Method for training a neural network to identify tool status and related methods and devices

Abstract

Method for training a neural network to recognize tool states and related methods and devices. In the method for training a neural network to recognize tool states based on image data, the neural network is trained to recognize tool states of a first tool type and image data of a second tool type are used, which are subjected to image processing, and the image data of the second tool type are converted into image data of the first tool type by means of the image processing, wherein the neural network is trained based on the converted image data. In the method for processing and / or manufacturing with the aid of the first tool type, the tool state of the first tool type is recognized by means of a neural network, and the neural network is trained according to this method. The device for processing and / or manufacturing with the aid of the first tool type includes a neural network, which is trained according to this method for training a neural network and / or the neural network is configured to implement this method for processing and / or manufacturing.

Description

Technical Field

[0001] The invention relates to a method for training a neural network for identifying a tool state based on image data, a method for machining and / or manufacturing, and a device. Background Art

[0002] In cutting manufacturing, a high correlation is due to the tools used and their wear: on the one hand, tool wear is a central cost factor in cutting manufacturing, on the other hand, important information about the machining process itself can be obtained by analyzing tool wear phenomena. In this way, machining parameters can be derived from tool wear in order to machine the workpiece more accurately or with less tool wear. For this purpose, it is necessary to accurately detect tool wear.

[0003] Optical methods are known for detecting the degree of wear of tools, which can differentiate between different types of wear, in particular surface wear or built-up edge or notch formation. On the other hand, it is known that not only the presence of wear-induced defects of the tool can be ascertained, but also more information about the respective defect, such as width and shape and area, can be determined.

[0004] Furthermore, it is known to use artificial intelligence methods to identify the type of wear.

[0005] However, in the artificial intelligence approach, neural networks need to be specially trained to identify defects.

[0006] Therefore, such methods could not be used previously for new tool types, since there was usually no training data at all for this purpose and the learning of new tool types therefore meant high labor costs and high barriers to entry for using new tool types.

[0007] This situation is particularly relevant for industrial applications, since there are a particularly wide variety of tool types for cutting operations. Tools, in particular cutting tools, differ not only in their geometry but also in the coatings used. Here, adaptation of the neural network is always necessary and, due to the learning of the neural network, is often complex and expensive. Summary of the invention

[0008] Against this background of the prior art, the object of the present invention is therefore to provide an improved method for training a neural network for identifying tool states. In addition, the object of the present invention is to specify an improved method and an improved device for machining and / or producing workpieces.

[0009] This object of the invention is achieved with a method for training a neural network for identifying tool states based on image data having the features specified in claim 1 and with a method for processing and / or manufacturing having the features specified in claim 9 and with a device having the features specified in claim 10. Preferred developments of the invention are described in the associated dependent claims, the subsequent description and the drawings.

[0010] In the method according to the invention for training a neural network for identifying tool states from image data, the neural network is trained to identify tool states of a first tool type and image data of a second tool type are used, the image data being subjected to image processing by means of which the image data of the second tool type are converted into image data of the first tool type. In the method according to the invention, the neural network is trained from the converted image data.

[0011] Advantageously, with the method according to the invention, image data of the second tool type can be used and converted into image data of the first tool type. Thus, in this way, image data matching the first tool type can be calculated. Thus, manual acquisition of image data of the first tool type can be dispensed with. Thus, according to the invention, the otherwise required outlay for generating image data of the first tool type is reduced to merely converting image data of the second tool type into image data of the first tool type. Based on the converted image data, a neural network can thus be trained in the same way as with actually detected image data of the first tool type.

[0012] In particular, in known solutions, the actually detected image data of the first tool type must first be classified by a domain expert in order to be able to assign tool states, such as in particular tool wear, to the actually detected image data. A manual evaluation must therefore be performed in the case of the actually detected image data. In the method according to the invention, these separate evaluations are dispensed with. Instead, the conversion of the image data of the second tool type into those of the first tool type is preferably performed automatically, for example with the aid of image processing software.

[0013] It is self-evident that the first tool type and the second tool type within the meaning of the present invention are expediently different from one another.

[0014] Preferably, in the method according to the invention, the image data are converted such that the shape of the second tool type is transformed into the shape of the first tool type.

[0015] Thus, tool types often differ precisely in shape, for example in the case of tools for subtractive machining. In particular, cutting edges and / or milling cutters and / or drills often differ in shape.

