METHOD, DEVICE, AND COMPUTER PROGRAM PRODUCT FOR DEMONSTRATING THE INFLUENCE OF CUTTING PARAMETERS ON A CUTTING EDGE

Description

Background of the invention

[0001] The invention relates to a method for analyzing a cut edge produced by a machine tool according to the preamble of claim 1 (see, e.g., J. STAHL ET AL: "Quick roughness evaluation of cut edges using a convolutional neural network", SPIE PROCEEDINGS, SPIE, US, Vol. 11172, July 16, 2019, pages 111720P-111720P). The invention further relates to an apparatus for carrying out the method and a computer program product, see claims 9 and 10.

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

[0003] However, most cutting processes and the influence of cutting parameters on the cut edges are not fully understood. This can be seen, for example, in the following article: Hügel, H., Graf, T. Laser in manufacturing: beam sources, systems, manufacturing processes. Wiesbaden: Vieweg + Teubner, 2009.

[0004] Petring, D., Schneider, F., Wolf, N. Some answers to frequently asked questions and open issues of laser beam cutting. In: International Congress on Applications of Lasers & Electro-Optics. ICALEOR 2012, Anaheim, California, USA: Laser Institute of America, 2012, pp. 43-48

[0005] Steen, WM, Mazumder, J. Laser Material Processing. London: Springer London, 2010.

[0006] Even experienced users of cutting systems are generally unable to predict how the cutting parameters will affect the appearance of the cut edge. Therefore, to improve the appearance of a cut edge, especially when problems arise due to changes in material quality and / or new processes using a new laser source, sheet thickness, etc., extensive test series with various varying cutting parameters must be carried out regularly. Object of the invention

[0007] In contrast, the object of the invention is to provide a method, a computer program product and a device that are capable of analyzing the influence of the cutting parameters on the cutting edge. Description of the invention

[0008] This problem is solved according to the invention by a method according to claim 1, a computer program product according to claim 10, and a device according to claim 9. The dependent claims describe preferred embodiments.

[0009] The problem according to the invention is thus solved by a method in which an image of a cutting edge produced by a machine tool, containing several recording pixels, is read in. An algorithm with a neural network is used to determine at least one cutting parameter, and in particular several cutting parameters, from the image. Subsequently, backpropagation is performed in the neural network to determine the relevance of the individual recording pixels for determining the previously determined cutting parameters. The image is then output with at least some recording pixels marked, the marking reflecting the previously determined relevance of the recording pixels. Preferably, all recording pixels are output and marked according to their relevance.

[0010] A user can therefore immediately see from the highlighted output how strongly the respective areas of the cut edge were influenced by the respective cutting parameter(s) and can then adjust this cutting parameter(s) to specifically change a certain area of ​​the cut edge.

[0011] The backpropagation of a neural network has been described, for example, in EP 3 654 248 A1.

[0012] Typically, backpropagation ("backpropagation-based mechanisms") is only used to verify whether a neural network has learned the correct relationships. This applies to neural networks that do not exhibit "superhuman performance." In such cases, humans can easily assess what constitutes the correct information. For example, it can be used to investigate whether a neural network that can distinguish dogs from cats actually recognizes the dog in an image when it indicates that it is a dog and not the meadow in which the dog is standing. For instance, the neural network might recognize a text in the image that appears in all horse images instead of a specific animal (e.g., a horse) (the so-called "Clever Hans problem"). In contrast, backpropagation is used here to understand or at least predict production processes or physical relationships during the cutting process.

[0013] A neural network is understood to be an architecture that includes at least one, and preferably several, data aggregation routines. A data aggregation routine can be designed to aggregate multiple "retrieved data" items into a new data package. This new data package can contain one or more numbers or vectors. Further data aggregation routines can provide the new data package, either fully or partially, with additional "retrieved data." "Retrieved data" can include, in particular, cutting parameters or data packages provided by one of the data aggregation routines. An architecture with multiple interconnected data aggregation routines is particularly preferred. Specifically, several hundred, and especially several thousand, such data aggregation routines can be interconnected. This significantly improves the quality of the neural network.

