Experiment supporting device, experiment supporting method, and program
The experiment support device and method efficiently collect machining data using a trial processing database and machine learning techniques, addressing inefficiencies in existing methods by ensuring optimal conditions are met with reduced costs and time.
Patent Information
- Application Number
- JP2023072144
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-26
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing methods for obtaining learning data for machine learning in machining processes face inefficiencies, as they either require numerous trial machinings leading to high economic and time costs, or fail to converge to optimal solutions due to insufficient data, necessitating a more efficient data collection method.
An experiment support device and method that utilizes a trial processing database, a trial machining condition search unit, a similarity judgment unit, and a termination processing unit to efficiently collect data by searching for similar input variables and terminating data generation when predetermined conditions are met, using machine learning techniques like Bayesian optimization.
This approach allows for the efficient collection of necessary data, reducing economic and time costs while ensuring optimal machining conditions are achieved by focusing on data similarity and convergence to optimal solutions.
Smart Images

Figure 2024157666000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to an experiment support device, an experiment support method, and a program. [Background technology]
[0002] In cutting processing using a processing device, in order to ensure a certain level of quality or to reduce economic or time costs, it is necessary to adjust processing conditions such as the position where the tool cuts into the workpiece, the rotation speed and feed speed of the tool, etc. In order to obtain such processing conditions, a method of estimating optimal processing conditions by machine learning is known (for example, Patent Document 1).
[0003] In order to estimate optimal machining conditions using machine learning, a large amount of learning data is required, and experiments must be conducted to actually machine workpieces under various machining conditions using the machining equipment. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2019-185591 A Summary of the Invention [Problem to be solved by the invention]
[0005] In such experiments, if there is a small amount of learning data, machine learning may not converge to an optimal solution, or the number of trial machining runs may not be estimated in advance. On the other hand, if a large amount of learning data is to be obtained, the number of trial machining runs using the actual machining device will increase, resulting in economic and time costs. For this reason, it is desirable to collect the necessary data efficiently.
[0006] This problem is not limited to the case of obtaining learning data for processing conditions of a processing device, but is common to a wide range of acquisition of learning data involving experiments.
[0007] The present disclosure has been made in consideration of the above-mentioned circumstances, and aims to provide an experiment support device, an experiment support method, and a program for supporting an experiment plan for efficiently collecting data. [Means for solving the problem]
[0008] In order to achieve the above-mentioned objective, the experimental support device disclosed herein includes a trial processing database that stores combinations of input variables in experiments to obtain learning data for machine learning in association with predetermined evaluation values; a trial processing condition search unit that searches for the next combination of input variables based on the combinations of input variables stored in the trial processing database; a similarity determination unit that calculates evaluation values indicating the similarity between the combinations of input variables searched for by the trial processing condition search unit and each of the combinations of input variables stored in the trial processing database and adds them to the trial processing database; and a termination processing unit that terminates the generation of input data for learning when the evaluation values calculated by the similarity determination unit satisfy predetermined conditions. Effect of the Invention
[0009] According to the present disclosure, when the evaluation value obtained by the similarity determination unit satisfies a predetermined condition, the generation of input data for learning is terminated, thereby assisting in experimental planning for efficiently collecting data in order to obtain optimal processing conditions for the processing device. [Brief description of the drawings]
[0010] [Figure 1] A diagram showing a scene in which an end mill of a processing device cuts a workpiece. [Diagram 2] FIG. 1 is a diagram showing a configuration of a processing condition adjustment system according to a first embodiment; [Diagram 3] A diagram showing the hardware configuration of the experiment support device and the machine learning device. [Figure 4] FIG. 13 is a diagram showing an example of a constraint condition setting unit. [Diagram 5]FIG. 1A is a diagram showing an initial state of a trial machining database according to a first embodiment; FIG. 1B is a diagram showing a state in which data is stored in the trial machining database according to the first embodiment; and FIG. 1C is a diagram showing a state in which a certain amount of data has been stored in the trial machining database according to the first embodiment. [Figure 6] Flowchart showing experiment support processing according to the first embodiment [Figure 7] FIG. 13 is a diagram showing a configuration of a processing condition adjustment system according to a second embodiment. [Figure 8] FIG. 13 is a diagram showing a trial machining database according to a second embodiment; [Figure 9] Flowchart showing an experiment support process according to the second embodiment [Figure 10] A diagram showing an example of variance evaluation of the objective variable for each input variable [Figure 11] FIG. 13 is a diagram showing a configuration of a processing condition adjustment system according to a third embodiment. [Figure 12] Flowchart showing an experiment support process according to the third embodiment DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0011] Hereinafter, the experiment support device 100 and the processing condition adjustment system 1 according to the embodiment of the present disclosure will be described in detail with reference to the drawings. Note that the present embodiment is for the purpose of explanation and does not limit the scope of the present disclosure. Therefore, a person skilled in the art can adopt an embodiment in which each or all of these elements are replaced with an equivalent, and such an embodiment is also included in the scope of the present disclosure.
[0012] (First embodiment) The following describes an example of a problem in which a workpiece 220 made of a material such as a metal or alloy is cut by an end mill 210 attached to a processing device as shown in Fig. 1. For example, the workpiece 220 is cut to produce a fixed scroll and an oscillating scroll that constitute a scroll compressor mounted in a multi-air conditioner for a building or a packaged air conditioner.
