A bending sequence calculation program that uses a machine learning method and a bending sequence calculation model created therefrom.

Abstract

This invention provides a machine learning method for determining a more appropriate bending sequence and a bending sequence calculation program that uses the resulting bending sequence calculation model. [Solution] A machine learning method according to one aspect of the present invention uses appropriate bending sequence data as training data. A work bending sequence calculation program according to another aspect of the present invention causes a computer to perform the steps of recording work shape data, recording processed product shape data, and creating proposed bending sequence data using a bending sequence calculation model obtained by the machine learning method based on the work shape data and processed product shape data.

Description

Technical Field

[0001] The present invention relates to a machine learning method and a bending sequence calculation program using a bending sequence calculation model generated thereby.

Background Art

[0002] A metal plate member is called a workpiece, which is bent into a complex shape by a press brake device or the like, and the bent workpiece (processed product) is used in a wide range of fields such as automobile, aerospace, construction, and electrical industries.

[0003] Meanwhile, against the background of higher functionality of products and diversification of designs in recent years, product shapes have become more complex. In order to realize such shapes, it has become common to use different dies each time bending is performed, and the sequence for bending complex shapes has also become extremely important.

[0004] As described above, as an apparatus for realizing a complex shape using a plurality of dies, for example, Patent Document 1 below discloses a long press brake device capable of exchanging a plurality of dies.

Prior Art Literature

Patent Literature

[0005]

Patent Document 1

Summary of the Invention

Problem to be Solved by the Invention

[0006] By the way, the aforementioned press brake device is merely an apparatus, and only performs bending processing in accordance with a preset workpiece position, dies, and sequence, so there remains room for study on optimization of the bending sequence. In particular, the more complex the shape is, the more difficult it is to find an optimal bending sequence.

[0007] Therefore, in view of the above problems, the present invention aims to provide a machine learning method for determining a more appropriate bending sequence and a bending sequence calculation program that uses the bending sequence calculation model created thereby. [Means for solving the problem]

[0008] A machine learning method according to one aspect of the present invention uses appropriate bending sequence data as training data.

[0009] Furthermore, although not limited to this perspective, it is preferable that the appropriate bending sequence data includes the appropriate number of bends data and at least one or more appropriate bending process data.

[0010] Furthermore, although not limited to this viewpoint, it is preferable that the appropriate bending process data includes at least one of the appropriate bending position data, appropriate bending angle data, appropriate bending curvature data, appropriate elongation value data, and appropriate mold data.

[0011] Furthermore, although not limited to this perspective, it is preferable that the appropriate mold data includes appropriate upper mold data and appropriate lower mold data.

[0012] Furthermore, although not limited to this perspective, it is preferable that the appropriate bending sequence data include data on the appropriate number of mold changes.

[0013] Furthermore, although not limited to this perspective, it is preferable that the appropriate bending sequence data includes bending sequence image data.

[0014] Furthermore, a workpiece bending sequence calculation program according to another aspect of the present invention causes a computer to perform the steps of recording workpiece shape data, recording processed product shape data, and creating proposed bending sequence data using a bending sequence calculation model obtained by the above-described machine learning method, based on the workpiece shape data and the processed product shape data. [Effects of the Invention]

[0015] In summary, the present invention provides a machine learning method for determining a more appropriate bending sequence and a bending sequence calculation program that uses the bending sequence calculation model created thereby. [Brief explanation of the drawing]

[0016] [Figure 1] This figure shows the external appearance of a press brake device used for processing workpieces. [Figure 2] This is an illustrative diagram of the bending sequence data in the embodiment. [Figure 3] This is an illustrative diagram of the bending sequence image data in the embodiment. [Figure 4] This is an illustrative diagram of the shape of the processed product in the embodiment. [Modes for carrying out the invention]

[0017] Embodiments of the present invention will be described in detail below with reference to the drawings. However, the present invention can be implemented in many different forms and is not limited to the specific examples described in the embodiments and examples below.

[0018] (Machine learning methods) The machine learning method according to this embodiment (hereinafter referred to as "this learning method") is used to create a bending sequence calculation model for a work bending sequence calculation program that creates data (hereinafter referred to as "work bending sequence data") that includes information on the sequence of processes for bending a metal plate-shaped workpiece using a press brake device.

[0019] Specifically, the present learning method is executed using a computer. The computer that executes the present learning method is not particularly limited, but preferably includes, but is not limited to, components of a general computer such as a central processing unit (CPU), non-volatile recording media such as a hard disk and flash memory, volatile recording media such as memory, a bus connecting these components, input devices such as a keyboard and a mouse, and a display device such as a monitor.

