Method for generating learned judgment criteria, method for generating press working result judger, method for estimating press working result, device for generating learned judgment criteria, and press working result judger

Abstract

To provide a learning-finish determination reference creation method which can determine a process state corresponding to a minute change.SOLUTION: A learning-finish determination reference creation method is a determination reference creation method for making a computer function so as to output a determination reference indicating a press-processing result on the basis of sensing data of a press-processing device includes: a learning data group acquisition step statistically and significantly including a normal sensing data group when performing press-processing by using a tool which is in a normal state, and also including an abnormal sensing data group not belonging to the normal sensing data group; and a determination reference creation step of performing no-teacher mechanical learning by using a learning data group, and creating and updating a plurality of determination references for classifying the learning data group, and creating and updating a first determination reference for classifying the learning data group to the normal sensing data group and a second determination reference for classifying the learning data group to the abnormal sensing data group.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a learned judgment criterion generating method, a press working result judger generating method, a press working result estimating method, a learned judgment criterion generating device, and a press working result judger. [Background technology]

[0002] 2. Description of the Related Art In press working devices, it has been considered to determine whether or not an abnormality has occurred and the cause of the abnormality by detecting the working status.

[0003] For example, the detection device described in Patent Document 1 outputs a signal representing the machining status based on a derived profile showing the relationship between a force value representing the force applied to the machining member from the workpiece and a distance value representing the distance between the workpiece and the machining member. The machining status indicates, for example, machining defects due to various causes. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2017-164751 Summary of the Invention [Problem to be solved by the invention]

[0005] The detection device described in Patent Document 1 still has room for improvement in terms of determining the process state in response to minute changes.

[0006] Therefore, the present disclosure provides a learned judgment criterion generation method, a press working result judger generation method, a press working result estimation method, a learned judgment criterion generation device, and a press working result judger that can judge process states that correspond to minute changes. [Means for solving the problem]

[0007] In one aspect of the present disclosure, a learned judgment criterion generation method is a judgment criterion generation method for causing a computer to function to output a first judgment criterion and a second judgment criterion that indicate the press processing results of a press processing device, based on sensing data that indicates the relationship between the position of the punch and the load on the workpiece of a press processing device that has a die and a punch facing the die and processes a workpiece placed on the die using the load of the punch.The judgment criterion generation method includes a training data group acquisition step of acquiring a training data group that statistically significantly includes a normal sensing data group that indicates the relationship between the position of the punch and the load on the workpiece when press processing is performed using a die, punch, and workpiece in a normal state, and includes an abnormal sensing data group that does not belong to the normal sensing data group, and a judgment criterion generation step of performing unsupervised machine learning using the training data group to generate and update multiple judgment criteria for classifying the training data group, wherein the judgment criterion generation step generates and updates a first judgment criterion that has been machine-learned to classify the training data group into a normal sensing data group and a second judgment criterion that has been machine-learned to classify the training data group into an abnormal sensing data group.

[0008] A method for generating a press working result judger according to one embodiment of the present disclosure is a method for generating a press working result judger based on a first judgment criterion and a second judgment criterion generated by the above-mentioned generation method, and includes the steps of acquiring a plurality of judgment criteria including the first judgment criterion and the second judgment criterion, and setting a label indicating normality or the cause of an abnormality for each of the acquired plurality of judgment criteria.

[0009] A press working result estimation method according to one aspect of the present disclosure is a method for estimating the results of press working based on sensing data obtained by a press working device during press working using a press working result judger generated by the above-mentioned press working result judger generation method, and includes the steps of acquiring sensing data from the press working device, estimating which of the judgment criteria the sensing data is similar to, and outputting a label attached to the judgment criterion that is similar to the sensing data.

[0010] A learned judgment criterion generation device according to one embodiment of the present disclosure is a device that generates multiple judgment criteria, including a first judgment criterion and a second judgment criterion, regarding the results of press processing by a press processing device that has a die and a punch facing the die and processes a workpiece placed on the die using the load of the punch, based on sensing data indicating the relationship between the position of the punch and the load on the workpiece.The device is equipped with: a memory unit that acquires a training data group that statistically significantly includes normal sensing data that indicates the relationship between the position of the punch and the load on the workpiece when press processing is performed using a die, punch, and workpiece in a normal state, and includes abnormal sensing data that does not belong to the normal sensing data group; and a judgment criterion generation unit that performs unsupervised machine learning using the training data group to generate multiple judgment criteria for classifying the training data group, and generates the first judgment criterion that has been trained to classify the training data group into a normal sensing data group, and a second judgment criterion that has been machine learned to classify the training data group into an abnormal data group.

[0011] A press working result determiner according to one embodiment of the present disclosure includes a plurality of judgment criteria including a first judgment criterion and a second judgment criterion generated by the learned judgment criterion generation device described above, and a label attached to each of the plurality of judgment criteria indicating whether the press working result of the press working device is normal or the cause of an abnormality, and determines which of the plurality of judgment criteria the input sensing data is similar to and outputs the label. [Effects of the Invention]

[0012] According to the present disclosure, it is possible to provide a learned judgment criterion generation method, a press working result judger generation method, a press working result estimation method, a learned judgment criterion generation device, and a press working result judger that can judge process states corresponding to minute changes. [Brief explanation of the drawings]