[0013] In this machining device, machining conditions can be set such as the spindle rotation speed A, which is the number of rotations per unit time of the end mill 210, the feed speed B, which is the speed at which the end mill 210 moves, the radial cutting depth C, which is the depth at which the end mill 210 cuts horizontally into the workpiece 220, and the axial cutting depth D, which is the depth at which the end mill 210 cuts vertically into the workpiece 220.
[0014] In this embodiment, a combination of machining conditions for minimizing the surface roughness, which indicates the uneven state of the cut surface 221 of the workpiece, is adjusted by machine learning. For this purpose, trial machining is performed multiple times under different machining conditions, and an experiment is performed to obtain the surface roughness after each trial machining.
[0015] As shown in Figure 2, the machining condition adjustment system 1 of this embodiment includes an experiment support device 100 that supports experimental planning, a machining device 200 that machines a workpiece 220 using an end mill 210, and a machine learning device 300 that estimates machining conditions through machine learning.
[0016] The experiment support device 100 supports the experiment plan by determining the processing conditions for the trial processing, acquiring data from the processing device 200, and deciding whether to end the experiment, etc. The processing device 200 acquires the processing conditions for the trial processing from the experiment support device 100, and after performing the trial processing under the processing conditions, acquires the surface roughness of the cut surface 221 after the trial processing by a sensor or the like provided in the processing device 200, and transmits it to the experiment support device 100.
[0017] The machine learning device 300 acquires experimental data from the experiment support device 100, and estimates machining conditions that minimize surface roughness by machine learning such as reinforcement learning. In reinforcement learning, in which an agent takes action in an environment and repeatedly receives a reward from the environment according to the result, the machine learning device 300 gives a larger reward to the agent as the objective variable becomes smaller, learns the reward to be given to the surface roughness for the machining conditions, which are actions, for each trial machining, and adjusts the action that maximizes the reward.
[0018] 3 is a diagram showing the hardware configuration of the experiment support device 100 and the machine learning device 300. The experiment support device 100 and the machine learning device 300 each include an input unit 10, a display unit 20, a control unit 30, a main memory unit 40, an auxiliary memory unit 50, and a communication unit 60, which are connected to each other via a bus 70.
[0019] The input unit 10 includes a mouse, a keyboard, etc. The display unit 20 includes a liquid crystal display, an organic EL (Electro Luminescence) display, etc.
[0020] The control unit 30 includes a CPU, which is, for example, an integrated circuit. The control unit 30 executes a program stored in the auxiliary storage unit 50 to realize various functions of the experiment support device 100 and the machine learning device 300, and executes the above-mentioned processes.
[0021] The main memory unit 40 includes a RAM (Random Access Memory). Programs are loaded into the main memory unit 40 from the auxiliary memory unit 50. The main memory unit 40 is used as a working area for the control unit 30.
[0022] The auxiliary storage unit 50 includes a non-volatile memory such as an EEPROM (Electrically Erasable Programmable Read-Only Memory), an SSD (Solid State Drive), and an HDD (Hard Disk Drive). In addition to programs, the auxiliary storage unit 50 stores various data used for processing by the control unit 30. The auxiliary storage unit 50 supplies the control unit 30 with data used by the control unit 30 in accordance with instructions from the control unit 30, and stores the data supplied from the control unit 30.
[0023] The communication unit 60 includes a network interface circuit for transmitting and receiving Ethernet frames from an external device. The communication unit 60 receives a signal from the outside and outputs data indicated by the signal to the control unit 30. The communication unit 70 also transmits a signal indicating the data output from the control unit 30 to the external device.
[0024] The input unit 10 and the display unit 20 may be configured as separate units from the experiment support device 100 and the machine learning device 300. For example, the input unit 10 and the display unit 20 may be connected to the experiment support device 100 and the machine learning device 300 from a control box of the experiment support device 100 and the machine learning device 300 via an extension cable or the like.
[0025] Furthermore, the functions of the experiment support device 100 and the machine learning device 300 can be realized by dedicated hardware or by a general computer system.
[0026] For example, the program executed by the control unit 30 may be stored in a non-transitory computer-readable recording medium, distributed, and the program may be installed on a computer to configure the experiment support device 100 and the machine learning device 300. Examples of such recording media include flexible disks, CD-ROMs (Compact Disc Read-Only Memory), DVDs (Digital Versatile Discs), and MOs (Magnetro-Optical Discs).
[0027] Also, the program may be stored in a disk device of a server device on a communication network such as the Internet, and may be downloaded to a computer, for example, by being superimposed on a carrier wave.
[0028] In addition, the experiment support process and the process by the machine learning device 300 described below can also be achieved by starting and executing a program while transferring it via a communication network.
[0029] Furthermore, the experiment support processing and processing by the machine learning device 300 can also be achieved by executing all or part of the program on a server device and executing the program while the computer sends and receives information regarding the processing via a communications network.
[0030] In addition, when the experimental support processing and the processing by the machine learning device 300 are shared and performed by the OS (Operating System) or by the OS working together with an application, only the parts other than the OS may be stored on a medium and distributed, or may be downloaded to a computer.
[0031] Furthermore, the means for realizing the functions of the experiment support device 100 and the machine learning device 300 are not limited to software, and some or all of them may be realized by dedicated hardware including circuits.
[0032] Returning to Figure 2, the experiment support device 100 includes a constraint condition setting unit 110 that sets constraint conditions such as the range of input variables, which are processing conditions, a trial processing database 120 that stores data regarding each trial processing, a trial processing condition search unit 130 that searches for input variables for the trial processing, a normalization processing unit 140 that normalizes or inversely converts the input variables, a similarity determination unit 150 that determines the similarity of the input variables, and a termination processing unit 160 that terminates the experiment.