[0016] Advantageously, in the method according to the present invention, the shape of the second tool type in the image data is transformed into the shape of the first tool type, so that the image data of the second tool type is divided into small image areas, and the small image areas are changed in their relative orientation and / or position and / or size within the image data of the second tool type, so that the changed image data reproduces the shape of the first tool type.

[0017] Alternatively or additionally and likewise preferably, in the method the image data are converted such that the colors of the second tool type are transferred to the colors of the first tool type.

[0018] Typically, tool types often differ in their tool materials and / or their coatings, in particular in the areas of the tool types used for processing. Different coatings and / or different materials of the second and first tool types can be realistically simulated by adapting the colors of the tool types. Thus, by adapting the colors of the second tool type, image data of the second tool type can be effectively converted into those of the first tool type.

[0019] Advantageously, in the method according to the invention, the first tool type and / or the second tool type is designed for subtractive machining of a workpiece.

[0020] Especially in the case of tool types designed for subtractive machining, the tool condition is a particularly cost-relevant variable. Thus, not only the timely replacement of the tool but also the longest possible service life of the tool has a significant impact on the machining costs of the workpiece. Therefore, a precise assessment of the tool condition of the tool type used is of great significance for economic efficiency during machining and / or production.

[0021] In the method according to the invention, the first and / or second tool type preferably forms a cutting and / or milling tool and / or a drilling tool. Suitably, the first and / or second tool type is / comprises a cutting blade and / or a milling cutter and / or a drill.

[0022] In the method according to the invention, the image data preferably originate from optical imaging. In the method according to the invention, the image data are suitably acquired using a camera, in particular a micro camera. The image data are expediently acquired digitally, so that the image data can be converted digitally, preferably by means of software.

[0023] In the method according to the invention, the tool state is preferably a wear state. In subtractive machining, the wear state of the tool has a high influence on the costs of machining the workpiece. For example, a premature replacement of the tool during machining results in high tool costs. A too late replacement of the tool can make the machining of the workpiece difficult and increase machining waste and the costs caused thereby. Therefore, the use of the method according to the invention for evaluating the wear state improves the cost efficiency during machining and / or manufacturing.

[0024] In the method according to the invention, the image data are preferably converted by changing color channels and / or compressing and / or expanding the image data and / or dividing and changing the image data, in particular changing and combining the image data with respect to position and / or size and / or orientation.

[0025] The method according to the invention for machining and / or producing is carried out with a first tool type, wherein a tool state of the first tool type is identified with the aid of a neural network which is trained according to the method described above.

[0026] Particularly preferably, the method for machining and / or manufacturing according to the invention is performed with the aid of a first tool type, wherein the training of the neural network is performed with the aid of the method according to the invention. That is, ideally, the training of the neural network forms part of the method for machining and / or manufacturing according to the invention.

[0027] Expediently, machining and / or production is adapted and / or interrupted as a function of the tool state.

[0028] The device according to the invention for machining and / or manufacturing with the aid of a first tool type comprises a neural network, which is trained in the manner explained according to the method according to the invention for training a neural network as described above, and / or the device is designed to implement the method according to the invention for machining and / or manufacturing as described above. Preferably, the device has a machining device with a first tool type, which machining device is preferably controlled as a function of the wear state and / or the device is designed to interrupt manufacturing and / or machining with the machining device in order to be able to replace the first tool type. Preferably, the device has a training device for training the neural network with the aid of the method according to the invention as described above. Preferably, the device comprises an image processing device and / or image processing software, which is set up and configured to convert image data of the second tool type into image data of the first tool type.

[0029] The device according to the invention advantageously comprises an image capture unit for capturing images of the first tool type and an evaluation unit having a neural network, which is designed and configured to detect the degree of wear of the first tool type. The image capture unit is advantageously a camera, preferably a micro camera. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The invention is explained in more detail below with reference to the exemplary embodiments shown in the drawings.

[0031] Figure 1 A device for carrying out the method according to the invention for machining a workpiece is shown schematically in a schematic diagram.

[0032] Figure 2 The method according to the invention for Figure 1 A flowchart of an embodiment of a method for processing a workpiece by an apparatus, and

[0033] Figure 3 The schematic diagram shows the implementation according to Figure 2 The method is a transformation of image data of the second cutting edge type into image data of the first cutting edge type. DETAILED DESCRIPTION

[0034] exist Figure 1 The device 10 according to the invention shown in FIG. 1 has a processing chamber 20 in which a workpiece 30 is processed according to the method BEAR according to the invention for processing a workpiece 30. The workpiece 30 is processed by milling the workpiece with the aid of a milling tool 40. The milling tool 40 has a cutting edge 50 which is worn when the workpiece 30 is milled.