[0013] [Figure 1A] FIG. 1 is a block diagram illustrating a learned criterion generation device according to a first embodiment; [Figure 1B] FIG. 1 is a block diagram showing a press working result determiner according to a first embodiment; [Figure 1C] Block diagram showing an example of a press processing device [Figure 2] A flowchart showing the generation of a criterion by the learned criterion generation device of FIG. 1A. [Figure 3] FIG. 3 is a diagram illustrating the criteria generated by the flowchart of FIG. 2; [Figure 4] A schematic diagram explaining how multiple criteria are generated from a set of training data. [Figure 5] A block diagram showing a press working result judger including the judgment criteria generated in the example of Figure 4. [Figure 6] Schematic diagram showing the waveforms of sensing data corresponding to each judgment criterion [Figure 7] Schematic diagram showing the waveforms of sensing data corresponding to each judgment criterion [Figure 8] Schematic diagram showing the waveforms of sensing data corresponding to each judgment criterion [Figure 9] Schematic diagram showing the waveforms of sensing data corresponding to each judgment criterion [Figure 10] Schematic diagram showing the waveforms of sensing data corresponding to each judgment criterion [Figure 11] Schematic diagram showing the waveforms of sensing data corresponding to each judgment criterion [Figure 12] Flowchart showing estimation of press working results using a press working result determiner DETAILED DESCRIPTION OF THE INVENTION

[0014] (Background to this disclosure) In press processing equipment, a workpiece is generally formed into a predetermined shape using metal molds called a die and a punch. Processing using press processing equipment is widespread in a wide range of manufacturing fields, such as home appliances, precision instruments, and automotive parts.

[0015] In press processing equipment, the number of shots is generally managed to manage the lifespan of the die due to wear or breakage. A shot in a press processing equipment refers to a processing operation performed with one stroke of the die of the press processing equipment. For example, when performing a punching process with a press processing equipment, the number of times the workpiece has been processed, i.e., the number of shots, is used to predict when the die should be replaced.

[0016] However, if the molds in the press processing equipment are not properly maintained, the die and punch may rub against each other, or two workpieces may be punched out unexpectedly, placing a greater load on the mold than expected. In this case, the mold may reach the end of its lifespan with a faster number of shots than expected, or it may suddenly break. This results in losses due to poor processing, which is an issue.

[0017] Therefore, a method of detecting an abnormality by detecting the machining status, such as the detection device disclosed in Patent Document 1, has been studied.

[0018] However, the detection device described in Patent Document 1 has the problem that it is difficult to set a threshold value for determining whether the press processing is normal or abnormal, and productivity is reduced due to the work of acquiring and analyzing data that associates abnormal press processing with the derived profile.

[0019] Therefore, the present inventors have studied a press working determination device, a press working determination program, and a press working device that can determine the process state, and have arrived at the following invention.

[0020] In one aspect of the present disclosure, a learned judgment criterion generation method is a judgment criterion generation method for causing a computer to function to output a first judgment criterion and a second judgment criterion that indicate the press processing results of a press processing device, based on sensing data that indicates the relationship between the position of the punch and the load on the workpiece of a press processing device that has a die and a punch facing the die and processes a workpiece placed on the die using the load of the punch.The judgment criterion generation method includes a training data group acquisition step of acquiring a training data group that statistically significantly includes a normal sensing data group that indicates the relationship between the position of the punch and the load on the workpiece when press processing is performed using a die, punch, and workpiece in a normal state, and includes an abnormal sensing data group that does not belong to the normal sensing data group, and a judgment criterion generation step of performing unsupervised machine learning using the training data group to generate and update multiple judgment criteria for classifying the training data group, wherein the judgment criterion generation step generates and updates a first judgment criterion that has been machine-learned to classify the training data group into a normal sensing data group and a second judgment criterion that has been machine-learned to classify the training data group into an abnormal sensing data group.

[0021] This configuration allows for the generation of learned criteria that can determine process conditions corresponding to minute changes, and also allows for the automatic generation of multiple criteria using a set of training data.

[0022] The first and second judgment criteria may be trained to distinguish between similarities between sensing data belonging to a normal sensing data group in the training data group and between similarities between sensing data belonging to an abnormal sensing data group, respectively.

[0023] With this configuration, it is possible to generate a judgment criterion that corresponds to minute process changes in press working.

[0024] The judgment criterion generation step may be a step of generating a first judgment criterion and a second judgment criterion using a Gaussian mixture model, and may include: (a) preparing a probability density function of a first Gaussian distribution for determining the first judgment criterion and a probability density function of a second Gaussian distribution for determining the second judgment criterion; (b) calculating a first weighting coefficient for the first Gaussian distribution and a second weighting coefficient for the second Gaussian distribution using the probability density for the first Gaussian distribution and the probability density for the second Gaussian distribution calculated for each of the sensing data included in the training data group; (c) updating the first judgment criterion and the second judgment criterion by updating the probability density function of the first Gaussian distribution and the probability density function of the second Gaussian distribution using the first weighting coefficient and the second weighting coefficient; (d) repeating step (b) and step (c) until a predetermined condition is satisfied; and (e) adopting the probability density function of the first Gaussian distribution and the probability density function of the second Gaussian distribution when the predetermined condition is satisfied as the first judgment criterion and the second judgment criterion.

[0025] With this configuration, it is possible to automatically generate a plurality of judgment criteria corresponding to a plurality of abnormality causes.

[0026] The step (b) includes calculating, for each piece of sensing data included in the training data group, a likelihood of the sensing data with respect to a first Gaussian distribution, calculating a likelihood of the sensing data with respect to a second Gaussian distribution, and calculating a first weighting coefficient and a second weighting coefficient using a ratio of the calculated likelihoods; may include:

[0027] With this configuration, it is possible to generate a criterion that can more accurately determine the press working results.

[0028] The step (b) may include calculating, using a Gaussian mixture model, a first weighting coefficient for the first judgment criterion for each piece of sensing data and a second weighting coefficient for the second judgment criterion for each piece of sensing data based on the sum of the probability densities for the first Gaussian distribution and the second Gaussian distribution before updating, and the step (c) may include updating the first judgment criterion by dividing the mean and variance of the product of the sensing data and the first weighting coefficient by the sum of the first weighting coefficients, and updating the second judgment criterion by dividing the mean and variance of the product of the sensing data and the second weighting coefficient by the sum of the second weighting coefficients.