[0033] The constraint condition setting unit 110 receives from the user the input variables and objective variables to be used in the experiment, as well as the constraint conditions in the experiment, and sets each of them as shown in Fig. 4. Here, the input variables refer to the machining conditions in each trial machining, and in this embodiment, they are the spindle speed A, feed rate B, radial cut-in amount C, and axial cut-in amount D of the end mill 210. The objective variable refers to the result of trial machining under certain input conditions, and in this embodiment, they are the surface roughness E of the cutting surface 221.
[0034] In the example shown in FIG. 4, an item type 111, an item 112, and a constraint condition 113 are set in the constraint condition setting unit 110. Here, the item type “input variable” is set to the item “spindle speed A [min -1]," "Feed rate B [mm / min]," "Radial cutting depth C [mm]," and "Axial cutting depth D [mm]" are set, and a range including the set lower limit and set upper limit value is set as a constraint condition for each. In addition, the item "Surface roughness E" is set as the item type "Objective variable." Since the objective variable is obtained as a result of trial machining, no particular constraint condition is set.
[0035] Furthermore, as the item type "Other", the items "Feed per blade F" and "Number of trials G" are set. The feed per blade F is, Feed per blade F [mm / rev] = Feed speed B [mm / min] / (Number of teeth × Spindle speed A [min -1 ]) The constraint is set so that the number of trials G is smaller than the set value set by the user. The number of trials G is the number of trial machinings to be performed in the experiment, and the constraint is set so that the number of trials G is equal to or smaller than the set value set by the user, that is, so that the number of trials G does not exceed the upper limit.
[0036] Returning to Fig. 2, the trial machining database 120 stores, for each of a plurality of trial machinings included in an experiment for adjusting the machining conditions in the machining device 200, input variables, which are machining conditions in the trial machining, and objective variables, which are the results of the trial machining under those machining conditions, in association with each other. As shown in Figs. 5(A) to (C), in the trial machining database 120, for each trial machining, a trial machining ID for identifying the trial machining, each input variable, objective variable, and data similarity are associated with each other. The data similarity is an index showing whether the input variables of the trial machining are similar to the input variables of past trial machining, and is set to "1" if they are similar, and "0" if they are not similar.
[0037] Returning to FIG. 2, the trial machining condition search unit 130 searches for input variables in the next trial machining to be attempted based on the input variables and the target variables in the past trial machining stored in the trial machining database 120. For example, Bayesian optimization is used as the algorithm of the trial machining condition search unit 130. By using Bayesian optimization, it becomes possible to focus on searching for an area where the target variable is likely to be minimized based on past information, rather than blindly searching for input variables. Acquisition functions for Bayesian optimization include PI (Probability of Improvement), EI (Expected Improvement), LCB (Lower Confidence Bound), etc.
[0038] Each time one trial machining is completed, the trial machining condition searching unit 130 searches for input variables in the next trial machining to be attempted, and adds the searched input variables to the trial machining database 120 via the normalization processing unit 140 described later.
[0039] The normalization processing unit 140 normalizes the input variables and the objective variables stored in the trial machining database 120, or inversely converts the input variables searched by the trial machining condition searching unit 130 and stores them in the trial machining database 120. The normalization processing unit 140 normalizes the past input variables and the objective variables stored in the trial machining database 120 to numerical values within a predetermined range. Here, normalization means matching the ranges of the input variables and the objective variables, for example, converting the input variables and the objective variables to a range of 0≦(input variable or objective variable)≦1. Inverse conversion means performing the reverse process of normalization on the results searched by the trial machining condition searching unit 130. Note that the input variables can be normalized by dividing them by a set upper limit value set in the constraint condition setting unit 110, and the objective variables can be normalized by dividing them by a predetermined value.
[0040] The similarity determination unit 150 calculates the similarity between the input variables in the current trial machining and the input variables in each of the past trial machining, and determines whether the input variables in the current trial machining are similar to the input variables in any of the past trial machining. The similarity is determined by calculating, for example, the Euclidean distance and determining whether it exceeds a predetermined threshold value, as described later.
[0041] The termination processing unit 160 terminates the experiment when the result of the determination by the similarity determining unit 150 satisfies a predetermined condition.
[0042] Fig. 6 is a flow chart for explaining the experiment support processing executed by the experiment support device 100. In the experiment for adjusting the processing conditions of the cutting processing by the processing device 200, the experiment support processing proposes processing conditions for the next trial processing that are expected to produce optimal results based on data from past trial processing, and ends the experiment when sufficient data has been collected. The experiment support processing will be explained below with reference to Fig. 6.
[0043] The experiment support process starts, for example, when an application or program for the experiment support process is launched. When the experiment support process starts, first, the constraint condition setting unit 110 accepts settings of input variables, objective variables, etc., shown in the example of Fig. 4, from the user, and determines whether or not the constraint conditions have been input (step S101). The experiment support process is in a standby state until the constraint condition setting unit 110 determines that the constraint conditions have been input by the user (No in step S101).
[0044] When the constraint condition setting unit 110 determines that the user has input a constraint condition (Yes in step S101), the trial machining database 120 is created (step S102). At this stage, as shown in FIG. 5(A), none of the items have been set.
[0045] Next, steps S103 to S111 are loop processes, which are repeated for each trial machining ID until a predetermined condition is satisfied. When the loop process starts (step S103), the normalization processing unit 140 first normalizes all input variables and objective variables in past trial machining stored in the trial machining database 120 (step S104). For the first trial machining, there is no past data, so step S104 is skipped.