[0014] The architecture can include a function with weighted variables. One, more specifically several, and preferably all, data aggregation routines can be designed to combine multiple "determined data" values ​​with a weighted variable, specifically to multiply them, thus transforming the "determined data" into "combined data" and then aggregating the "combined data" into a new data package, specifically by adding it. In the neural network, data can be multiplied by weights. The information from multiple neurons can be added. Furthermore, the neural network can have a nonlinear activation function.

[0015] The cutting edge features contained in the recording can themselves be data packages, in particular several structured data, especially data vectors or data arrays, which themselves can represent "determined data", especially for the data aggregation routines.

[0016] To determine suitable weighted variables, i.e., to train the neural network, the procedure can be carried out with data, in particular cutting parameters, whose association with recordings is known.

[0017] The neural network is preferably designed as a multi-layered convolutional neural network (CNN). The convolutional neural network can have convolutional layers and pooling layers. Pooling layers are typically positioned between two successive convolutional layers. Alternatively, or additionally, pooling can be performed after each convolution.

[0018] In addition to convolutional and pooling layers, a CNN can have a fully connected layer, particularly at the very end of the neural network. The convolutional and pooling layers extract features, while the fully connected layer can assign the features to the cutting parameters.

[0019] The neural network can have multiple filters per layer. The structure of a convolutional neural network can be found, for example, in the following articles, especially the first one listed below: LeCun Y, Bengio Y, Hinton G (2015) Deep learning; Nature 521:436{444, DOI 10.1038 / nature14539; 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; 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; 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; Gonçalves 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; Karatas A, Kölsch D, Schmidt S, Eier M, Seewig J (2019) Development of a convolutional autoencoder using deep neural networks for defect detection and generating ideal references for cutting edges; Munich, Germany, DOI 10.1117 / 12.2525882; Stahl J, Jauch C (2019) Quick roughness evaluation of cut edges using a convolutional neural network; In: Proceedings SPIE 11172, Munich, Germany, DOI 10.1117 / 12.2519440. .

[0020] For backpropagation, layer-wise relevance propagation (LRP) has proven to be particularly effective and easy to implement. Further information on layer-wise relevance propagation can be found in the following articles: Bach S, Binder A, Montavon G, Klauschen F, Müller 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; W Samek, A Binder, G Montavon, S Lapuschkin, K Müller (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; Montavon G, Lapuschkin S, Binder A, Samek W, Müller KR (2017) Explaining Non-Linear Classification Decisions with Deep Taylor Decomposition; Pattern Recognition 65:211{222, DOI 10.1016 / j. patcog.2016.11.008; Montavon G, Lapuschkin S, Binder A, Samek W, Müller KR (2015) Explaining Non-Linear Classification Decisions with Deep Taylor Decomposition; arXiv preprint URL https: / / arxiv.org / pdf / 1512.02479.pdf; Montavon G, Binder A, Lapuschkin S, Samek W, Müller KR (2019) Layer-Wise Relevance Propagation: An Overview.In: Samek W, Montavon G, Vedaldi A, Hansen LK, Müller KR (eds) Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, Springer, Cham, Switzerland, pp 193{209. .

[0021] Backpropagation, particularly in the form of layer-wise relevance propagation, is preferably based on deep Taylor decomposition (DTD). Deep Taylor decomposition is described in more detail in the following article: Montavon G, Lapuschkin S, Binder A, Samek W, Müller KR (2017) Explaining Non-Linear Classification Decisions with Deep Taylor Decomposition; Pattern Recognition 65:211, 222, DOI 10.1016 / j.patcog.2016.11.008.

[0022] The implementation can be done, for example, in Python using the libraries 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 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.

[0023] Preferably, the output is presented as a heatmap. The heatmap can use two colors, particularly red and blue, to indicate particularly relevant and particularly irrelevant pixels, respectively. Pixels of average relevance can be indicated by intermediate shades between the two colors or by gray. This makes the output particularly intuitive to understand.