[0029] With this configuration, it is possible to generate a criterion that can more accurately determine the press working results.

[0030] The criterion generating step may generate the first criterion and the second criterion using the K-Means algorithm.

[0031] With this configuration, it is possible to generate a criterion that can more accurately determine the press working results.

[0032] The criterion generating step may be repeatedly executed until the similarity between the first criterion and the second criterion in the updated first criterion and the updated second criterion becomes equal to or less than a predetermined threshold.

[0033] This configuration allows multiple criteria to be generated.

[0034] A method for generating a press working result judger according to one embodiment of the present disclosure is a method for generating a press working result judger based on a first judgment criterion and a second judgment criterion generated by the above-mentioned generation method, and includes the steps of acquiring a plurality of judgment criteria including the first judgment criterion and the second judgment criterion, and setting a label indicating normality or the cause of an abnormality for each of the acquired plurality of judgment criteria.

[0035] With this configuration, it is possible to generate a press working result judger that can determine whether the press working is normal or, if abnormal, identify the cause of the abnormality depending on which judgment criterion the sensing data falls under.

[0036] A press working result estimation method according to one aspect of the present disclosure is a method for estimating the results of press working based on sensing data obtained by a press working device during press working using a press working result judger generated by the above-mentioned press working result judger generation method, and includes the steps of acquiring sensing data from the press working device, estimating which of the judgment criteria the sensing data is similar to, and outputting a label attached to the judgment criterion that is similar to the sensing data.

[0037] With this configuration, it is possible to identify the cause of an abnormality corresponding to a minute change in the process based on the sensing data during press working, thereby improving the productivity of press working.

[0038] A learned judgment criterion generation device according to one embodiment of the present disclosure is a device that generates multiple judgment criteria, including a first judgment criterion and a second judgment criterion, regarding the results of press processing by a press processing device that has a die and a punch facing the die and processes a workpiece placed on the die using the load of the punch, based on sensing data indicating the relationship between the position of the punch and the load on the workpiece.The device is equipped with: a memory unit that acquires a training data group that statistically significantly includes normal sensing data that indicates the relationship between the position of the punch and the load on the workpiece when press processing is performed using a die, punch, and workpiece in a normal state, and includes abnormal sensing data that does not belong to the normal sensing data group; and a judgment criterion generation unit that performs unsupervised machine learning using the training data group to generate multiple judgment criteria for classifying the training data group, and generates the first judgment criterion that has been trained to classify the training data group into a normal sensing data group, and a second judgment criterion that has been machine learned to classify the training data group into an abnormal data group.

[0039] This configuration provides a method for generating learned criteria that can determine process conditions corresponding to minute changes, and also enables automatic generation of multiple criteria using a set of training data.

[0040] A press working result determiner according to one embodiment of the present disclosure includes a plurality of judgment criteria including a first judgment criterion and a second judgment criterion generated by the learned judgment criterion generation device described above, and a label attached to each of the plurality of judgment criteria indicating whether the press working result of the press working device is normal or the cause of an abnormality, and determines which of the plurality of judgment criteria the input sensing data is similar to and outputs the label.

[0041] With this configuration, it is possible to identify the cause of an abnormality corresponding to a minute change in the process based on the sensing data during press working, thereby improving the productivity of press working.

[0042] Hereinafter, embodiments of the present disclosure will be described in detail with appropriate reference to the drawings. However, more detailed description than necessary may be omitted. For example, detailed description of already well-known matters or redundant description of substantially identical configurations may be omitted. This is to avoid unnecessary redundancy in the following description and to facilitate understanding by those skilled in the art. Note that the inventors provide the accompanying drawings and the following description to enable those skilled in the art to fully understand the present disclosure, and do not intend for them to limit the subject matter described in the claims.

[0043] (Embodiment 1) [Overall configuration] Fig. 1A is a block diagram showing a learned judgment criterion generating device 100 according to the first embodiment. Fig. 1B is a block diagram showing a press working result determiner 200 according to the first embodiment. Fig. 1C is a block diagram showing a press working device 300.

[0044] A learned judgment criterion generating device 100 and a press working result judger 200 according to this embodiment will be described with reference to FIGS. 1A to 1C.

[0045] The learned judgment criterion generation device 100 shown in Fig. 1A is a device that generates a plurality of judgment criteria related to the results of press working of the press working device 300 shown in Fig. 1C based on sensing data acquired during press working by the press working device 300. The learned judgment criterion generation device 100 can be constructed using a computer system such as a PC or a workstation. The learned judgment criterion generation device 100 includes a memory unit 11 and a judgment criterion generation unit 12.

[0046] The memory unit 11 accumulates sensing data acquired from the press working device 300 as a learning data group 13. The sensing data is acquired by receiving it from the press working device 300 using, for example, a communication unit (not shown). The sensing data is data acquired by a sensor (group) such as a load sensor provided on the die 31 or the punch 32, or on both, when press working is performed using the die 31, punch 32, and workpiece 33 in a normal state or a state in which an abnormality exists. The learning data group 13 includes a normal sensing data group and an abnormal sensing data group that does not belong to the normal sensing data group.

[0047] The storage unit 11 can store the training data group 13, and is configured by, for example, a hard disk (HDD), an SSD, a memory, and the like.