[0046] Next, the trial machining condition searching unit 130 searches for a combination of input variables by a preset optimization process, for example, Bayesian optimization, based on the past input variables and the objective variable normalized by the normalization processing unit 140 (step S105). The trial machining condition searching unit 130 searches for input variables in the trial machining to be attempted next within the normalized numerical values, for example, within the range of 0 to 1 mentioned above. Note that, since there is no past data for the first trial machining, for example, a search may be performed from the edge of a Euclidean space with each input variable as one dimension.
[0047] When the trial machining condition searching unit 130 searches for a combination of each input variable, the normalization processing unit 140 inversely converts the searched combination of input variables (step S106). Then, the normalization processing unit 140 adds the inversely converted combination of input variables to the trial machining database 120. In other words, the trial machining database 120 stores raw data for the machining operation of the machining device 200.
[0048] For example, when the combination of the inversely transformed input variables for the second trial machining is added to the trial machining database 120, it becomes as shown in the example in Fig. 5(B). The experiment support device 100 transmits a trial machining instruction specifying the combination of the added input variables to the machining device 200 (step S107).
[0049] Thereafter, the processing device 200 receives a trial processing instruction from the experiment support device 100, performs trial processing under input conditions specified in the trial processing instruction, obtains the surface roughness which is the objective variable, and then transmits a response specifying the objective variable to the experiment support device 100. The experiment support device 100 determines whether or not a response specifying the objective variable has been obtained from the processing device 200 (step S108). The experiment support process is in a standby state until the experiment support device 100 determines that a response specifying the objective variable has been obtained from the processing device 200 (No in step S108).
[0050] When the experiment support device 100 determines that a response specifying an objective variable has been acquired from the machining device 200 (Yes in step S108), the objective variable specified in the response is stored in the trial machining database 120. After that, the similarity determination unit 150 calculates the similarity between the input variables in the current trial machining and the input variables in each of all the past trial machining (step S109). The similarity is calculated by the Euclidean distance L between the input variables (Ai, Bi, Ci, Di) in the i-th trial machining of the current trial machining and the input variables (Aj, Bj, Cj, Dj) in the j-th past trial machining. ij The Euclidean distance L ij teeth, L ij =√{(Ai-Aj) 2 +(Bi-Bj) 2 +(Ci-Cj) 2 +(Di-Dj) 2} j=1~(i-1) It can be calculated as follows.
[0051] The similarity determination unit 150 uses the Euclidean distance L ijis equal to or less than a predetermined threshold, the input variables (Ai, Bi, Ci, Di) in the current trial machining which is the i-th trial machining are similar to the input variables (Aj, Bj, Cj, Dj) in the past trial machining which is the j-th trial machining, and if not, the input variables are not similar. The similarity determination unit 150 stores "1" in the data similarity of the trial machining database 120 when it determines that the input variables in the current trial machining are similar to the input variables in any of the past trial machining, and stores "0" when it determines that the input variables in the current trial machining are not similar to the input variables in any of the past trial machining. The similarity determination unit 150 also stores "0" in the data similarity of the first trial machining. By repeating the loop process of steps S103 to S111, the trial machining database 120 as shown in FIG. 5(C) is created.
[0052] Next, the end processing unit 160 judges whether the current state satisfies a predetermined condition (step S110). The end processing unit 160 may judge that the current state satisfies the predetermined condition when the data similarity "1" appears consecutively for a predetermined number of times or more (Yes in step S110). That is, in this case, when the similarity judgment unit 150 judges that the input variables in the current trial machining are similar to the input variables in any of the past trial machinings consecutively for a predetermined number of times or more (Yes in step S110), the end processing unit 160 exits the loop and ends the experiment (step S112). For example, if the current trial machining database 120 is in a state as shown in the example of FIG. 5(C), the similarity "1" is not consecutive, so the experiment continues.
[0053] Alternatively, the termination determination unit 160 may determine that the current state satisfies the predetermined condition when the data similarity "1" has appeared a predetermined number of times or more up to now (Yes in step S110). That is, in this case, when the similarity determination unit 150 determines that the input variables in the current trial machining are similar to the input variables in any of the past trial machinings a predetermined number of times or more (Yes in step S110), the termination processing unit 160 exits the loop and terminates the experiment (step S112).
[0054] Alternatively, when the number of trial machining operations exceeds the upper limit number of trials G set by the constraint condition setting unit 110 shown in the example of Fig. 4, the end determination unit 160 determines that the predetermined condition is satisfied (Yes in step S110), exits the loop, and ends the experiment (step S112). When the experiment ends, the experiment support process ends.
[0055] On the other hand, if the end processing unit 160 determines that the current state does not satisfy the predetermined condition (No in step S110), the process proceeds to step S111, and steps S103 to S111 are repeated for the next trial machining.
[0056] The contents of the trial machining database 120 obtained by the experiment support process are transmitted to the machine learning device 300. The machine learning device 300 estimates machining conditions, which are a combination of input variables, by, for example, reinforcement learning, so that the surface roughness E, which is the objective variable, is minimized. Since the learning data is selected by the above-mentioned method, the amount of learning data is neither too small nor too large, and appropriate learning is possible. In addition, the optimal machining conditions may be estimated by extracting a combination of input variables when the value of the surface roughness E, which is the objective variable of the trial machining database 120 obtained by the experiment support process, is minimized.
[0057] As described above, according to the first embodiment, the experiment support device 100 ends an experiment in which trial machining is performed under various machining conditions searched for by machine learning such as Bayesian optimization, when a certain amount of highly similar data is obtained. This allows the data required to obtain optimal machining conditions to be collected in a focused manner, without collecting too much similar data, thereby reducing economic and time costs and allowing efficient data collection.