[0020] The computer may be a so-called notebook personal computer or a desktop personal computer, and may also be a mobile information terminal that has become widespread in recent years, specifically a so-called smartphone or a tablet terminal. However, in view of processing a large amount of data quickly and accurately, the computer is preferably a stationary computer rather than a mobile information terminal.

[0021] In addition, the present learning method can also be implemented via telecommunication such as the Internet. Specifically, it is possible to connect a plurality of computers via telecommunication, and cause these different computers to execute the learning steps in the present learning method.

[0022] As will be described later, by using a bending sequence calculation model for a workpiece bending sequence calculation program obtained as a result of the present learning method, a workpiece bending sequence can be calculated to generate proposed bending sequence data. Then, by controlling a press brake device as shown in FIG. 1 in accordance with the proposed bending sequence data to bend the workpiece, a processed product having a final shape can be obtained.

[0023] (Data Preparation Step) First, this learning method is characterized by using appropriate bending sequence data as training data. However, as part of the machine learning process of this learning method, it is important to first prepare a large amount of appropriate bending sequence data, that is, it has a data preparation step. Here, "appropriate bending sequence data" refers to data that contains information about the appropriate bending sequence for a workpiece. Here, "bending sequence" refers to the order of bending processes from a flat workpiece to its final processed product shape through multiple bending processes. "Appropriate bending sequence" refers to a bending sequence that is suitable to adopt from among the many combinations (paths) of bending processes performed from a single workpiece to a single final processed product shape, considering factors such as the number of bending processes and the number of times the die is changed. Preferably, it refers to an optimized bending sequence. The selection of this appropriate bending sequence data may be determined empirically by the person in charge of bending the workpiece, or it may be determined by weighting the number of bending processes and the number of times the die is changed, and the bending sequence that results in the minimum or maximum sum of these values ​​may be considered the appropriate bending sequence. Furthermore, while it is preferable that there is, in principle, only one correct bending sequence for a given final product shape, it is possible to have two or more. In other words, this learning method includes evaluating whether the bending sequence is correct (or incorrect), and supervised learning is preferable.

[0024] Furthermore, the bending sequence data used in this learning method can be created separately by a workpiece bending sequence calculation program. This workpiece bending sequence calculation program can create multiple bending sequence data sets by taking the workpiece shape before processing and the processed product shape (the final form) as input data. However, these multiple bending sequence data sets do not include data that evaluates the bending sequence, i.e., information on whether it is appropriate or not. Therefore, in this learning method, by using appropriate bending sequence data, which includes information on whether the bending sequence is appropriate, as the correct data, a machine learning method for finding a more appropriate bending sequence can be realized.

[0025] The specific details of these will become clear from the description below, but it is preferable to include at least one set of appropriate bending process data. Here, "appropriate bending process data" refers to optimized bending process data included in the appropriate bending sequence data, and "bending process data" refers to data that includes information about the bending process. The information about the bending process here is not limited, but it is preferable to include at least one of the following: bending position data including information about the bending position of the workpiece, bending angle data including information about the bending angle of the workpiece, bending curvature data including information about the bending curvature of the workpiece, elongation value data including information about the elongation value of the workpiece, mold data including information about the mold used in this bending process, and data including information about the bending sequence (information about which bending process is performed). As a result, in the case of this appropriate bending process data, it will include at least one of the appropriate bending position data, appropriate bending angle data, appropriate bending curvature data, appropriate elongation value data, and appropriate mold data. Figure 2 shows an illustrative diagram of the bending sequence data. As shown in this figure, the bending sequence data includes multiple bending process data, each of which includes the bending position, bending angle, bending curvature, etc. By following these, the workpiece can be bent into a predetermined processed product shape.

[0026] Furthermore, it is preferable that the appropriate bending process data also include bending sequence image data. Bending sequence image data is data that includes information about the bending sequence, and specifically includes information such as the shape of at least one of the workpieces before and after each bending process, the position where the workpiece is supported by the upper die, lower die, back gauge, etc., and image information that visualizes this process. This learning method has the advantage that by including this bending sequence image data in the appropriate bending process data, it becomes possible to determine a more appropriate bending sequence. In addition, in this learning method, not only is this bending sequence image data associated with each bending process data, but multiple bending sequence image data may also be arranged in the order of the bending processes and compiled into a video showing the entire flow from the workpiece to the final processed product shape. Here, an example of bending sequence image data is shown in Figure 3.

[0027] Furthermore, in this learning method, the appropriate bending sequence data may include bending sequence image data as described above, and may also include workpiece shape data before bending and processed product shape data. "Workpiece shape data before bending" refers to data that includes information about the workpiece shape before bending, and it is preferable that this information includes, for example, information about length, width, thickness, and material. "Processed product shape data" refers to data that includes information about the processed product shape, which is the final form, and it is preferable that this data includes not only information about the shape of the processed product, but also information such as bending locations and bending angles. Figure 4 shows an image of the processed product shape data in this learning method. In the figure, (a) shows a cross-section of the processed product, and (b) shows a perspective view of the processed product.