[0024] The recording is preferably a photograph, especially preferably a color photograph, particularly in the form of an RGB photograph, or a 3D point cloud.

[0025] 3D point clouds are somewhat more complex to create because they contain depth information. This depth information can be obtained during image acquisition, particularly through light sectioning or triangulation from different angles. However, it has been shown that color photographs are especially suitable and sufficient, as the neural network recognizes the different sectioning parameters primarily by the varying colors of the section edges.

[0026] The recording is made with a photo and / or video camera. According to the invention, the camera is part of the machine tool to ensure a consistent recording situation. Alternatively or additionally, the camera can be part of a photo booth to reduce environmental influences during recording.

[0027] According to the invention, the method comprises creating the cut edge using a machine tool. The cutting process of the machine tool can be a thermal cutting process, in particular a plasma cutting process, preferably a laser cutting process. In the case of a laser cutting process, the determined cutting parameters preferably include beam parameters, in particular focus diameter or laser power; transport parameters, in particular focus position, nozzle focus distance or feed rate; gas dynamic parameters, in particular gas pressure or nozzle-workpiece distance; and / or material parameters, in particular gas purity or melting temperature of the workpiece. These cutting parameters have proven to be particularly influential for the appearance of the cut edge. The object of the invention is further achieved by a computer program for performing the calculation operation described herein, see claim 10.The computer program product can be partially, or even entirely, cloud-based to allow access to multiple users. Furthermore, more comprehensive training of the neural network can be performed by multiple users.

[0028] The problem according to the invention is finally solved by a device for carrying out the method described herein, wherein the device comprises the machine tool, in particular in the form of a laser cutting machine, see claim 9.

[0029] Further advantages of the invention will become apparent from the description and the drawing. The embodiments shown and described are not to be understood as an exhaustive list, but rather as examples illustrating the invention as defined in the claims. Detailed description of the invention and drawing

[0030] Fig. 1 shows a schematic representation of a machine tool in the form of a laser cutting machine to illustrate essential cutting parameters. Fig. 2 shows a schematic overview of the method according to the invention with the process steps: A) Creating the cutting edge with several cutting parameters; B) Creating an image of the cutting edge; C) Reading in the image; D) Analyzing the image with a neural network to determine the cutting parameters; E) Backpropagation of the neural network to determine the relevance of the image pixels with respect to the determined cutting parameters; and F) Labeled representation of the relevant and irrelevant image pixels. Fig. 3 shows images of different cutting edges. Fig. 4 schematically shows the operation of the neural network and the backpropagation.Figure 5 shows images of two cutting edges in the left column and the determined cutting parameters and outputs of the images in the subsequent columns, highlighting the image pixels relevant to the determined cutting parameters.

[0031] Fig. 1 shows part of a machine tool 10 in the form of a laser cutting machine. A cutting head. 12 It passes over a workpiece 14 under laser and gas irradiation of the workpiece 14. Here, a cutting edge is created. 16 The cutting edge 16 is generated in particular by the following cutting parameters. 18 Influenced: Gas pressure 20, feed 22, Nozzle workpiece distance 24, Nozzle focus distance 26 and / or focus position 28.

[0032] The influence of the individual cutting parameters 18 on the appearance of the resulting cut edge 16 is largely incomprehensible, even to experts. For example, if scoring occurs on the cut edge 16, the cutting parameters 18 must be varied until the scoring disappears. This variation involves significant material, energy, and time consumption, and often results in the creation of new artifacts. Therefore, there is a need to provide a method and a device that selectively assigns cutting parameters 18 to the characteristics of a cut edge 16. These cutting parameters 18 can then be changed to modify the characteristics of the cut edge 16. The invention thus solves a problem that cannot be solved by human users due to its complexity ("superhuman performance").