[0058] Moreover, according to the first embodiment, before the trial machining condition searching unit 130 searches for machining conditions for the next trial machining by Bayesian optimization or the like, data from the past trial machining is normalized. This makes it easier for the machining conditions searched for by the trial machining searching unit 130 to converge to an optimal solution, which leads to reducing the cost of experiments.
[0059] Second embodiment Next, a second embodiment of the present disclosure will be described. The same reference numerals are used for configurations and processes equivalent to those of the first embodiment. In the first embodiment, the machining conditions searched for by Bayesian optimization or the like are added to the trial machining database every time trial machining is performed, but in the second embodiment, the trial machining database is created in advance by an experimental design method.
[0060] As shown in FIG. 7, the experiment support device 101 includes a trial machining database 121 instead of the trial machining database 120 in the first embodiment, and further includes a database creation unit 170 that creates the trial machining database 121. The trial machining creation unit 170 creates the trial machining database 121 shown in the example of FIG. 8 in advance by an experimental design method. Here, it is assumed that an L9 orthogonal table with four factors and three levels is created. Each of the four factors, the spindle speed A, the feed rate B, the diameter cutting amount C, and the axial cutting amount D, stores three levels of values, (a1, a2, a3), (b1, b2, b3), (c1, c2, c3), and (d1, d2, d3). In the L9 orthogonal table, the same number of three levels of values appear in the columns of each input variable.
[0061] For example, the lower limit, median, and upper limit of the range of the constraint condition 113 set for each input variable by the constraint condition setting unit 110 are selected as the values of these three levels. That is, the lower limit of the constraint condition 113 may be selected for a1, b1, c1, and d1, the average value of the constraint condition 113 may be selected for a2, b2, c2, and d2, and the upper limit of the constraint condition 113 may be selected for a3, b3, c3, and d3.
[0062] 9 is a flowchart illustrating the experiment support processing executed by the experiment support apparatus 101. The experiment support processing will be described below with reference to FIG.
[0063] The experiment support process starts, for example, when an application or program for the experiment support process is launched. When the experiment support process starts, first, the constraint condition setting unit 110 accepts settings of input variables, objective variables, etc., shown in the example of Fig. 4, from the user, and determines whether or not the constraint conditions have been input (step S201). The experiment support process is in a standby state until the constraint condition setting unit 110 determines that the constraint conditions have been input by the user (No in step S201).
[0064] When the constraint condition setting unit 110 determines that the constraint condition has been input by the user (Yes in step S201), the database creation unit 170 creates the trial machining database 121 by an algorithm of the design of experiments (step S202).
[0065] Subsequently, steps S203 to S207 are loop processes, which are repeated for each trial processing ID until a predetermined condition is satisfied. When the loop process starts (step S203), first, the experiment support device 101 transmits a trial processing instruction specifying a combination of input variables linked to the first trial processing ID to the processing device 200 (step S204).
[0066] Thereafter, the processing device 200 receives a trial processing instruction from the experiment support device 101, performs trial processing under input conditions specified in the trial processing instruction, obtains the surface roughness which is the objective variable, and then transmits a response specifying the objective variable to the experiment support device 101. The experiment support device 101 determines whether or not a response specifying the objective variable has been obtained from the processing device 200 (step S205). The experiment support process is in a standby state until the experiment support device 101 determines that a response specifying the objective variable has been obtained from the processing device 200 (No in step S205).
[0067] On the other hand, when the experiment support device 101 determines that a response specifying an objective variable has been acquired from the processing device 200 (Yes in step S205), the objective variable specified in the response is stored in the trial processing database 121, and then the end determination unit 160 determines whether or not the current state satisfies a predetermined condition (step S206). For example, when the number of trial processings exceeds the upper limit number of trials G set by the constraint condition setting unit 110 shown in the example of Fig. 4, the end determination unit 160 determines that the predetermined condition has been satisfied (Yes in step S206), exits the loop, and ends the experiment (step S208). When the experiment ends, the experiment support process ends.
[0068] On the other hand, if the end processing unit 160 determines that the current state does not satisfy the predetermined condition (No in step S206), the process proceeds to step S207, and steps S203 to S207 are repeated for the next trial machining.
[0069] The contents of the trial machining database 121 obtained by the experimental processing in this manner may be transmitted to the machine learning device 300, and machining conditions, which are a combination of input variables, may be estimated by machine learning so as to minimize the surface roughness E, which is the objective variable.
[0070] Here, it is assumed that the contents of the trial machining database 121 are subjected to a variance analysis, and the results shown in Fig. 10 are obtained. Here, it is assumed that the factors that have a strong influence on minimizing the surface roughness, which is the objective variable, are the spindle speed A and the radial cutting depth C. This allows the user to determine that when conducting another experiment, it is acceptable to narrow down the spindle speed A and the radial cutting depth C to conduct the experiment and machine learning using the machine learning device 300, and thus the results can be used as pre-learning.
[0071] As described above, according to the second embodiment, the experiment is terminated when a predetermined condition is met, such as exceeding the upper limit of the number of trial machining attempts set by the user. This makes it possible to narrow down the items to be considered in a re-experiment while avoiding a situation in which the number of trial machining attempts cannot be estimated in advance, and to collect data efficiently.
