Method and apparatus for revealing the effect of cutting parameters on the cutting edge

Abstract

The invention relates to a method for identifying cutting parameters (18) which are of particular importance for specific features of a cutting edge (16). Here, a recording (32) of the cutting edge (16) is analyzed by an algorithm (34) having a neural network (36) in order to determine (38) the cutting parameters (18). Those recording pixels (42a, b) which play an important role for determining the cutting parameters (18) are identified by backpropagation (40) of the analysis. An output (50) in the form of a representation of the important recording pixels (42a, b), in particular in the form of a heat map, reveals to a user of the method which cutting parameters (18) need to be changed in order to improve the cutting edge (16). The invention also relates to a computer program product and a device for performing the method, respectively.

Description

TECHNICAL FIELD

[0001] The invention relates to a method for analyzing a cutting edge created by a machine tool. The invention also relates to a computer program product and a device for carrying out the method. BACKGROUND

[0002] It is known to optimize the cutting of workpieces. By way of example, DE 10 2017 105 224 A1 discloses the use of a neural network to regulate a laser cutting process.

[0003] However, the influence of most cutting processes or cutting parameters on the cutting edge is not yet fully understood. This can be inferred from the following articles, for example:

[0004] Hügel, H., Graf, T. Laser in der Fertigung: Strahlquellen, Systeme, Fertigungsverfahren [Lasers in manufacturing: beam sources, systems, manufacturing methods]. Wiesbaden: Vieweg + Teubner, 2009.

[0005] Petring, D., Schneider, F., Wolf, N. Some answers to frequently asked questions and open issues of laser beam cutting. In: International Conference on Applications of Lasers and Electro-Optical Systems. ICALEO OR 2012, Anaheim, CA, USA: Laser Institute of America, 2012, pp. 43-48

[0006] Steen, W. M., Mazumder, J. Laser Material Processing. London: Springer London, 2010.

[0007] Even experienced users of cutting devices often cannot predict how a cutting parameter will influence the appearance of a cutting edge. Therefore, in order to improve the appearance of a cutting edge, especially in the case of problems due to a change in material quality and / or in the case of a new process using a new laser source, a new sheet metal thickness, etc., it is necessary to carry out a complex test series at regular intervals in which the cutting parameters are varied in different ways. SUMMARY

[0008] By contrast, it is the task of the present invention to provide a method, computer program product and device which enable an analysis of the influence of cutting parameters on the cutting edge.

[0009] The task is solved by a method for analyzing a cutting edge created by a machine tool according to the invention, a computer program product according to the invention and a device according to the invention.

[0010] The task of the present invention is therefore solved by a method in which a recording of a cutting edge created by a machine tool is read, the recording having a plurality of recording pixels. At least one cutting parameter, in particular a plurality of cutting parameters, is determined from the recording using an algorithm having a neural network. Subsequently, a backpropagation is carried out in the neural network in order to determine a relevance of the individual recording pixels for determining the previously determined cutting parameter. The recording is then output, wherein at least some of the recording pixels are labeled, the label reflecting the relevance of the previously determined recording pixels. Preferably, all recording pixels are output and labeled according to their relevance.

[0011] The user can therefore immediately recognize from the labeled output how strongly the respective region of the cutting edge is influenced by one or more respective cutting parameters and can then carry out an adjustment of the one or more cutting parameters in order to change the specific region of the cutting edge in a targeted manner.

[0012] The backpropagation of the neural network is disclosed, for example, by EP 3 654 248 A1, the content of which is hereby incorporated by reference in its entirety into the present specification.

[0013] Generally, such a backpropagation (“backpropagation-based mechanism”) is only used to check whether the neural network has learned the correct relationship. What is meant here is a neural network without “superhuman performance”. In this case, a human can assess without any problems what is the correct information. For example, it can be checked whether a neural network that can distinguish between dogs and cats really recognizes a dog in the image when indicating the presence of a dog, and not the grass on which the dog is standing. By way of example, it can happen that the neural network does not recognize a specific animal in the image (e.g. a horse), but rather a letter that can be seen in the image, in all images of horses (so-called “clever Hans problem”). By contrast, in the present case, the backpropagation is used in order to understand or at least be able to predict the production process or physical relationship in the cutting process.