[0048] The judgment criterion generation unit 12 generates and updates a plurality of judgment criteria including a first judgment criterion c1 and a second judgment criterion c2. Specifically, unsupervised machine learning is performed using the training data group 13 stored in the storage unit 11. As a result, the judgment criterion generation unit 12 generates a plurality of judgment criteria including a first judgment criterion c1 that has been trained to classify the training data group 13 into a normal sensing data group, and a second judgment criterion c2 that has been trained to classify the training data group 13 into an abnormal sensing data group.

[0049] In the learned criterion generation device 100, the sensing data stored in the storage unit 11 is used as a learning data group 13, and a learning phase, which will be described later, is executed.

[0050] The criterion generating unit 12 can be configured with, for example, a microcomputer, a CPU, an MPU, a GPU, a DSP, an FPGA, or an ASIC. The functions of the criterion generating unit 12 may be configured with hardware alone, or may be realized by combining hardware and software.

[0051] 1B is a determiner that acquires new sensing data from the pressing device 300, determines whether the data corresponds to one of a plurality of learned judgment criteria, and outputs a label corresponding to the corresponding judgment criterion. The pressing result determiner 200 includes a plurality of judgment criteria including a first judgment criterion c1 and a second judgment criterion c2 generated by the learned judgment criterion generation device 100, and a plurality of labels L1, L2 corresponding to the plurality of judgment criteria, respectively. The sensing data is acquired by receiving it from the pressing device 300 using, for example, a communication unit (not shown).

[0052] The press working result determiner 200 executes an inference phase in which it predicts whether the press working device 300 is normal and, if an abnormality exists, what type of abnormality it is, by using the sensing data from the press working device 300 and a plurality of learned judgment criteria. The learned judgment criterion generation device 100 and the press working result determiner 200 may be realized by the same hardware.

[0053] The press working result determiner 200 can be configured with, for example, a microcomputer, a CPU, an MPU, a GPU, a DSP, an FPGA, or an ASIC. The functions of the press working result determiner 200 may be configured with hardware only, or may be realized by combining hardware and software.

[0054] 1C has a die 31 and a punch 32 facing the die 31, and is an apparatus that processes a workpiece 33 placed on the die 31 by the load of the punch 32. In this embodiment, a load sensor 35 and a position sensor 36 are used as sensors 34 arranged in the press working apparatus 300. The sensing data of the press working apparatus 300 is process data during press working detected by the sensor 34. Specifically, the sensing data is waveform data that indicates the relationship between the position of the punch 32 and the load applied to the workpiece 33 when press working is performed using the die 31, punch 32, and workpiece 33.

[0055] The load sensor 35 preferably has high sensitivity capable of detecting minute changes in load so that minute process conditions are reflected in the criteria c1 and c2. For this reason, a quartz piezoelectric sensor is suitable as the load sensor 35. The load sensor 35 is preferably disposed so that it supports the die 31 or punch 32.

[0056] Similarly, the position sensor 36 preferably has high resolution to accurately measure the position of the punch 32 so that minute process conditions are reflected relative to the criteria c1 and c2. For this reason, an eddy current sensor or a capacitance sensor is suitable for the position sensor 36. In particular, when the thickness of the workpiece 33 is on the order of several tens of μm, it is preferable that the resolution of the position sensor 36 be 1 μm or less. Furthermore, it is preferable that the position sensor 36 be a sensor with low measurement noise.

[0057] The learned judgment criterion generation device 100, the press working result judger 200, and the press working device 300 are connected to each other so that they can communicate with each other via wired or wireless communication. They may be installed in the same factory or in two or more different locations on the factory premises. Communication may be performed using a public line such as the Internet and / or a dedicated line.

[0058] [Judgment criteria generation method] Fig. 2 is a flowchart showing generation of a criterion by the learned criterion generation device 100 of Fig. 1A. Fig. 3 is a diagram schematically showing a criterion generated by the flowchart of Fig. 2. A method for generating a criterion by the learned criterion generation device 100 will be described with reference to Figs. 2 and 3.

[0059] The judgment criteria can be generated by performing unsupervised machine learning using an algorithm that classifies the training data group 13 into two or more classes. Specific examples of the algorithm that can be used include the K-Means method and a Gaussian mixture model. In this embodiment, an example using a Gaussian mixture model will be described.

[0060] First, the memory unit 11 stores a learning data group 13 input from an external device (learning data group acquisition step, step S11). The learning data group 13 is a sensing data group indicating the relationship between the position of the punch 32 and the load applied to the workpiece 33 during press processing. The sensing data constituting the learning data group 13 is waveform data indicating the relationship between the position of the punch 32 during one press processing shot and the corresponding magnitude of the load on the workpiece 33, measured by the load sensor 35 and position sensor 36 of the press processing device 300. Examples of waveform data are shown in FIGS. 6 to 11, which will be described later. In this embodiment, the waveform data is acquired by acquiring the magnitude of the load on the workpiece 33 for each of positions 1 to L (L is a natural number) of the punch 32 in the sensing data. That is, data on the position of the punch 32 and the load on the workpiece 33 is acquired L times during one press processing shot, and this data is used as one piece of sensing data. In one example, L is 100.

[0061] The training data group 13 includes a normal sensing data group and an abnormal sensing data group that does not belong to the normal sensing data group. Here, the normal sensing data group is sensing data obtained when press processing is performed using a die 31, punch 32, and workpiece 33 in a normal state. The abnormal sensing data group is sensing data that does not belong to the normal sensing data group. The normal state may be defined, for example, by processing with a small total number of shots in the press processing device 300. The significance level is generally 5%. In a preferred example of this embodiment, the normal sensing data group in the training data group 13 should be 70% or more, in other words, the abnormal sensing data group should be within 30% of the total. In this case, it is possible to prevent the abnormal sensing data group from affecting the generation of the first judgment criterion c1 for classifying the normal sensing data group.