[0072] Third embodiment Next, a third embodiment of the present disclosure will be described. The same reference numerals are used for configurations and processes equivalent to those of the first and second embodiments. In the first embodiment, whether or not to end the experiment was determined based on whether a certain amount of data with high similarity was obtained, but in the third embodiment, the determination is made based on whether or not the optimal solutions derived by the experiment support device 102 and the machine learning device 300 respectively match.
[0073] As shown in FIG. 11, the experiment support device 102 includes an extraction unit 180 that extracts, from the trial machining database 120, a combination of input variables in which the objective variable satisfies a predetermined condition. Moreover, the machine learning device 300 acquires the trial machining database 120 from the experiment support device 102, and performs reinforcement learning or the like to estimate a combination of input variables that results in a response variable satisfying a predetermined condition. The predetermined condition is, for example, that the surface roughness E, which is the response variable, takes a minimum value.
[0074] The termination processing unit 160 terminates the experiment when the combination of input variables extracted by the extraction unit 180 matches the combination of input variables estimated by the machine learning device 300. Here, "match" does not only mean perfect match, but also includes a certain degree of match, that is, the concept of "similarity." Whether or not they are similar is determined by a certain similarity calculation, such as calculating the Euclidean distance between the combinations of two input variables and determining whether or not it exceeds a predetermined threshold.
[0075] 12 is a flowchart illustrating the experiment support processing executed by the experiment support device 102. The experiment support processing will be described below with reference to FIG.
[0076] The experiment support process starts, for example, when an application or program for the experiment support process is launched. When the experiment support process starts, first, the constraint condition setting unit 110 accepts settings of input variables, objective variables, etc., shown in the example of Fig. 4, from the user, and determines whether or not the constraint conditions have been input (step S301). The experiment support process is in a standby state until the constraint condition setting unit 110 determines that the constraint conditions have been input by the user (No in step S301).
[0077] When the constraint condition setting unit 110 determines that the user has input a constraint condition (Yes in step S301), the trial machining database 120 is created (step S302). At this stage, as shown in FIG. 5(A), none of the items have been set.
[0078] Next, steps S303 to S312 are loop processes, which are repeated for each trial machining ID until a predetermined condition is satisfied. When the loop process starts (step S303), the normalization processing unit 140 first normalizes all input variables and objective variables in past trial machining stored in the trial machining database 120 (step S304). For the first trial machining, there is no past data, so step S304 is skipped.
[0079] Next, the trial processing condition search unit 130 searches for a combination of input variables to be tried in the next trial processing, based on the past input variables and the objective variable normalized by the normalization processing unit 140, and a combination of input conditions that can be expected to minimize the objective variable (step S305).
[0080] When the trial machining condition searching unit 130 searches for a combination of each input variable, the normalization processing unit 140 inversely converts the searched combination of input variables (step S306). Then, the normalization processing unit 140 adds the inversely converted combination of input variables to the trial machining database 120. When a new combination of input variables is added to the trial machining database 120, the experiment support device 102 transmits a trial machining instruction specifying the added combination of input variables to the machining device 200 (step S307).
[0081] Thereafter, the processing device 200 receives a trial processing instruction from the experiment support device 102, performs trial processing under input conditions specified in the trial processing instruction, obtains the surface roughness which is the objective variable, and then transmits a response specifying the objective variable to the experiment support device 102. The experiment support device 102 determines whether or not a response specifying the objective variable has been obtained from the processing device 200 (step S308). The experiment support process is in a standby state until the experiment support device 102 determines that a response specifying the objective variable has been obtained from the processing device 200 (No in step S308).
[0082] On the other hand, when the experiment support device 102 determines that a response specifying an objective variable has been obtained from the processing device 200 (Yes in step S308), after the objective variable specified in the response is stored in the trial processing database 120, the extraction unit 180 extracts the combination of input variables when the objective variable takes the minimum value (step S309).
[0083] Next, the experiment support device 102 transmits a learning instruction specifying the contents of the current trial machining database 120 to the machine learning device 300. When the machine learning device 300 receives the learning instruction from the experiment support device 102, it estimates a combination of input variables when the surface roughness E, which is the objective variable, takes the minimum value from the trial machining database 120 specified in the learning instruction. Then, the machine learning device 300 transmits a response specifying the estimated combination of input variables to the experiment support device 102. By receiving the response specifying the estimated combination of input variables from the machine learning device, the experiment support device 102 acquires the combination of input variables when the objective variable takes the minimum value (step S310).
[0084] Thereafter, the experiment support device 102 determines whether the combination of input variables extracted by the extraction unit 180 matches the combination of input variables estimated by the machine learning device 300 (step S311). If it is determined that the combination of input variables extracted by the extraction unit 180 matches the combination of input variables estimated by the machine learning device 300 (Yes in step S311), the loop is exited and the termination processing unit 160 ends the experiment (step S313). Then, the experiment support process ends.
[0085] On the other hand, if it is determined that the combination of input variables extracted by the extraction unit 180 does not match the combination of input variables estimated by the machine learning device 300 (No in step S311), the process proceeds to step S312, and steps S303 to S312 are repeated for the next trial processing.
[0086] As described above, according to the third embodiment, the experiment is terminated when the optimal solutions of the combination of input variables calculated by the experiment support device 102 and the machine learning device 300 match. This makes it possible to prevent a situation in which a so-called local optimal solution is estimated as the optimal solution.
[0087] (Variation 1) In the first embodiment, it has been described that the contents of the trial machining database 120 obtained by the implementation support process may be transmitted to the machine learning device 300, and the optimal machining conditions may be estimated by the machine learning device 300. Here, the experiment support device 100 may delete data related to the trial machining having a data similarity of "0" before transmitting the contents of the trial machining database 120 to the machine learning device 300. By deleting the data before transmitting it to the machine learning device 300, it is possible to suppress the load on the network and the load on a central processing unit (CPU) during machine learning.