[0014] Herein, a neural network is understood to be an architecture having at least one data aggregation routine, in particular a plurality of data aggregation routines. The data aggregation routines can be designed to aggregate a plurality of "sought data" to form a new data package. The new data package can comprise one or more numbers or vectors. Further data aggregation routines can be provided completely or partially as "sought data" to the new data package. The "sought data" can in particular be a cutting parameter or a data package provided by one of the data aggregation routines. It is particularly preferred that the architecture is configured with a plurality of connected data aggregation routines. In particular, several hundred, in particular several thousand, such data aggregation routines can be connected to one another. Thereby, the quality of the neural network is greatly improved.

[0015] Herein, the architecture can have a function with a weighting variable. One data aggregation routine, in particular a plurality of data aggregation routines, particularly preferably all data aggregation routines, can be designed to combine, in particular multiply, each of a plurality of "sought data" with a weighting variable, thereby transforming the "sought data" into "combined data", in order to then aggregate, in particular add, the "combined data" to form a new data package. In the neural network, data can be multiplied by a weight. Information of a plurality of neurons can be added. Furthermore, the neural network can have a non-linear activation function.

[0016] Herein, the cutting edge features contained in the record can themselves be data packages, in particular a plurality of structured data, in particular data vectors or data arrays, which themselves can constitute "sought data", in particular for the data aggregation routines.

[0017] In order to determine suitable weighting variables, i.e. in order to train the neural network, the method can be run with data, in particular cutting parameters, the association of which with the records is known separately.

[0018] Herein, the neural network is preferably constituted in the form of a convolutional neural network (CNN) having a plurality of layers. The convolutional neural network can have convolutional layers and pooling layers. The pooling layers are usually arranged between two successive convolutional layers. As an alternative or in addition thereto, pooling can be performed after each convolution.

[0019] In addition to the convolutional layers and the pooling layers, the CNN can also have fully connected layers, in particular at the end of the neural network. The convolutional layers and the pooling layers extract features, and the fully connected layers can assign the features to the cutting parameters.

[0020] Each layer of the neural network can have a plurality of filters. The structure of a convolutional neural network can be derived, for example, from the following articles, in particular the first one mentioned below:

[0021] LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521 :436{444, DOI 10.1038 / nature14539;

[0022] Lin H, Li B, Wang X, Shu Y, Niu S (2019) Automated defect inspection of LED chip using deep convolutional neural network. J Intell Manuf 30:2525{2534, DOI 10.1007 / s10845-018-1415-x;

[0023] Fu G, Sun P, Zhu W, Yang J, Cao Y, Yang MY, Cao Y (2019) A deep-learning-based approach for fast and robust steel surface defects classification. Opt Laser Eng 121 :397{405, DOI 10.1016 / j.optlaseng.2019.05.005;

[0024] Lee KB, Cheon S, Kim CO (2017) A Convolutional Neural Network for Fault Classification and Diagnosis in Semiconductor Manufacturing Processes. IEEE T Semiconduct M 30:135{142, DOI 10.1109 / TSM.2017.2676245;

[0025] DA, Stemmer MR, Pereira M (2020) A convolutional neural network approach on bead geometry estimation for a laser cladding system. Int J Adv Manuf Tech 106: 1811-1821, DOI 10.1007 / s00170-019-04669-z;

[0026] Karatas A, D, Schmidt S, Eier M, Seewig J (2019) Development of a convolutional autoencoder using deep neuronal networks for defect detection and generating ideal references for cutting edges. Munich, Germany, DOI 10.1117 / 12.2525882;

[0027] Stahl J, Jauch C (2019) Quick roughness evaluation of cut edges using a convolutional neural network. In: SPIE Conference Proceedings 11172, Munich, Germany, DOI 10.1117 / 12.2519440.