[0062] In order to statistically significantly include normal sensing data groups, the learning data group 13 should be a sensing data group for N shots (N is a natural number) from the start of processing by the stamping device 300. This is because it is assumed that the die 31, punch 32, and workpiece 33 are in a normal state at the start of processing. In addition, in order to obtain a sufficient amount of learning data group 13 for generating a judgment criterion, N is preferably 10,000 or more.

[0063] Next, the judgment criterion generation unit 12 generates a judgment criterion (judgment criterion generation step, step S12). The generation of the judgment criterion is a step of performing unsupervised machine learning using the training data group 13 and generating and updating a plurality of judgment criteria for classifying the training data group 13. The first judgment criterion c1 is a criterion for determining whether the training data group 13 is normal data, and the second judgment criterion c2 is a criterion for determining whether the training data group 13 is abnormal data.

[0064] In this embodiment, the first judgment criterion c1 and the second judgment criterion c2 are trained to identify the similarity between sensing data belonging to a normal sensing data group in the training data group 13 and the similarity between sensing data belonging to an abnormal sensing data group, respectively.

[0065] To determine the first and second criterion c1 and c2, the criterion generation unit 12 first executes a step of preparing a probability density function of a Gaussian distribution as an initial setting of the criterion (step S13). Specifically, the criterion generation unit 12 prepares a first Gaussian distribution g1 and a second Gaussian distribution g2 using a mean and a variance that define the probability density function of the Gaussian distribution, which are designated in advance by, for example, a user. The first Gaussian distribution g1 and the second Gaussian distribution g2 become the initial setting of the criterion (see FIG. 3).

[0066] The initial setting of the judgment criterion is basically a Gaussian distribution calculated from two or more arbitrary sensing data that make up the training data group. However, since the above-mentioned first judgment criterion c1 is a judgment criterion that indicates that the process is normal, learning is performed using a training data group that statistically significantly includes normal sensing data groups.

[0067] The probability density function f1 for the first Gaussian distribution g1 is given by equation (1): μ1 represents the mean of the first Gaussian distribution g1, and σ1 represents the variance of the first Gaussian distribution g1.

[0068]

number

[0069] Similarly, the probability density function f2 for the second Gaussian distribution g2 is given by equation (2): μ2 denotes the mean of the second Gaussian distribution g2, and σ2 denotes the variance of the second Gaussian distribution g2.

[0070]

number

[0071] Using the above equations (1) and (2), the probability density function of the Gaussian distribution for criterion 1 and the probability density function of the Gaussian distribution for criterion 2 at point j in waveform i of the training data group 13 are defined as f1 and f2, respectively. i indicates the ith (i is 1 to N) sensing data in the training data group 13, and j indicates the position j (j is 1 to L) of the ith sensing data. Hereinafter, the ith sensing data may also be referred to as sensing data i.

[0072] The probability density for the first Gaussian distribution g1 and the probability density for the second Gaussian distribution g2 are used to calculate the first weighting coefficient α for the first Gaussian distribution g1. i1 , and the second weighting factor α for the second Gaussian distribution g2 i2 (Step S14). i1 and the second weighting factor α i2 is calculated by calculating the likelihood of the sensing data for the first Gaussian distribution g1 and the likelihood of the sensing data for the second Gaussian distribution g2, and using the ratio of these likelihoods. Specifically, the first weighting coefficient α i1 and the second weighting factor α i2 is the sum of the probability densities for the first Gaussian distribution g1, P i1 and the sum of the probability densities for the second Gaussian distribution g2, P i2 It is calculated based on the following.

[0073] The sum of the probability densities for the first Gaussian distribution g1 is P i1 is calculated using equation (3).

[0074]

number

[0075] The sum of the probability densities P for the second Gaussian distribution g2 i2 is calculated using equation (4).

[0076]

number

[0077] First weighting coefficient α for the first Gaussian distribution g1 i1 is the sum of the probability distributions P for the first Gaussian distribution g1 and the second Gaussian distribution g2. i1 , P i2 Based on this, it is calculated using equation (5).

[0078]

number

[0079] Similarly, the second weighting coefficient α for the second Gaussian distribution g2 i2 is the sum of the probability distributions P for the first Gaussian distribution g1 and the second Gaussian distribution g2. i1 , P i2 Based on this, it is calculated using equation (6).

[0080]

number

[0081] For all sensing data i, a weight coefficient α i1 , α i2 Calculate.

[0082] Next, the first weighting coefficient α i1 and the second weighting factor α i2 The first Gaussian distribution g1 and the second Gaussian distribution g2 are updated by updating the probability density function f1 of the first Gaussian distribution g1 and the probability density function f2 of the second Gaussian distribution g2 using the above equation. As shown in Fig. 3, the first Gaussian distribution g1 is updated to obtain the first update criterion u1, and the second Gaussian distribution g2 is updated to obtain the second update criterion u2.

[0083] The first Gaussian distribution g1 and the second Gaussian distribution g2 are updated using the weight coefficient α i1 , α i2 This is done by updating the mean and variance of the first Gaussian distribution g1 and the second Gaussian distribution g2 using

[0084] The mean and variance are calculated for each position j of the sensing data. i1 Average of the product of FM1 j is calculated by equation (7), where F ij denotes the load at position j of sensing data i.

[0085]

number

[0086] Similarly, the sensing data i and the second weighting coefficient α i2 Average FM2 of the product j is calculated using equation (8).

[0087]

number

[0088] Sensing data i and the first weighting coefficient α i1 Variance of the product FV1 j is calculated using equation (9).

[0089]

number

[0090] Similarly, the sensing data i and the second weighting coefficient α i2 Variance of the product FV2 j is calculated using equation (10).