[0033] Fig. 2Figure 1 shows an overview of the method according to the invention. In process step A), the cutting edge 16 is generated by the machine tool 10 using the cutting parameters 18. In process step B), the cutting edge 16 (see process step A)) is measured using a camera. 30 The camera 30 can be designed as a still camera and / or a video camera. In process step C), the recording is taken. The camera 30 can be designed as a still camera and / or a video camera. 32 read in. In process step D), recording 32 is processed by an algorithm. 34 analyzed. Algorithm 34 features a neural network. 36 The neural network 36 serves to determine 38 the cutting parameters 18. The determined cutting parameters 18 can be compared with the set cutting parameters 18 (see process step A)), for example to determine a defect in the machine tool 10 (see process step A)).

[0034] In process step E), backpropagation is carried out using algorithm 34. 40 in the neural network 36. Through the backpropagation 40 of the cutting parameters 18 to the recording 32, the relevance of individual recording pixels is determined. 42a, 42b the recording 40 during the determination of the cutting parameters 18 in process step D). In process step F), the recording pixels 42a, b are displayed (where in Fig. 2For clarity, only the recording pixels 42a and b are marked with a reference symbol and their respective relevance. In this case, the particularly relevant recording pixel 42a is marked in a first color (e.g., red), and the particularly irrelevant recording pixel 42b in a second color (e.g., blue or gray). Due to formatting requirements, the different colors are represented by different patterns (hatching) in this description. Based on the particularly relevant recording pixel 42a, a user can immediately see which areas of the recorded cutting edge 16 (see process step A)) are particularly affected by the respective determined cutting parameter 18 (see process step D)).

[0035] Fig. 3 shows three exemplary images 32a, 32b, 32c, where the recordings 32a-c with different cutting parameters 18 (see Fig. 1 ) were created: Recording 32a: Gas pressure 20 15 bar Feed rate 22 21m / min Nozzle workpiece distance 24 1.5mm Nozzle focus distance 26 -2mm

[0036] Image 32b, in contrast, was taken with an increased nozzle focus distance 26. Image 32c was taken with a reduced feed rate 22 compared to image 32a. From Fig. 3 It is evident that the influence of the cutting parameters 18 (see Fig. 1 ) is not immediately apparent to human users from the recordings 32a-c.

[0037] Fig. 4 Figure 34 schematically shows algorithm 34 or neural network 36. Neural network 36 is in the form of a convolutional neural network with multiple blocks. 44a, 44b, 44c, 44d, 44e The structure includes an input block 44a. Blocks 44b-e each have three convolutional layers. 46a, 46b, 46c, 46d, 46e, 46f, 46g, 46h, 46i, 46j, 46k, 461 Blocks 44a-e have filters. 48a, 48b, 48c, 48d, 48eEach layer of input block 44a has 32 filters 48a. The layers of block 44b also have 32 filters 48b. The layers of block 44c have 64 filters 48c. The layers of block 44d have 128 filters 48d, and the layers of block 44e have 256 filters 48e. The filters 48a-e can reduce the resolution of the image 32 (e.g., from 200 pixels x 200 pixels to 7 pixels x 7 pixels) while simultaneously increasing the depth (or number of channels). The filters 48a-e of the third layer of each block 44a-e reduce the resolution. Here, convolutional layers are used for pooling. The depth increases from one block 44a-e to the next. For example, block 44b consists of three convolutional layers, each with 32 filters. In the first two images, the spatial resolution is 112 x 112 pixels. From the second to the third, the spatial resolution decreases from 112 x 112 pixels to 56 x 56 pixels.At the transition from block 44b (last layer) to block 44c (first layer), the depth increases from 32 to 64. The spatial resolution remains the same.

[0038] The neural network 36 thus enables the determination 38 of the cutting parameters 18. For backpropagation 40, layer-wise relevance propagation is used. Its results are presented in Fig. 5 depicted.

[0039] Fig. 5 The upper column shows image 32a and the lower column image 32b. Images 32a and b are reproduced multiple times in each column, with the second column highlighting the image pixels 42a and b strongly or slightly influenced by the feed rate 22, the third column highlighting the image pixels 42a and b strongly or slightly influenced by the focus position 28, and the fourth column highlighting the image pixels 42a and b strongly or slightly influenced by the gas pressure 20. The outputs 50 They may be in the form of heatmaps.