[0028] For backpropagation, the layer-wise relevance propagation (LRP) has proven to be particularly targeted and can be implemented in a simple manner at the same time. The layer-wise relevance propagation can be derived in particular from the following articles:

[0029] Bach S, Binder A, Montavon G, Klauschen F, Muller KR, Samek W (2015) On Pixel- Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLoS ONE 10: e0130140, DOI 10.1371 / journal.pone.0130140;

[0030] W Samek, A Binder, G Montavon, S Lapuschkin, K Muller (2017) Evaluating the Visualization of What a Deep Neural Network Has Learned. IEEE T Neur Net Lear 28: 2660{2673, DOI 10.1109 / TNNLS.2016.2599820;

[0031] Montavon G, Lapuschkin S, Binder A, Samek W, Muller KR (2017) Explaining NonLinear Classification Decisions with Deep Taylor Decomposition. Pattern Recognition 65: 211{222, DOI 10.1016 / j.patcog.2016.11.008;

[0032] Montavon G, Lapuschkin S, Binder A, Samek W, Muller KR (2015) Explaining NonLinear Classification Decisions with Deep Taylor Decomposition. arXiv preprint URL https: / / arxiv.org / pdf / 1512.02479.pdf;

[0033] Montavon G, Binder A, Lapuschkin S, Samek W, Muller KR (2019) Layer-Wise Relevance Propagation: An Overview. In: Samek W, Montavon G, Vedaldi A, Hansen LK, Muller KR (eds) Explainable AI: Interpreting, Explaining and Visualizing Deep Learning. Springer, Cham, pp 193-209.

[0034] Backpropagation, in particular in the form of layer-wise relevance propagation, is preferably based on deep Taylor decomposition (DTD). Deep Taylor decomposition can in particular be derived from the following article:

[0035] Montavon G, Lapuschkin S, Binder A, Samek W, Muller KR (2017) Explaining Nonlinear Classification Decisions with Deep Taylor Decomposition. Pattern Recognition 65:211{222, DOI 10.1016 / j.patcog.2016.11.008.

[0036] For example, the implementation can be realized in Python in the form of the library TensorFlow 1.13.1 (see Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, Corrado G, Davis A, Dean J, Devin M, Ghemawat S, Goodfellow I, Harp A, Irving G, Isard M, Jia Y, Jozefowicz R, Kaiser L, Kudlur M, Zheng X (2016) TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems. arXiv preprint arXiv:1603.04467. URL https: / / arxiv.org / pdf / 1603.04467.pdf) and Keras 2.2.4 (see Chollet F (2015) Keras. URL https: / / keras.io). Furthermore, the Python library "iNNvestigate" (Alber, M. et al.: iNNvestigate neural networks! J. Mach. Learn. Res. 20(93), 1-8 (2019)) can be used.

[0037] More preferably, the output is realized in the form of a heat map. The heat map can have two colors, in particular red and blue, which identify particularly relevant and particularly irrelevant record pixels, respectively. Record pixels of average relevance can be identified by an intermediate color tone or gray between the two colors. Thus, the output is particularly intuitively understandable.

[0038] The record is preferably a photo, particularly preferably a color photo, in particular in the form of an RGB photo or a 3D point cloud. The creation of 3D point clouds is slightly more complex, as they include depth information. The depth information can be obtained during the creation of the record, in particular by light sectioning or by triangulation from different angles. However, it has been found that color photos are particularly suitable or sufficient, as the neural network mainly recognizes the various cutting parameters from the different colors of the cutting edges.

[0039] The record can be created by a camera and / or a video camera. Preferably, the video camera is part of the machine tool in order to ensure a continuous recording situation. As an alternative or in addition thereto, the camera can be part of a photo booth in order to reduce the environmental influences during the creation of the record.

[0040] More preferably, the method of the application comprises creating the cutting edge by means of a machine tool. The cutting method of the machine tool can be a thermal cutting method, in particular a plasma cutting method, preferably a laser cutting method.

