Method for augmenting training data for training machine learning models, and system

Abstract

The invention relates to a method for augmenting training data for training machine learning models (25), which method is used for suggestions (16) for recommending process parameters and / or for predicting a quality of produced components and / or products, wherein the suggestions are forwarded to an environment (14) in which input variables (22) for the machine learning model (25) are detected, said input variables are interpreted in a space (26) of the input variables (22) as original points, wherein, on the basis of a predefined distribution of a measurement uncertainty, hyper-cuboids (21) are determined for each original point, and in each hyper-cuboid (21) a set of further new artificial points are derived for further artificial input variables (24) and have a respective relationship to their corresponding original point; and the original points are applied with their set of corresponding new artificial points for the training of the machine learning model (25), whereby respective result variables are derived by means of the model (25). The invention further relates to a system (1).

Description

[0001] Methods for augmenting training data for training machine learning models and systems

[0002] The invention relates to a method for augmenting training data for training machine learning models according to claim 1. The invention further relates to a system.

[0003] Machine learning algorithms are typically used to implement process control. However, since process parameters are rarely changed, data is often only available for a limited number of parameter combinations. In this context, all process parameters can be combined into a single vector. If the value of one or more process parameters is changed, this new vector will differ from the original one in one or more components. Both of these vectors can be interpreted as different combinations of process parameters. Furthermore, the training data for machine learning models can become invalid if, for example, changes are made to manufacturing processes.

[0004] This shows that in production, it can happen that there is (too) little data available for a training model for machine learning.

[0005] The object of the invention is to avoid or reduce downtime when finding process parameters for problem solving.

[0006] This problem is solved by means of a method with the features of claim 1 and by means of a system according to the invention. Advantageous embodiments and further developments can be found in the dependent claims and in the description.

[0007] One aspect of the invention relates to a method for augmenting training data for machine learning models, which is used to suggest process parameter recommendations and / or predict the quality of manufactured components and / or products, wherein the suggestions are forwarded to an environment. Here, input variables for the machine learning model are acquired and represented as original points in a space of input variables. In particular, process parameters, semi-finished product properties, and / or state variables in the manufacturing process of the components are defined as input variables. The augmentation method is intended to be carried out especially for semi-finished product properties and state variables. That is, each input variable is assigned an original point in the space, thereby filling the space with the respective original points.Based on a predefined distribution of measurement uncertainty, a respective hypercube is determined for each original point.

[0008] Thus, artificial points in a multidimensional space, in particular a space with more than three dimensions, can be created from the input sizes for the machine learning model to enable a more complex set of data for training the machine learning.

[0009] Within each hypercube, a set of additional new artificial points is derived for further artificial input variables. These artificial points have a corresponding label to their respective original point and can thus be viewed as a set of new artificial points similar to the original point. For example, such a label could be assigned to all artificial points corresponding to a given hypercube, using the same label as the original point of that hypercube. The original points, along with their set of corresponding new artificial points, are then used to train the machine learning model, thereby deriving the respective output variables.

[0010] In other words, a method for augmenting training data for machine learning models is proposed for recommending process parameters and / or predicting the quality of manufactured components or products. It is then proposed to derive further artificial input variables for each dataset of the original training data, consisting of the input variables. This involves varying the input variables to generate the artificial input variables. Thus, for each dataset of the training data, n additional artificial datasets can be generated, each consisting of a variation of the respective original input variables and the corresponding new artificial input variables.

[0011] This makes it possible to meaningfully expand the training dataset and thus improve the performance of machine learning algorithms. A particular advantage of this method is that this type of training data augmentation can be performed with very little computational effort.

[0012] In an advantageous embodiment of the invention, the respective hypercubes are derived using a predetermined significance level. Measured quantities can then be acquired and augmented (measured quantities are introduced as input variables into the space containing the respective hypercubes), whereby a hypercube is determined based on a distribution of measurement uncertainty and derived using a desired and technically appropriate significance level. For example, a significance level of 90% could preferably be chosen.

[0013] In a further advantageous embodiment of the invention, it is provided that, by means of a restriction, only certain artificial points within the hypercube are used. Accordingly, the restriction consists in the fact that the training is only carried out with the new artificial input parameters that lie within the restriction area within the hypercube, in order to enable a complex training variant. For example, the dimensions of the hypercube can be reduced to apply fewer or only certain artificial input parameters.

[0014] In a further advantageous embodiment of the invention, a multivariate distribution of measurement uncertainty is determined, thereby taking into account relationships between the individual input variables. The generation of the artificial input variables can thus also be carried out in a different way. For example, a distribution of measurement uncertainty can again be determined. In this case, it is therefore a multivariate distribution of measurement uncertainty that also takes into account relationships between the individual input variables. Artificial points are then generated from this multivariate distribution, which can also be recorded and used as artificial input variables. The multivariate distribution again refers to one of the original points of the original training dataset or to a dataset consisting of the initial input variables.In a further advantageous embodiment of the invention, it is provided that overlapping spaces of overlapping hypercubes within the space are detected, for example, by means of an algorithm. Overlaps occur when the hypercubes are so large that adjacent points do not have enough space between them to allow a clear separation. These overlaps can also arise or even increase when the dimensions of the hypercubes are enlarged, particularly when the points are close together. In such cases, it is provided that these spaces, in which points of two hypercubes can be located, are identified in order to avoid distorted or inaccurate results.

[0015] In a further advantageous embodiment of the invention, points located within the overlap space are deleted. These overlap areas can lead to certain regions of the space being overrepresented. Deleting points in these areas / overlap spaces helps, for example, to avoid distortions or imbalances in the artificial input variables.