[0040] The respective cutting parameter 18 (see Fig. 1 The least affected recording pixels 42b primarily serve to check the plausibility of the outputs 50. Preferably, only those pixels particularly affected by the respective cutting parameter 18 (see Fig. 1 ) influenced recording pixel 42a highlighted to make it easier for a user to handle the outputs 50.

[0041] In summary, the invention relates to a method for identifying cutting parameters 18 that are particularly important for certain features of a cutting edge 16, taking into account all figures of the drawing. An image 32, 32a-c of the cutting edge 16 is analyzed by an algorithm 34 with a neural network 36 to determine 38 the cutting parameters 18. By backpropagation 40 of this analysis, those image pixels 42a, b that play a significant role in determining the cutting parameters 18 are identified. An output 50 in the form of a representation of these significant image pixels 42a, b, in particular in the form of a heatmap, shows a user of the method which cutting parameters 18 must be changed to improve the cutting edge 16. The invention further relates to a computer program product or a device for carrying out the method. Reference symbol list

[0042] 10 Machine tool 12 Cutting head 14 Workpiece 16 Cutting edge 18 Cutting parameters 20 Gas pressure 22 Feed 24 Nozzle-workpiece distance 26 Nozzle-focus distance 28 Focus position 30 Camera 32, 32a-c Capture 34 Algorithm 36 Neural network 38 Determination of cutting parameters 18 40 Backpropagation 42a, b Capture pixels 44a-e Blocks of the neural network 36 46a Layers of the neural network 36 48a-e Filters of the neural network 36 50 Output

Claims

1. A method of analyzing a cut edge (16) created by a machine tool (10), comprising the method steps: A) creating the cut edge (16) using at least one cutting parameter (18) with the machine tool (10); B) creating at least one image (32, 32a-c) of the cut edge (16) with a camera (30); C) scanning the image (32, 32a-c) of the cut edge (16), wherein the image (32, 32a-c) comprises a plurality of image pixels (42a, b); characterized by the method steps: D) analyzing the image (32, 32a-c) with a trained neuronal network (36) to determine the at least one cutting parameter (18); E) backpropagation (40) of the neuronal network (36) to determine the relevance of the image pixels (42a, b) analyzed to determine the determined cutting parameters (18); F) output of the image (32, 32a-c) with identification of the particularly relevant and / or particularly irrelevant image pixels (42a, b) determined in method step E); wherein steps C) to F) are performed by a computer.

2. The method according to claim 1, in which the analysis in method step D) is carried out using a convolutional neural network with a plurality of layers (46a-l), in particular with a plurality of filters (48a-e) per layer (46a-l).

3. The method according to claim 1 or 2, wherein the backpropagation (40) in method step E) is performed by layer-wise relevance propagation.

4. The method according to any one of the preceding claims, wherein the assignment of relevance in method step E) is based on deep Taylor decomposition.

5. The method according to any one of the preceding claims, wherein the output in method step F) is in the form of a heatmap.

6. The method according to any one of the preceding claims, wherein the image (32, 32a-c) in method step C) is present in the form of an RGB photo or a 3D point cloud.

7. The method according to any one of the preceding claims, wherein the camera (30) is a camera (30) of the machine tool (10).

8. The method according to one of the preceding claims, wherein the machine tool (10) is designed in the form of a laser cutting machine, wherein the following cutting parameters (18) in particular are determined in method step D): • beam parameters, in particular focal diameter and / or laser power; • transport parameters, in particular focus position (28), nozzle focus distance (26) and / or feed rate (22); • gas dynamics parameters, in particular gas pressure (20) and / or nozzle-workpiece distance (24); and / or • material parameters, in particular gas purity and / or melting temperature of the workpiece (14).

9. An apparatus with a machine tool (10), in particular in the form of a laser cutting machine, a computer and a camera (30), wherein the computer is programmed with a neuronal network so as to perform the steps of the method according to any one of the preceding claims.

10. A computer program product comprising instructions that cause the apparatus of claim 9 to perform the method steps according to any one of claims 1 to 8, wherein the computer program product comprises the neuronal network (36).