[0041] In the case of a laser cutting method, the cutting parameters determined preferably include the beam parameters, in particular the focal point diameter and / or the laser power, the transport parameters, in particular the focal point position, the nozzle-focal point distance and / or the feed, the gas dynamics parameters, in particular the gas pressure and / or the nozzle-workpiece distance, and / or the material parameters, in particular the gas purity and / or the melting temperature of the workpiece. These cutting parameters have proven to have a particular influence on the appearance of the cutting edge.

[0042] The task of the application is also achieved by a computer program product for carrying out the calculation operations described here. The computer program product can be configured, in particular completely, in a cloud-based manner in order to enable access to the computer program product by a plurality of users. Furthermore, a more comprehensive training of the neural network can take place by a plurality of users.

[0043] Finally, the task of the application is achieved by a device for carrying out the method described here, wherein the device comprises a machine tool, in particular in the form of a laser cutting machine.

[0044] Here, the device can comprise the camera described here.

[0045] Further advantages of the application result from the description and the drawings. Likewise, according to the application, the above-mentioned features and those to be further explained can be used individually, respectively, on their own or in any desired combination as multiple uses. The embodiments shown and described are not to be understood as exhaustive, but rather have exemplary properties for outlining the application. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A schematic diagram of a machine tool in the form of a laser cutting machine is shown to illustrate the basic cutting parameters.

[0047] Figure 2 A schematic overview of the method according to the application is shown, which comprises the following method steps:

[0048] A) creating a cutting edge with a plurality of cutting parameters;

[0049] B) creating a record of the cutting edge;

[0050] C) reading the record;

[0051] D) analyzing the record by means of a neural network in order to determine the cutting parameters;

[0052] E) backpropagating the neural network for determining a relevance of recorded pixels with respect to the determined cutting parameters; and

[0053] F) an indication of relevant and / or irrelevant recorded pixels.

[0054] Figure 3 A recording of various cutting edges is shown.

[0055] Figure 4 The functioning of the neural network, or more specifically the backpropagation, is schematically shown.

[0056] Figure 5 In the left column a recording of two cutting edges is shown, and in the further columns the determined cutting parameters and the recorded output is shown, wherein the recorded pixels relevant to the determined cutting parameters are highlighted. DETAILED DESCRIPTION

[0057] Figure 1 A portion of a machine tool 10 in the form of a laser cutting machine is shown. Here, a cutting head 12 passes above a workpiece 14, wherein the workpiece 14 is subjected to laser- and gas radiation. Here, a cutting edge 16 is produced. The cutting edge 16 is influenced, inter alia, by the following cutting parameters 18: gas pressure 20, feed 22, nozzle-workpiece distance 24, nozzle-focus distance 26 and / or focus position 28.

[0058] The influence of the individual cutting parameters 18 on the appearance of the obtained cutting edge 16 is largely unclear even to experts. If, for example, stripes appear on the cutting edge 16, the cutting parameters 18 must be changed until the stripes disappear, wherein, with such a change, on the one hand, material consumption, energy consumption and time consumption are incurred, and on the other hand, such a change often results in a new artefact. There is therefore a need to provide a method and a device by which the cutting parameters 18 are assigned to features of the cutting edge 16 in a targeted manner. These cutting parameters 18 can then be changed in order to change the features of the cutting edge 16. The present invention therefore solves the problem that a human user cannot solve the problem due to the complexity of the problem ("superhuman performance").

[0059] Figure 2An overview of the method according to the application is shown. In method step A), a cutting edge 16 is produced by a machine tool 10 using a cutting parameter 18. In method step B), the cutting edge 16 (see method step A)) is recorded using a camera 30. The camera 30 can be configured in the form of a photo camera and / or a video camera. In method step C), the created recording 32 is read. In method step D), the recording 32 is analyzed by an algorithm 34. The algorithm 34 has a neural network 36. The neural network 36 is used to derive 38 the cutting parameter 18. The derived cutting parameter 18 can be compared with the set cutting parameter 18 (see method step A)) for example in order to derive a defect of the machine tool 10 (see method step A)).

