waveform, injection velocity waveform, head pressure waveform, and rod pressure waveform may be added to the state waveforms.

[0104] In the above embodiment, the product for which the judgment model M judges whether or not there is a defect is an alloy product formed by die-casting using the molding device 70, but this is not limited to this. For example, the product for which the judgment model M judges whether or not there is a defect may be a resin molded product formed by injection molding, or may be a product manufactured by other manufacturing equipment such as processing equipment or parts assembly equipment. In this case, the judgment model M is created by the above-mentioned method using state waveforms of parts involved in the manufacturing of the product in these manufacturing equipment, and the state waveforms acquired from these manufacturing equipment are input into the created judgment model M to judge whether or not there is a defect in the product.

[0105] Furthermore, the presence or absence of defects in products manufactured in a chemical (manufacturing) plant may be determined using the judgment model M. In this case, too, the judgment model M is created using state waveforms of parts involved in the manufacture of products in the chemical (manufacturing) plant, and the presence or absence of defects in the products can be determined by inputting state waveforms acquired from the chemical (manufacturing) plant into the created judgment model M.

[0106] Alternatively, fault diagnosis and lifespan prediction of the target device may be performed using the judgment model M. For example, a judgment model M may be created using state waveforms based on measurable state data such as vibrations and noise of each part of the target device, and the state waveforms based on the state data acquired from the target device may be input into the created judgment model M, thereby performing fault diagnosis and lifespan prediction of the target device.

[0107] Furthermore, natural disasters may be predicted using the judgment model M. For example, a judgment model M may be created using a state waveform based on the flow rate of a river and vibrations around the river, and the state waveform based on the acquired flow rate of the river and vibrations around the river may be input into the created judgment model M to predict a river breach.

[0108] Alternatively, object detection may be performed using the judgment model M. For example, a judgment model M may be created using state waveforms based on two or more frequencies from radar such as aircraft radar, rain cloud radar, or fish finder, and the state waveforms based on two or more frequencies of the acquired radar may be input to the created judgment model M to detect the presence or absence of an object, the presence or absence of an object having a predetermined shape (e.g., a building), or even a predetermined individual (e.g., a specific person or animal).

[0109] In the above embodiment, the presence or absence of defects in the alloy product is determined in the defect determination process of S4 using the determination model M created in the model creation process of S2 in Fig. 5(a), but this is not limited to this. For example, the presence or absence of defects in the alloy product may be determined in the defect determination process of S4 using the determination model M created in a computer other than PC1. Furthermore, the determination model M created in the model creation process of S2 may be sent to a computer other than PC1, and that computer may execute a process equivalent to the defect determination process of S4 to determine the presence or absence of defects in the alloy product.

[0110] In the above embodiment, PC1 receives the integrated waveform created by the data collecting device 20 from the data collecting device 20 via the communication device 15, creates the judgment model M in the model creation process of S2, and judges whether the alloy product is defective in the defect judgment process of S4. However, the integrated waveform is not limited to being received from the data collecting device 20, and for example, PC1 may collect state waveforms from the forming device 70 and calculate the integrated waveform, or the forming device 70 itself may create the integrated waveform and transmit the created integrated waveform from the forming device 70 to PC1.

[0111] Furthermore, the acquisition of the integrated waveform from the data collection device 20 or the forming device 70 by the PC1 is not limited to via the communication device 15. For example, the integrated waveform created by the data collection device 20 or the forming device 70 may be stored in a portable storage device, the portable storage device may be connected to the PC1, and the integrated waveform may be acquired from the portable storage device.

[0112] In the above embodiment, PC1 receives an integrated waveform, which is a combination of state waveforms into one piece of data, from the data collecting device 20 and uses it to create the judgment model M and to determine whether or not the alloy product is defective, but this is not limited to this. For example, it is also possible to omit the creation of the integrated waveform by the data collecting device 20, and send the state waveforms of each part directly from the data collecting device 20 to PC1, and create the judgment model M and determine whether or not the alloy product is defective based on the state waveforms received by PC1.

[0113] In the above embodiment, the control program 11a is stored in the HDD 11 of the PC 1 and is configured to run on the PC 1. However, this is not necessarily limited to this, and the control program 11a may be configured to run on other computers such as a PC (personal computer), a mobile phone, a smartphone, or a tablet terminal. Furthermore, the control program 11a may be stored in a ROM or the like, and the present invention may be applied to a dedicated device (defect determination device) that executes only the control program 11a. [Explanation of symbols]