[0016] In a further advantageous embodiment of the invention, points are generated using a sampling method. This sampling method allows, for example, n points to be evenly distributed in space, ensuring that every area of ​​the space is adequately covered. Thus, no regions of the space are neglected, and a particularly good and representative representation of the entire space is provided.

[0017] Further features of the invention will become apparent from the claims, the figures, and the description of the figures. The features and combinations of features mentioned above in the description, as well as the features and combinations of features mentioned below in the description of the figures and / or shown in the figures alone, are not only usable in the combinations specified, but also in other combinations or on their own.

[0018] The invention will now be explained in more detail with reference to a preferred embodiment and the drawings. The drawings show:

[0019] Fig. 1 is a diagram illustrating a system for manufacturing components, in which a method for augmenting training data for training machine learning models can be carried out by means of an arranged computing device.

[0020] Fig. 1 shows in a diagram a system 10 for manufacturing components, in which a method for augmenting (in space 26) training data for training (in model 25) of models 25 for machine learning can be carried out by means of an arranged computing device 12.

[0021] In this procedure, proposals 16 are applied for a recommendation of process parameters and / or for a prediction of the quality of produced components and / or products, whereby the proposals can be forwarded to an environment 14 by means of the electronic computing device 12.

[0022] First, input variables 22 for the machine learning model 25 can be recorded and considered as original points in a space 26 of input variables 22. Based on a predefined distribution of measurement uncertainty, a respective hypercube 21 is determined for each original point, and in each respective hypercube 21, a set of further new artificial points for further artificial input variables 24 is derived. Each of these new artificial points can have a respective reference or label to its corresponding original point for a unique assignment.

[0023] The original points can thus be used with their set of corresponding new artificial points for training the model 25 for machine learning.

[0024] For example, this process can be integrated into the production of body components or components 20 in press shops, a multi-step process. One such step would be cutting blanks 16 from a coil on a coil line. These blanks 16 are then sorted into stacks for storage before processing in the press line. In the press line, the sheet metal or the cut blanks 16 are typically deep-drawn, followed by further steps such as trimming or reshaping.

[0025] The materials processed in press shops often exhibit fluctuations in their properties. For example, sheet thickness, lubricant quantity, surface roughness, and elastoplastic material properties can vary. To achieve the required quality of the manufactured components, the process parameters must be adjusted accordingly. In most press shops, this is done based on experience by the machine operators, which can lead to downtime and significant costs.

[0026] Typical process parameters, such as those affecting the press, include draw cushion forces, the kinematics of the ram movement, the positioning of the blank holder, and the application of lubricant. Further process parameters include tool parameters such as draw aids, gas filling of gas springs, and the position of guides.

[0027] However, there is a challenge: Often, only limited data is available for training machine learning models, as process parameters rarely change. Therefore, it can happen that there is insufficient data for training, especially when manufacturing processes change.

[0028] These process parameters can be acquired by the electronic computing device 12 as input variables 22 for the machine learning model 25. The device is configured to interpret these parameters as points in the space 26 of the input variables 22 and to assign further artificial points within a predefined hypercube 21 to these points, thereby enabling at least partial training of the machine learning model 25. Furthermore, a set of points for training the machine learning model 25 can be expanded from these new artificial points and the associated additional new artificial input variables 24.

[0029] Thus, the Model 25 for machine learning can gather enough information to make accurate predictions or provide appropriate process parameter recommendations.

[0030] Reference symbol list

[0031] Electronic computing system

[0032] Vicinity

[0033] Suggestions

[0034] circuit boards

[0035] components

[0036] Hypercube

[0037] Input sizes new artificial input sizes

[0038] Model

[0039] Space

Claims

Patent claims 1. Method for augmenting training data for training models (25) for machine learning, which is applied for suggestions (16) for recommending process parameters and / or for predicting the quality of produced components and / or products, wherein the suggestions are forwarded to an environment (14) in which: - Input parameters (22) for the model (25) for machine learning are recorded, - these are considered as original points in a space (26) of input quantities (22), - wherein a respective hypercube (21) is determined for each original point based on a given distribution of measurement uncertainty, - and in each hypercube (21) a set of further new artificial points for further artificial input quantities (24) are derived, each having a respective reference to its corresponding original point; and - the original points with their set of corresponding new artificial points are applied to the training of the machine learning model (25), thereby deriving respective outcome quantities using the model (25).

2. Method according to claim 1, characterized in that the respective hypercubes (21) are derived using a predetermined significance.

3. Method according to claim 1 or 2, characterized in that only certain artificial points within the hypercube (21) are used by means of a restriction.

4. Method according to one of the preceding claims, characterized in that within the space (26) the distribution of the measurement uncertainty of all input variables (22) is constant.

5. Method according to one of the preceding claims, characterized in that a multivariate distribution of the measurement uncertainty is determined, whereby relationships between the individual input variables (22) are taken into account.

6. Method according to one of the preceding claims, characterized in that Overlap spaces of overlapping hypercubes (21) within the space (26) are captured.

7. Method according to claim 6, characterized in that Points located within the overlap area will be deleted.

8. Method according to one of the preceding claims, characterized in that Points are generated using a sampling method.

9. System (10) for manufacturing components (20), in which a method for augmenting training data for training models (25) for machine learning according to one of the preceding claims 1 to 8 can be carried out by means of an arranged computing device (12) and which is applicable for suggestions (16) for a recommendation of process parameters and / or for a prediction of the quality of produced components and / or products, wherein an electronic computing device (12) arranged to the system (10) is configured to transmit the suggestions (16) to an environment (14).

Images