[0091]

number

[0092] In equations (9) and (10), A1 is the first weighting coefficient α of the sensing data i. i1 A2 is the average of the second weighting coefficient α i2Also, μ j is the average of the weights at each position j of the sensing data i.

[0093] Average FM1 calculated using equations (7) to (10) j , FM2 j and variance FV1 j , FV2 j is divided by the sum of the weighting coefficients of all the sensing data i, thereby generating the updated first update reference u1 and second update reference u2.

[0094] The calculation of the weighting coefficient in step S14 and the update of the determination criteria in step S15 are repeatedly executed until a predetermined condition is satisfied.

[0095] In this embodiment, the mean FM1 of the first Gaussian distribution g1 j and variance FV1 j , and the mean FM2 of the second Gaussian distribution g2 j and variance FV2 j The condition is satisfied when the amount of change is equal to or less than a predetermined threshold. The threshold can be set, for example, to 1% or less.

[0096] That is, after step S15, if the changes in the mean and variance of the first Gaussian distribution g1 and the mean and variance of the second Gaussian distribution g2 each exceed 1%, the process returns to step S14 (No in step S16).If the changes in the mean and variance of the first Gaussian distribution g1 and the mean and variance of the second Gaussian distribution g2 each are 1% or less (Yes in step S16), the updated first update criterion u1 and second update criterion u2 are adopted as the first judgment criterion c1 and the second judgment criterion c2, respectively (step S17).

[0097] As described above, the judgment criterion generation unit 12 generates two judgment criteria from the input learning data group 13. The judgment criterion generation unit 12 calculates the similarity between the two generated judgment criteria (first judgment criterion c1 and second judgment criterion c2) and repeats generating judgment criteria until the similarity exceeds a predetermined threshold. In this way, by repeating the generation of judgment criteria, multiple judgment criteria can be generated.

[0098] 4 is a schematic diagram illustrating the generation of multiple judgment criteria from the learning data group 13. The judgment criterion generation unit 12 receives the learning data group 13 as input and generates a first judgment criterion c1 and a second judgment criterion c2. In FIG. 4, the normal judgment criterion 1 corresponds to the above-mentioned first judgment criterion c1, and the abnormal judgment criterion 1 corresponds to the above-mentioned second judgment criterion c2.

[0099] The judgment criterion generation unit 12 calculates the similarity between normality judgment criterion 1 and abnormality judgment criterion 1. For example, Euclidean distance or the sum of the Gaussian probability distributions can be used as the similarity. If the similarity between normality judgment criterion 1 and abnormality judgment criterion 1 is equal to or less than a predetermined threshold, the corresponding data 13a classified as normality judgment criterion 1 is input, and the processing of steps S13 to S17 in FIG. 2 is executed to further classify normality judgment criterion 1 into normality judgment criterion 2 and abnormality judgment criterion 2. Similarly, abnormality judgment criterion 1 is classified into abnormality judgment criterion 5 and abnormality judgment criterion 6.

[0100] If the similarity between the two generated judgment criteria is higher than a predetermined threshold, the judgment criterion generator 12 determines that the generation of the judgment criteria has converged. As a result of repeating the generation of two judgment criteria in this way, seven judgment criteria are generated: normal judgment criterion 2, abnormality judgment criterion 3, abnormality judgment criterion 4, abnormality judgment criterion 5, abnormality judgment criterion 8, abnormality judgment criterion 9, and abnormality judgment criterion 10 in FIG. 4.

[0101] The judgment criteria generated by the judgment criterion generating unit 12 are used in the press working result judger 200. FIG. 5 is a block diagram showing the press working result judger 200 including the judgment criteria generated in the example of FIG. 4. As shown in FIG. 5, a label is attached to each judgment criterion. Since each judgment criterion shown in FIG. 5 is a criterion for classifying input data, it does not indicate whether the press working result is normal or, if abnormal, the cause of the abnormality. Therefore, in the press working result judger 200, a label indicating whether each judgment criterion is normal or the cause of the abnormality is attached to each judgment criterion.

[0102] Label Settings 6 to 11 are schematic diagrams showing waveforms of sensing data corresponding to each judgment criterion. The label setting conditions for each abnormality judgment criterion will be described with reference to Fig. 6 to Fig. 11. In Fig. 6 to Fig. 11, the dashed lines represent waveforms when the press working result is judged to be normal, and the solid lines represent waveforms when the press working result is judged to be abnormal.

[0103] Figure 6 shows waveform data when tool (punch) wear occurs. When tool wear occurs, stress concentration at the tip of the punch decreases. As a result, the amount of punch travel required to punch the workpiece increases.

[0104] The position where the load on the workpiece begins to be applied is the punching start position S1, and the position where the load on the workpiece after punching starts is the same as before punching is the punching end position E1. In this case, the workpiece thickness T1 is the size from S1 to E1. If tool wear has occurred, the punching end position E2 will be lower than the normal judgment standard punching end position E1 by 5% or more of the workpiece thickness T1. In other words, if the distance D1 from E1 to E2 is 5% or more of T1, a label indicating tool wear is set.

[0105] Figure 7 shows waveform data when tool chipping occurs. When tool chipping occurs, stress concentration at the chipping location decreases. As a result, the load required to locally break the workpiece increases. When tool chipping occurs, the load required for punching increases. For this reason, in the waveform data, an inflection point FP1 exists between the peak load occurrence position P1 and the punching end position E3, where the load increases compared to the normal judgment standard. If there is one or more inflection points between the peak load occurrence position P1 and the punching end position E3, a tool chipping label is set.