[0088] (Variation 2) In the above embodiment, it has been described that the spindle speed A, feed rate B, radial cut-in amount C, and axial cut-in amount D of the end mill 210 are selected as input variables, and the surface roughness E of the cut surface 221 is selected as the objective variable, but the variables selected as the input variables and objective variable are not limited to these. For example, the flow rate of the coolant used during cutting, the characteristic values of tools such as the end mill 210, the characteristic values of the workpiece 220, etc. may be added to the input variables, and the cutting resistance value, spindle current, etc. may be added to the objective variables.
[0089] In addition, multiple objective variables may be expressed in a mathematical formula to set a new objective variable. For example, objective variables 1 and 2 expressed in the following mathematical formulas may be set as objective variables using a cutting resistance value, which is a value that can be acquired by a sensor or the like provided in the processing device 200 and indicates the resistance that the end mill 210 receives from the workpiece 220 during cutting. Objective variable 1 = Cutting resistance value x geometric tolerance Objective variable 2 = Cutting resistance value / Cutting amount (Cutting amount = axial cutting amount D x diameter cutting amount C x feed rate B)
[0090] When estimating input variables as a multi-objective optimization problem, multi-objective optimization algorithms such as PSO (Particle Swarm Optimization) and genetic algorithms are used.
[0091] (Variation 3) In the above-described embodiments, the experiment support device 100, the experiment support device 101 or the experiment support device 102, the processing device 200, and the machine learning device 300 are treated as separate entities, but these may be configured as an integrated unit, or any two of them may be configured as an integrated unit.
[0092] (Variation 4) In the above embodiment, the present disclosure has been described by taking as an example a case where learning data of processing conditions of a processing device is obtained, but the present disclosure is not limited thereto. The present disclosure can be widely used to efficiently obtain learning data involving various experiments.
[0093] (Appendix 1) a trial machining database that stores a combination of input variables in an experiment for obtaining learning data for machine learning in association with a predetermined evaluation value; a trial machining condition searching unit that searches for a next combination of input variables based on the combination of input variables stored in the trial machining database; a similarity determination unit that calculates the evaluation value indicating the degree of similarity between the combination of input variables searched for by the trial machining condition search unit and each combination of input variables stored in the trial machining database, and adds the evaluation value to the trial machining database; a termination processing unit that terminates generation of the learning input data when the evaluation value determined by the similarity determination unit satisfies a predetermined condition; An experiment support device equipped with:
[0094] (Appendix 2) acquiring a target variable indicating a result of an experiment performed under the condition of the combination of input variables generated by the trial machining condition search unit, and storing the target variable in the trial machining database in association with the input variables used in the experiment; the trial machining condition search unit searches for a next combination of input variables based on the combination of input variables stored in the trial machining database and a corresponding objective variable; 2. The experimental support device according to claim 1.
[0095] (Appendix 3) When the similarity determination unit determines that the input variables in the current trial machining are similar to the input variables in any of the past trial machining operations consecutively for a predetermined number of times or more, the termination processing unit terminates the experiment. 3. An experiment support device according to claim 1 or 2.
[0096] (Appendix 4) When the similarity determination unit determines that the input variables in the current trial machining are similar to the input variables in any of the past trial machining a predetermined number of times or more, the termination processing unit terminates the experiment. 4. An experiment support device according to any one of appendices 1 to 3.
[0097] (Appendix 5) A constraint condition setting unit that sets a range of the input variable, The trial machining condition search unit searches for input variables in the trial machining to be attempted next within the set range. 5. An experiment support device according to any one of appendix 1 to 4.
[0098] (Appendix 6) The constraint condition setting unit further sets an upper limit on the number of trial machining operations included in the experiment, The termination processing unit terminates the experiment when the number of times of the trial machining exceeds the set upper limit number of times. 6. The experimental support device according to claim 5.
[0099] (Appendix 7) a normalization processing unit that normalizes the past input variables and the objective variables stored in the trial machining database to numerical values within a predetermined range, The trial machining condition search unit searches for input variables in the trial machining to be attempted next based on the normalized past input variables and the objective variable. 7. An experiment support device according to any one of appendix 1 to 6.
[0100] (Appendix 8) a database creation unit that creates the trial machining database in advance by an experimental design method; A constraint condition setting unit that sets an upper limit number of trial machining operations included in the experiment, The termination processing unit terminates the experiment when the number of times of the trial machining exceeds the set upper limit number of times. 8. An experiment support device according to any one of appendix 1 to 7.
[0101] (Appendix 9) an extraction unit that extracts, from the trial machining database, a combination of the input variables when the objective variable satisfies a predetermined condition; and a machine learning device that acquires the trial machining database and estimates a combination of the input variables when the objective variable satisfies a predetermined condition, the termination processing unit terminates the experiment when the extracted combination of input variables matches the estimated combination of input variables. 9. An experiment support device according to any one of appendices 1 to 8.
[0102] (Appendix 10) The computer A next combination of input variables is searched for based on the combination of input variables stored in a trial machining database that stores the combination of input variables in an experiment for obtaining learning data for machine learning in association with a predetermined evaluation value; determining the evaluation value indicating the degree of similarity between the searched combination of input variables and each of the combinations of input variables stored in the trial machining database, and adding the evaluation value to the trial machining database; if the obtained evaluation value satisfies a predetermined condition, the generation of the learning input data is terminated. Experimental support methods.