[0028] Furthermore, in the above, it is preferable that the appropriate mold data includes "appropriate upper mold data" and "appropriate lower mold data." Here, "appropriate upper mold data" refers to data that includes information about an appropriate upper mold, and "appropriate lower mold data" refers to data that includes information about an appropriate lower mold. Since identifying the upper and lower molds is essential in the bending process, it is important to include this in the mold data. Note that mold data other than appropriate will include upper mold data and lower mold data.

[0029] Furthermore, in this learning method, it is preferable that the appropriate bending sequence data includes appropriate bending count data and at least one appropriate bending process data. Here, "appropriate bending count data" refers to the number of appropriate bending process data included in the appropriate bending sequence data. Note that in the case of non-appropriate bending sequence data, "bending count data" refers to data that includes information about the number of bends in that bending sequence.

[0030] Furthermore, in this learning method, it is preferable that the appropriate bending sequence data includes data on the appropriate number of mold changes. Here, "appropriate number of mold changes data" refers to data that includes information on the number of mold changes included in the appropriate bending sequence data. Note that when "number of mold changes data" is used for bending sequence data other than the appropriate one, it refers to data that includes information on the number of mold changes in that bending sequence.

[0031] Furthermore, in this learning method, in addition to the "appropriate bending sequence data" mentioned above, it is preferable to also use inappropriate bending sequence data as training data for the appropriate bending sequence data. Here, "inappropriate bending sequence data" refers to bending sequence data other than those judged to be appropriate among the bending sequence data for transforming a workpiece into its final processed product shape. Specifically, it refers to bending sequence data other than the appropriate bending sequence data among the multiple bending sequence data created to transform a flat workpiece into a processed product shape before bending. By using this inappropriate bending sequence data as training data, it becomes possible to determine a more appropriate bending sequence.

[0032] Furthermore, in this learning method, it is preferable to use the bending sequence calculation model for the workpiece bending sequence calculation program created by this learning method to calculate the workpiece bending sequence, determine whether the created proposed bending sequence data is appropriate or not, and add the evaluation data to the training data as appropriate bending sequence data or inappropriate bending sequence data. In this way, it becomes possible to find an even more appropriate bending sequence.

[0033] (Data preprocessing step) Furthermore, in this learning method, it is important to perform data preprocessing on the collected appropriate bending sequence data before the process described later; in other words, it includes a data preprocessing step. Performing data preprocessing can further improve the accuracy of machine learning.

[0034] The data preprocessing steps here are not limited to those mentioned above, but it is preferable to include more detailed steps such as a data cleaning step, a normalization step, and a feature extraction step.

[0035] Here, "data cleanup" can be exemplified by unifying the format of properly bent-sequence data, correcting missing or outlier data, and deleting duplicate data.

[0036] Furthermore, "normalization" refers to the process of unifying the scale of data, reducing variability between features, and streamlining learning to improve its effectiveness. For example, scaling, which limits the maximum and minimum values ​​of each data point to a predetermined range, is an example of normalization.

[0037] Furthermore, "feature extraction processing" refers to the process of extracting characteristic values ​​(features) that are useful for machine learning from the included data. This feature extraction processing can utilize publicly known methods, such as aggregation, filtering, and principal component analysis, but is not limited to these.

[0038] (Model selection step) Furthermore, in this learning method, it is preferable to have a model selection step, i.e., a model selection step, after performing the data preprocessing described above. In machine learning, selecting an appropriate learning model can further enhance the effectiveness of the machine learning process. While one learning model may be used, it is also preferable to employ multiple learning models.

[0039] Here, the models are not limited, but it is preferable to adopt an appropriate model depending on the task to be solved. For example, for regression tasks, examples include linear regression models, ridge regression models, LASSO regression models, XGBoost models, and neural network models. For classification tasks, examples include logistic regression models, support vector machine models, decision tree models, random forest models, XGBoost models, and neural network models. For clustering tasks, examples include K-means models, hierarchical clustering models, and DBSCAN models. For time series forecasting tasks, examples include ARIMA models, SARIMA models, LSTM models, and Prophet models. For anomaly detection tasks, examples include one-class SVM models, isolation forest models, and local outlier factor models.

[0040] (Model training steps) Furthermore, this learning method includes a model learning step after the model selection step, in which the model is specifically trained. Specifically, in this step, the selected model is trained to learn patterns from the appropriate bending sequence data, enabling the prediction of appropriate bending sequence data.