[0106] Figure 8 shows waveform data when double punching or foreign matter contamination occurs. If a punched workpiece or other foreign matter is present at the top or bottom of the workpiece, the position S2 where the load on the workpiece begins to be applied moves above the punching start position S1, which is the normal judgment standard. If the distance D2 between the position S2 where the load on the workpiece begins to be applied and the punching start position S1, which is the normal judgment standard, is 5% or more of the workpiece thickness T1, a foreign matter contamination label is set. In particular, if the distance D2 between the position S2 where the load on the workpiece begins to be applied and the punching start position S1, which is the normal judgment standard, is 90% or more of the workpiece thickness T1, a double punching label is set.

[0107] Figure 9 shows waveform data when scrap clogging has occurred. When workpiece scraps become stuck inside the die, even after punching of the workpiece is complete, the punch comes into contact with the stuck scraps, generating a load. For this reason, in the waveform data, the position E4 where the load no longer applies is located below the punching end position E1, which is the normal judgment standard. Therefore, if the distance D3 between the position E4 where the load no longer applies and the punching end position E1, which is the normal judgment standard, is 5% or more of the workpiece thickness T1, a scrap clogging label is set.

[0108] Figures 10 and 11 show waveform data when a tool breaks. When a tool breaks, the point where the tool (punch) comes into contact with the workpiece is missing, so no load is generated. Therefore, when the peak load PL1, which is the normal criterion, is set to 100%, a label indicating a tool break is set if the load is 1% or less of the peak load PL1 between the punching start position S1 and the punching end position E1, which are normal criterion (Figure 10).

[0109] Alternatively, if a point occurs between the punching start position S1 and the punching end position E1 of the normal judgment standard where the load is 30% or less of the peak load PL1 of the normal judgment standard, a label of tool breakage is set (Fig. 11). In this way, the same "tool breakage" label may be set for two or more different judgment standards.

[0110] Of the generated multiple judgment criteria, a label indicating that the press working result is normal is set for the normal judgment criterion, and a label indicating the cause of the abnormality is set for the abnormal judgment criterion according to the characteristics of the waveform data as explained in Figures 6 to 11. In this way, the press working result judger 200 is generated.

[0111] [Press processing result estimation method] Fig. 12 is a flowchart showing estimation of the press working result using the press working result determiner 200. A method of estimating the press working result by the press working device 300 using the press working result determiner 200 will be described with reference to Fig. 12.

[0112] First, the press working result determiner 200 acquires sensing data from the press working device 300 (step S21). The sensing data from the press working device 300 is waveform data that indicates the relationship between the position of the punch and the load on the workpiece for one press working shot.

[0113] The press working result determiner 200 calculates the similarity of the acquired sensing data with respect to each determination criterion (step S22). As the similarity, for example, Euclidean distance, Mahalanobis distance, or the sum of Gaussian distribution probability density can be used.

[0114] The press working result determiner 200 determines the determination criterion that has the highest similarity between the sensing data (step S23). The press working result determiner 200 outputs a label attached to the determined determination criterion (step S24).

[0115] [effect] According to the above-described embodiments, it is possible to provide a learned judgment criterion generation method, a press working result judger generation method, a press working result estimation method, a learned judgment criterion generation device, and a press working result judger that can judge process states corresponding to minute changes.

[0116] The method and device for generating learned decision criteria disclosed herein generate multiple decision criteria by classifying sensing data accumulated as a group of training data using unsupervised machine learning. Multiple decision criteria that correspond to minute changes in process state can be automatically generated by machine learning.

[0117] By using a judger including multiple judgment criteria generated in this way, when mass production is performed using automatic operation of a press processing device, it becomes possible to judge the press processing results using sensing data obtained during mass production. That is, it becomes possible to determine whether press processing has completed normally and, if an abnormality occurs, to identify the cause. This makes it possible to detect the cause of an abnormality that occurs in the press processing device, prevent losses due to defects, and perform maintenance without delay, thereby improving productivity. Furthermore, when an abnormality occurs in press processing, it becomes possible to quickly identify the cause, thereby reducing downtime of the press processing device.

[0118] In the above-described embodiment, the step of classifying the training data group 13 into two criteria is repeatedly executed, but the step of classifying the training data group 13 into three or more criteria may also be repeatedly executed.

[0119] Furthermore, in the above-described embodiment, an example has been described in which the training data group 13 is classified using a Gaussian mixture model, but it is also possible to use an appropriate classification algorithm such as the K-Means method.

[0120] In the above-described embodiment, examples of labels to be displayed when the press working result is abnormal are tool wear, tool chipping, double punching, foreign matter contamination, and tool breakage, but the labels for the cause of the abnormality are not limited to these. Appropriate labels can be set according to the press working device. [Industrial Applicability]

[0121] The learned judgment criterion generation method, press working result judger generation method, press working result estimation method, learned judgment criterion generation device, and press working result judger disclosed herein can be widely used for generating judgment criteria for estimating the press working results of a press working device and for estimating the press working results. [Explanation of symbols]

[0122] 11 Storage section 12 Judgment criteria generation section 13 Training data set 31 Die 32 Punch 33 Work 34 Sensors 35 Load sensor 36 Position Sensor 100 Judgment criteria generation device 200 Press processing result judgement device 300 Press processing equipment c1 1st judgment criterion c2 Second judgment criterion

Claims

1. A method for generating judgment criteria for causing a computer to function to output first and second judgment criteria that indicate press processing results of a press processing device, the press processing device having a die and a punch facing the die, which processes a workpiece placed on the die by a load of the punch, based on sensing data that indicates a relationship between a position of the punch and a load applied to the workpiece, the method comprising: a learning data group acquisition step of acquiring a learning data group that statistically significantly includes a normal sensing data group that indicates the relationship between the position of the punch and the load applied to the workpiece when press working is performed using the die, the punch, and the workpiece in a normal state, and that includes an abnormal sensing data group that does not belong to the normal sensing data group; a criterion generation step of performing unsupervised machine learning using the training data group to generate and update a plurality of criterion for classifying the training data group, the criterion generation step generating and updating the first criterion that has been machine-learned to classify the training data group into the normal sensing data group, and the second criterion that has been machine-learned to classify the training data group into the abnormal sensing data group; Including, the determination criterion generating step includes, when a similarity between the generated first determination criterion and the generated second determination criterion is equal to or less than a predetermined threshold, generating a third determination criterion for classifying a first data group classified under the first determination criterion as an input into the normal sensing data group and a fourth determination criterion for classifying the first data group as an abnormal sensing data group, and generating a fifth determination criterion for classifying a second data group classified under the second determination criterion as an input into the abnormal sensing data group and a sixth determination criterion for classifying the second data group as an input into the abnormal sensing data group. A method for generating learned criteria.