[0103] (Appendix 11) On the computer, a trial machining condition search unit that searches for a next combination of input variables based on a combination of input variables stored in a trial machining database that stores a combination of input variables in an experiment for obtaining learning data for machine learning in association with a predetermined evaluation value; a similarity determination unit that calculates the evaluation value indicating the similarity between the combination of input variables searched for by the trial machining condition search unit and each combination of input variables stored in the trial machining database, and adds the evaluation value to the trial machining database; a termination processing unit that terminates generation of the learning input data when the evaluation value determined by the similarity determination unit satisfies a predetermined condition; A program that functions as a
[0104] Various embodiments and modifications of the present disclosure are possible without departing from the broad spirit and scope of the present disclosure. The above-described embodiments are for explaining the present disclosure and do not limit the scope of the present disclosure. That is, the scope of the present disclosure is indicated by the claims, not the embodiments. Various modifications made within the scope of the claims and the scope of the disclosure equivalent thereto are considered to be within the scope of the present disclosure. [Explanation of symbols]
[0105] 1 Machining condition adjustment system, 10 input unit, 20 display unit, 30 control unit, 40 main memory unit, 50 auxiliary memory unit, 60 communication unit, 70 bus, 100, 101, 102 experiment support device, 110 constraint condition setting unit, 111 item type, 112 item, 113 constraint condition, 120, 121 trial machining database, 130 trial machining condition search unit, 140 normalization processing unit, 150 similarity judgment unit, 160 termination processing unit, 170 database creation unit, 180 extraction unit, 200 machining device, 210 end mill, 220 workpiece, 221 cutting surface, 300 machine learning device, A spindle speed, B feed rate, C diameter cutting amount, D axis cutting amount
Claims
1. a trial machining database that stores a combination of input variables in an experiment for obtaining learning data for machine learning in association with a predetermined evaluation value; a trial machining condition searching unit that searches for a next combination of input variables based on the combination of input variables stored in the trial machining database; a similarity determination unit that calculates the evaluation value indicating the similarity between the combination of input variables searched for by the trial machining condition search unit and each combination of input variables stored in the trial machining database, and adds the evaluation value to the trial machining database; a termination processing unit that terminates generation of the learning input data when the evaluation value determined by the similarity determination unit satisfies a predetermined condition; An experiment support device equipped with:
2. acquiring a target variable indicating a result of an experiment performed under the condition of the combination of input variables generated by the trial machining condition search unit, and storing the target variable in the trial machining database in association with the input variables used in the experiment; The trial machining condition search unit searches for a next combination of input variables based on the combination of input variables stored in the trial machining database and a corresponding objective variable. The experiment support device according to claim 1.
3. When the similarity determination unit determines that the input variables in the current trial machining are similar to the input variables in any of the past trial machining operations consecutively for a predetermined number of times or more, the termination processing unit terminates the experiment. The experiment support device according to claim 1.
4. When the similarity determination unit determines that the input variables in the current trial machining are similar to the input variables in any of the past trial machining a predetermined number of times or more, the termination processing unit terminates the experiment. The experiment support device according to claim 1.
5. A constraint condition setting unit that sets a range of the input variable, The trial machining condition search unit searches for input variables in the trial machining to be attempted next within the set range. The experiment support device according to claim 1.
6. The constraint condition setting unit further sets an upper limit on the number of trial machining operations included in the experiment, The termination processing unit terminates the experiment when the number of times of the trial machining exceeds the set upper limit number of times. The experiment support device according to claim 5.
7. a normalization processing unit that normalizes the past input variables and the objective variables stored in the trial machining database to numerical values within a predetermined range, The trial machining condition search unit searches for input variables in the trial machining to be attempted next based on the normalized past input variables and the objective variable. The experiment support device according to claim 1.
8. a database creation unit that creates the trial machining database in advance by an experimental design method; A constraint condition setting unit that sets an upper limit number of trial machining operations included in the experiment, The termination processing unit terminates the experiment when the number of times of the trial machining exceeds the set upper limit number of times. The experiment support device according to claim 1.
9. an extraction unit that extracts, from the trial machining database, a combination of the input variables when the objective variable satisfies a predetermined condition; and a machine learning device that acquires the trial machining database and estimates a combination of the input variables when the objective variable satisfies a predetermined condition, the termination processing unit terminates the experiment when the extracted combination of input variables matches the estimated combination of input variables. The experiment support device according to claim 1.
10. The computer A next combination of input variables is searched for based on the combination of input variables stored in a trial machining database that stores the combination of input variables in an experiment for obtaining learning data for machine learning in association with a predetermined evaluation value; determining the evaluation value indicating the degree of similarity between the searched combination of input variables and each of the combinations of input variables stored in the trial machining database, and adding the evaluation value to the trial machining database; if the obtained evaluation value satisfies a predetermined condition, the generation of the learning input data is terminated. Experimental support methods.
11. On the computer, a trial machining condition search unit that searches for a next combination of input variables based on a combination of input variables stored in a trial machining database that stores a combination of input variables in an experiment for obtaining learning data for machine learning in association with a predetermined evaluation value; a similarity determination unit that calculates the evaluation value indicating the similarity between the combination of input variables searched for by the trial machining condition search unit and each combination of input variables stored in the trial machining database, and adds the evaluation value to the trial machining database; a termination processing unit that terminates generation of the learning input data when the evaluation value determined by the similarity determination unit satisfies a predetermined condition; A program that functions as a
Citation Information
Patent Citations
Experiment assisting device and experiment assisting method
JP2019185591A