[0060] In method step E), the algorithm 34 implements a backpropagation 40 in the neural network 36. When the cutting parameter 18 is derived in method step D), the relevance of the individual recording pixels 42a, 42b of the recording 40 is established by the backpropagation 40 of the cutting parameter 18 with respect to the recording 32. In method step F), the recording pixels 42a, b (for the sake of clarity, only the recording pixels 42a, b are provided with reference numerals in the Figure 2 current case, a first color (for example red) is used to identify particularly relevant recording pixels 42a and a second color (for example blue or gray) is used to identify particularly irrelevant recording pixels 42b. Due to formal provisions, different colors are represented by different patterns (hatching) in the present description. On the basis of the particularly relevant recording pixels 42a, the user can directly identify which regions of the recorded cutting edge 16 (see method step A)) are particularly influenced by the respectively derived cutting parameter 18 (see method step D)).

[0061] Figure 3 Three recordings 32a, 32b, 32c are exemplary shown, wherein these recordings 32a-c are created with different cutting parameters 18 (see Figure 1 ):

[0062] Recording 32a:

[0063] In contrast, the recording 32b is created with an increased nozzle-focus distance 26. In comparison to the recording 32a, the recording 32c is created with a reduced feed 22. From Figure 3 it is evident that a human user cannot directly derive the influence of the cutting parameter 18 (see Figure 1 ) from the recordings 32a-c.

[0064] Figure 4The algorithm 34 or the neural network 36 is illustrated schematically. The neural network 36 is configured in the form of a convolutional neural network having a plurality of blocks 44a, 44b, 44c, 44d, 44e. Here, an input block 44a is provided. The blocks 44b-e each have three convolutional layers 46a, 46b, 46c, 46d, 46e, 46f, 46g, 46h, 46i, 46j, 46k, 46l. The blocks 44a-e have filters 48a, 48b, 48c, 48d, 48e. Each layer of the input block 44a has 32 filters 48a. The layers of the block 44b likewise have 32 filters 48b. The layers of the block 44c have 64 filters 48c. The layers of the block 44d have 128 filters 48d, and the layers of the block 44e have 256 filters 48e. The filters 48a-e can result in a reduction of the resolution of the recording 32 (for example from 200 pixels x 200 pixels to 7 pixels x 7 pixels) while the depth (or number of channels) increases. The filters 48a-e of the third layer of each block 44a-e result in a reduction of the resolution. Here, the convolutional layers are also used for pooling. The depth increases from one block 44a-e to the next. By way of example, the block 44b consists of three convolutional layers, each having 32 filters. In the first two convolutional layers, the spatial resolution is 112 x 112 pixels. From the second convolutional layer to the third convolutional layer, the spatial resolution is reduced from 112 x 112 pixels to 56 x 56 pixels. When transitioning from the block 44b (last layer) to the block 44c (first layer), the depth increases from 32 to 64. The spatial resolution remains unchanged.

[0065] The neural network 36 thus enables the cut parameters 18 to be determined 38. In the present case, the layer-wise relevance propagation is used in the backpropagation 40. The result is shown in Figure 5 .

[0066] Figure 5 The recording 32a is shown in the upper column and the recording 32b is shown in the lower column. In each column, the recording 32a, b is reproduced a plurality of times, the recording pixels 42a, b that are greatly or little influenced by the feed 22 are highlighted in the second column, the recording pixels 42a, b that are greatly or little influenced by the focus position 28 are highlighted in the third column, and the recording pixels 42a, b that are greatly or little influenced by the gas pressure 20 are highlighted in the fourth column. Here, the output 50 can exist in the form of a heat map.

[0067] The recording pixels 42b that are particularly little influenced by the respective cut parameter 18 (see Figure 1 ) are primarily used to check the plausibility of the output 50. Preferably, in the output 50, only the recording pixels 42a that are particularly greatly influenced by the respective cut parameter 18 (see Figure 1 ) are highlighted in order to facilitate the user in processing the output 50.