[0041] (Model evaluation step) Furthermore, this learning method preferably includes a model evaluation step in which the model is evaluated after model training. In this case, it is preferable to perform the evaluation by creating evaluation metric data. Here, "evaluation metric" refers to information used to evaluate the performance of the model, for example, in a regression task, MSE, RMSE, R 2 Examples include, but are not limited to, those listed above.

[0042] In summary, this learning method is a machine learning method for determining a more appropriate bending sequence. This process creates a bending sequence calculation model, which can then be used in the workpiece bending sequence calculation program described later.

[0043] (Workpiece bending sequence calculation program) Furthermore, in this embodiment, as described above, the bending sequence of a workpiece can be calculated using the bending sequence calculation model created as a result of this learning method. More specifically, it is possible to provide a workpiece bending sequence calculation program (hereinafter referred to as "this program") that causes a computer to perform the steps of recording workpiece shape data, recording processed product shape data, and creating proposed bending sequence data using the bending sequence calculation model obtained by the above machine learning method based on the workpiece shape data and the processed product shape data.

[0044] Specifically, this program is recorded on a non-volatile recording medium such as a computer's hard disk, and when requested by the user, it is read to a volatile recording medium such as memory and executed, thereby enabling the computer to execute a work bending sequence calculation method (hereinafter referred to as "this calculation method") which includes the steps of (S1) recording work shape data, (S2) recording processed product shape data, and (S3) creating proposed bending sequence data based on the work shape data and the processed product shape data using a bending sequence calculation model obtained by the machine learning method described above. The following describes each step of this calculation method implemented by this program.

[0045] This calculation method first includes the step of recording workpiece shape data (S1). Here, "workpiece shape data" refers to data that includes information about the shape of the workpiece before bending, and specifically includes, but is not limited to, data on dimensions such as length, width, and thickness of a flat workpiece, and data on material.

[0046] Furthermore, this calculation method includes a step (S2) of recording the shape data of the workpiece. Here, "shape data of the workpiece" refers to data that includes information about the shape of the workpiece. The information about the shape of the workpiece is not limited to this, but it is preferable to have, for example, bending position data that includes information about the bending position of the workpiece, bending angle data that includes information about the bending angle of the workpiece, bending curvature data that includes information about the bending curvature of the workpiece, and elongation value data that includes information about the elongation value of the workpiece, corresponding to each of the multiple bending points.

[0047] Furthermore, this calculation method includes a step of creating proposed bending sequence data using a bending sequence calculation model obtained by the machine learning method described above, based on (S3) workpiece shape data and processed product shape data. Here, the obtained "proposed bending sequence data" refers to the bending sequence data as a proposed example created by this calculation method, and corresponds to the appropriate bending sequence data used as training data in this learning method. In addition, this proposed bending sequence data can adopt a data structure similar to the bending sequence data described above, and preferably includes, for example, proposed bending count data, proposed bending process data, and proposed die change count data, and it is preferable that the proposed bending process data further includes "proposed bending position data", "proposed bending angle data", "proposed bending curvature data", "proposed elongation value data", and "proposed die data".

[0048] Furthermore, this step can be divided into two steps: (S3-1) creating multiple bending sequence data, and (S3-2) selecting one of the multiple bending sequence data as the proposed bending sequence data. In this step, the conventional bending sequence calculation program is used to create multiple bending sequence data based on the workpiece shape data and the processed part shape data in a brute-force manner. Subsequently, from the created multiple bending sequence data, it is possible to create a proposed bending sequence data that is presumed to be the appropriate bending sequence data using a bending sequence calculation model.

[0049] In summary, this embodiment provides a machine learning method for determining a more appropriate bending sequence and a bending sequence calculation program that uses the bending sequence calculation model created thereby. [Industrial applicability]

[0050] The present invention has industrial applicability as a machine learning method and a bending sequence calculation program.

Claims

1. A machine learning method that uses data on the optimal bending sequence as training data.

2. The machine learning method according to claim 1, wherein the appropriate bending sequence data includes appropriate bending count data and at least one or more appropriate bending process data.

3. The machine learning method according to claim 2, wherein the appropriate bending process data includes at least one of appropriate bending position data, appropriate bending angle data, appropriate bending curvature data, appropriate elongation value data, and appropriate mold data.

4. The machine learning method according to claim 3, wherein the appropriate mold data includes appropriate upper mold data and appropriate lower mold data.

5. The machine learning method according to claim 1, wherein the appropriate bending sequence data includes data on the number of appropriate mold changes.

6. The machine learning method according to claim 1, wherein the appropriate bending sequence data includes bending sequence image data.

7. On the computer, Steps to record workpiece shape data, Steps to record the shape data of the processed product, A workpiece bending sequence calculation program for performing the step of creating proposed bending sequence data using a bending sequence calculation model obtained by the machine learning method described in claim 1, based on the workpiece shape data and the processed product shape data.

Images