2. the first judgment criterion and the second judgment criterion are trained to identify similarities between sensing data belonging to the normal sensing data group in the learning data group and similarities between sensing data belonging to the abnormal sensing data group, respectively. The method of claim 1 .

3. The determination criterion generating step is a step of generating the first determination criterion and the second determination criterion using a Gaussian mixture model, (a) providing a probability density function of a first Gaussian distribution for determining the first criterion and a probability density function of a second Gaussian distribution for determining the second criterion; (b) using a probability density for the first Gaussian distribution and a probability density for the second Gaussian distribution calculated for each piece of sensing data included in the learning data group; calculating a first weighting factor for the first Gaussian distribution and a second weighting factor for the second Gaussian distribution; (c) updating the first and second judgment criteria by updating a probability density function of the first Gaussian distribution and a probability density function of the second Gaussian distribution using the first and second weighting factors; and (d) repeating steps (b) and (c) until a predetermined condition is met; and (e) adopting a probability density function of the first Gaussian distribution and a probability density function of the second Gaussian distribution when the predetermined condition is satisfied as the first judgment criterion and the second judgment criterion; Including, The method of claim 1 .

4. In the step (b), for each of the sensing data included in the learning data group, calculating a likelihood of the sensing data with respect to the first Gaussian distribution; Calculating a likelihood of the sensing data with respect to the second Gaussian distribution; and calculating the first weighting factor and the second weighting factor using a ratio of the likelihoods calculated respectively; Including, The method of claim 3.

5. the step (b) includes calculating, using the Gaussian mixture model, the first weighting coefficient for the first determination criterion for each of the sensing data and the second weighting coefficient for the second determination criterion for each of the sensing data, based on a sum of probability densities for the first Gaussian distribution and the second Gaussian distribution before updating; the step (c) includes updating the first judgment criterion by dividing the average and variance of the product of the sensing data and the first weighting coefficient by the sum of the first weighting coefficients, and updating the second judgment criterion by dividing the average and variance of the product of the sensing data and the second weighting coefficient by the sum of the second weighting coefficients; The method of claim 4.

6. the criterion generating step generates the first criterion and the second criterion using a K-Means method. The method of claim 4.

7. the generating step of generating a judgment criterion is repeatedly executed until a similarity between the first judgment criterion and the second judgment criterion in updating the first judgment criterion and updating the second judgment criterion exceeds a predetermined threshold. The method of any one of claims 1 to 6.

8. A method for generating a press working result determiner based on the first determination criterion and the second determination criterion generated by the generation method according to any one of claims 1 to 7, comprising: obtaining a plurality of criteria including the first criterion and the second criterion; setting a label indicating normality or a cause of abnormality for each of the acquired multiple judgment criteria; Including, Method for generating press processing result judger.

9. A method for estimating a result of press working based on the sensing data during press working of the press working device, using the press working result determiner generated by the method for generating a press working result determiner according to claim 8, comprising: acquiring the sensing data of the press working device; a step of estimating which of the criteria the sensing data is similar to, and outputting a label attached to the criteria that is similar to the sensing data; Including, Method for estimating press working results.

10. A press processing device has a die and a punch facing the die, and processes a workpiece placed on the die using a load of the punch. The press processing device generates a plurality of judgment criteria, including a first judgment criterion and a second judgment criterion, regarding a result of press processing by the press processing device, based on sensing data indicating a relationship between a position of the punch and a load applied to the workpiece, a memory unit for storing a learning data group, the learning data group including statistically significant normal sensing data that indicates the relationship between the position of the punch and the load applied to the workpiece when press working is performed using the die, the punch, and the workpiece in a normal state, and including abnormal sensing data that does not belong to the normal sensing data group; a criterion generation unit that performs unsupervised machine learning using the training data group and generates the plurality of criterion for classifying the training data group, the criterion generation unit generating the first criterion that has been trained to classify the training data group into a normal sensing data group and the second criterion that has been trained to classify the training data group into the abnormal data group; Equipped with the judgment criterion generation unit generates, when a similarity between the generated first judgment criterion and the generated second judgment criterion is equal to or less than a predetermined threshold, a third judgment criterion for classifying the first data group classified under the first judgment criterion as an input into the normal sensing data group and a fourth judgment criterion for classifying the first data group classified under the first judgment criterion as an input into the abnormal sensing data group, and generates a fifth judgment criterion for classifying the second data group classified under the second judgment criterion as an input into the abnormal sensing data group and a sixth judgment criterion for classifying the second data group as an input into the abnormal sensing data group. Learned criteria generator.

11. the plurality of criterion including the first criterion and the second criterion generated by the learned criterion generation device according to claim 10; a label attached to each of the plurality of judgment criteria, which indicates that the press working result of the press working device is normal or indicates the cause of an abnormality; Equipped with determining which of the plurality of criteria the input sensing data is similar to, and outputting the label; Press processing result judgement device.

Images