[0068] In combination with all of the figures in the drawings, the invention relates generally to a method for identifying cutting parameters 18 that are of particular importance for a specific feature of a cutting edge 16. Here, recordings 32, 32a-c of the cutting edge 16 are analyzed by an algorithm 34 having a neural network 36 for deriving 38 the cutting parameters 18. Through backpropagation 40 of the analysis, those recording pixels 42a,b that play an important role for determining the cutting parameters 18 are identified. An output 50 in the form of a representation of these important recording pixels 42a,b, in particular in the form of a heat map, reveals to a user of the method which cutting parameters 18 need to be changed to improve the cutting edge 16. The invention also relates to a computer program product and a device for performing the method, respectively.

[0069] List of reference signs

[0070] 10 machine tool

[0071] 12 cutting head

[0072] 14 workpiece

[0073] 16 cutting edge

[0074] 18 cutting parameter

[0075] 20 gas pressure

[0076] 22 feed

[0077] 24 nozzle-workpiece distance

[0078] 26 nozzle-focus distance

[0079] 28 focus position

[0080] 30 camera

[0081] 32, 32a-c recording

[0082] 34 algorithm

[0083] 36 neural network

[0084] 38 deriving the cutting parameters 18

[0085] 40 backpropagation

[0086] 42a,b recording pixel

[0087] 44a-e block of the neural network 36

[0088] 46a-l layer of the neural network 36

[0089] 48a-e filter of the neural network 36

[0090] 50 output

Claims

1. A method for analyzing a cut edge (16) created by a machine tool (10), the method comprising the following steps: A) The cutting edge (16) is created by the machine tool (10) using at least one cutting parameter (18); B) Create at least one record (32, 32a-c) of the cut edge (16) using a camera (30), C) Read at least one record (32, 32a-c) of the cut edge (16), the record (32, 32a-c) having a plurality of record pixels (42a, b); D) Analyze the record (32, 32a-c) using a trained neural network (36) to obtain the at least one cutting parameter (18); E) Backpropagation (40) of the neural network (36) to obtain the correlation of these recorded pixels (42a, b) analyzed in order to determine the cut parameters (18) to be obtained; F) If the particularly relevant and / or particularly irrelevant record pixels (42a, b) identified in method step E) are identified, output records (32, 32a-c). in, The machine tool (10) is constructed in the form of a laser cutting machine. In step D), the following cutting parameters (18) are obtained: material parameters, which include gas purity and melting temperature of workpiece (14).

2. The method as described in claim 1, wherein, The analysis in step D) is performed using a convolutional neural network with multiple layers (46a-l).

3. The method as described in claim 1 or 2, wherein, Backpropagation (40) in method step E) is achieved through layer-by-layer correlation propagation.

4. The method as described in claim 1 or 2, wherein, The assignment of correlations in step E) is based on deep Taylor decomposition.

5. The method as described in claim 1 or 2, wherein, The output of step F) is implemented in the form of a heatmap.

6. The method as described in claim 1 or 2, wherein, The records (32, 32a-c) in step C) exist in the form of RGB photographs or 3D point clouds.

7. The method as described in claim 1 or 2, wherein, The camera (30) is the camera (30) of the machine tool (10).

8. The method as claimed in claim 1 or 2, wherein, In step D), the following cutting parameters (18) are obtained: • Beam parameters; • Transmission parameters; and / or • Gas dynamic parameters.

9. The method of claim 2, wherein, Each layer (46a-l) of the convolutional neural network has multiple filters (48a-e).

10. The method of claim 8, wherein, The beam parameters include the focal diameter and / or laser power.

11. The method of claim 8, wherein, The transmission parameters include focal position (28), nozzle-focal distance (26), and / or feed (22).

12. The method of claim 8, wherein, The gas dynamic parameters include gas pressure (20) and / or nozzle-workpiece distance (24).

13. A computer program product for performing method steps C) to F) as described in any one of claims 1 to 12, wherein, The computer program product includes the neural network (36).

14. An apparatus comprising: The device includes a machine tool (10), a computer, and a computer program product as described in claim 13 for performing the method as described in any one of claims 1 to 12, wherein the device includes the camera (30), and wherein the machine tool (10) is configured as a laser cutter.

Images