[0114] 1 PC (computer, defect judgment device) 11a Control program (model creation program, defect judgment program) 70 Molding equipment (manufacturing equipment) 71 Mold 76 Plunger Tip 77 Rod 79 Injection Cylinder 80 Injection piston (piston) Aw,Bw state waveform ABt interaction waveform S20 Waveform acquisition step S22 Centering step S23, S25, S27, S29, S31, S33 Interaction waveform creation steps S24, S26, S28, S30, S32, S34 Model creation steps S62 Acquisition means, acquisition step S64 Determination means, determination step

Claims

1. a waveform acquiring step of acquiring state waveforms based on transitions of states detected at two or more parts involved in the production of a product in a manufacturing apparatus; an alternating waveform creating step of creating an alternating waveform by mixing two or more state waveforms acquired in the waveform acquiring step; A model creation method characterized by comprising a model creation step of creating a judgment model for determining whether or not the product is defective by machine learning the interaction waveform created in the interaction waveform creation step.

2. 2. The model creation method according to claim 1, wherein the interaction waveform creation step creates an interaction waveform by multiplying two or more state waveforms acquired in the waveform acquisition step.

3. The interactive waveform creating step includes: a centering step of performing centering for each of the two or more state waveforms acquired in the waveform acquiring step by subtracting each value in the state waveform by an average value of the state waveform; 2. The model creation method according to claim 1, wherein the state waveforms centered in the centering step are used to create interaction waveforms.

4. 2. The model creation method according to claim 1, wherein the model creation step creates a determination model by a random forest using the interaction waveform created in the interaction waveform creation step.

5. the manufacturing apparatus is a molding apparatus that manufactures the product by molding a material using a mold, the waveform acquisition step acquires, as state waveforms, an injection speed waveform based on a change in speed of a rod connected to a plunger tip that pushes the material into the mold, and a vacuum waveform based on a change in pressure inside the mold; the alternating waveform creating step creates an alternating waveform by mixing the injection velocity waveform and the vacuum waveform acquired in the waveform acquiring step, 5. The model creation method according to claim 1, wherein the model creation step creates a determination model for determining whether or not there is a casting cavity defect in the product by machine learning the interaction waveform created in the interaction waveform creation step.

6. the manufacturing apparatus is a molding apparatus that manufactures the product by molding a material using a mold, the waveform acquisition step acquires, as state waveforms, an injection velocity waveform based on a transition in velocity of a rod connected to a plunger tip that pushes the material into the mold, and a rod pressure waveform based on a transition in pressure on the rod side in an injection cylinder that is connected to the rod and in which a piston that pushes the rod is disposed; the interactive waveform creating step creates an interactive waveform by mixing the injection velocity waveform and the rod pressure waveform acquired in the waveform acquiring step, 5. The model creation method according to claim 1, wherein the model creation step creates a judgment model for determining whether or not a leakage defect exists in the product by machine learning the interaction waveform created in the interaction waveform creation step.

7. the manufacturing apparatus is a molding apparatus that manufactures the product by molding a material using a mold, the waveform acquisition step acquires, as state waveforms, a head pressure waveform based on a transition of pressure on a side opposite to a rod in an injection cylinder in which a piston that pushes the rod is disposed and which is connected to a rod that is connected to a plunger tip that pushes the material into the mold, and a vacuum waveform based on a transition of pressure inside the mold; the alternating waveform creating step creates an alternating waveform by mixing the head pressure waveform and the vacuum waveform acquired in the waveform acquiring step, 5. The model creation method according to claim 1, wherein the model creation step creates a judgment model for determining whether or not a leakage defect exists in the product by machine learning the interaction waveform created in the interaction waveform creation step.

8. A model creation program that causes a computer to execute a model creation process to create a determination model for determining whether or not a product is defective, a waveform acquiring step of acquiring state waveforms based on transitions of states detected at parts involved in the production of the product in the manufacturing apparatus, the state waveforms being detected at two or more of the parts; an alternating waveform creating step of creating an alternating waveform by mixing two or more state waveforms acquired in the waveform acquiring step; A model creation program that causes the computer to execute a model creation step of creating the judgment model by machine learning the interaction waveform created in the interaction waveform creation step.

9. an acquisition means for acquiring state waveforms based on transitions of states detected at parts involved in the manufacture of a product in a manufacturing device, the state waveforms being detected at two or more of the parts; A defect judgment device characterized by comprising a judgment means for judging whether or not the product is defective using two or more state waveforms acquired by the acquisition means and a judgment model created by the model creation method described in claim 1.

10. A defect determination program that causes a computer to execute a defect determination process for determining whether or not a product is defective, an acquiring step of acquiring state waveforms based on transitions of states detected at parts involved in the production of a product in a manufacturing apparatus, the state waveforms being detected at two or more of the parts; A defect judgment program characterized by causing the computer to execute a judgment step of judging whether or not the product is defective using two or more state waveforms acquired in the acquisition step and a judgment model created by the model creation method described in claim 1.

Images