[0035] Furthermore, a camera 60 is arranged in the process chamber 20 of the device 10, which records images BNS of the cutting edge 50 of the milling tool 40 between the machining phases and evaluates them with the aid of an evaluation device AUSW. In the exemplary embodiment shown, the camera 60 is a micro camera, which detects a high-resolution image BNS of the cutting edge 50 in an image acquisition step BEE. In this case, the camera 60 primarily detects the areas of the cutting edge 50 that are typically subject to wear. In the exemplary embodiment shown, these are the main rear face and the cutting face of the cutting edge 50.

[0036] Firstly, the evaluation device AUSW determines from the image BNS of the cutting edge 50 whether the cutting edge 50 can continue to be used or possibly must be replaced. Furthermore, the evaluation device AUSW determines from the image BNS whether and what kind of adaptation of the processing parameters of the device 20 is required if the workpiece 30 is further processed with the cutting edge 50.

[0037] For evaluating the image BNS of the cutting edge 50 , the evaluation device AUSW comprises a neural network NN which is trained in a first step by means of a method NMG according to the invention for training a neural network for identifying tool states from image data.

[0038] According to the present invention, Figure 2 The neural network NN is trained as shown in . Figure 2 In the method shown in , the cutting edge 50 belongs to a first cutting edge type, which has not been used to train the neural network NN of the evaluation device AUSW. However, there are images BHS of historical cutting edges, i.e. images BHS of the second cutting edge type, which have been classified into different wear degrees MHS by means of process experts.

[0039] For the automatic learning of the neural network NN for the first cutting edge type, a plurality of images BNS of different cutting edges 50 of the first cutting edge type in different wear states must be detected. For example, such wear states can be assigned to categories such as "new state", "state in process" and "worn state". In the exemplary embodiment shown, the images BNS cover all the categories under consideration. The images BNS of new cutting edges 50 are temporarily stored for training the neural network NN.

[0040] However, instead of recording actual images BNS of the cutting edge 50 of the first cutting edge type for different wear states, manipulated image data BNS are used in the method according to the invention to train the neural network NN:

[0041] For this purpose, historical data HD about the cutting edges of the second cutting edge type are used, which have been detected in the past and are prepared in the memory of the evaluation device AUSW. These historical data HD each consist of a tuple of historical images BHS of the cutting edges of the second cutting edge type and the associated categories MHS of wear states, which correspond to the above-described categories of wear states. These categories MHS of wear states are the categories MHS identified by the process expert based on the images BHS of the cutting edges of the second cutting edge type.

[0042] In the evaluation device AUSW, first in a first characterization step ECMNS, characteristic features CMNS of the cutting edge 50 of the first cutting edge type are now detected, for example by evaluating the images BNS by software, for example by means of a software implemented by means of the programming language Python. The software analyzes a collection of images BNS of the cutting edge 50 of the first cutting edge type and calculates the characteristic features CMNS, in the illustrated embodiment, shape and color, of the cutting edge 50. These characteristic features CMNS of the cutting edge 50 of the first tool type are distinguished in that they appear almost unchanged in all images BNS of the same cutting edge 50 of the first cutting edge type.

[0043] In the second characterization step ECMHS, the characteristic features CMHS of the cutting edge of the second cutting edge type are detected by means of the evaluation device AUSW. The characteristic features CMHS are determined by means of an evaluation of the image BHS of the second cutting edge type stored in the memory of the evaluation device AUSW by software. Here, the shape and color of the cutting edge of the second cutting edge type also form the characteristic features CMHS of the second cutting edge type. The characteristic features CMHS of the second cutting edge type appear almost identically in all images BHS of the cutting edge of the second cutting edge type. However, the characteristic features CMHS of the second cutting edge type and the characteristic features CMNS of the first cutting edge type are significantly different from each other. Therefore, the corresponding characteristic features CMNS, CMHS are characteristic and selective for the first and second cutting edge types, respectively.

[0044] Based on the individual features CMNS of the first cutting edge type and the individual features CMHS of the second cutting edge type, a transfer function TF is now derived with the aid of a software module ATF for deriving transfer functions, which transforms the image BHS of the second cutting edge type into the image BNS of the first cutting edge type. With the aid of the transfer function, the image BHS of the cutting edge of the second cutting edge type can thus be used to train the neural network NN of the evaluation unit AUSW for evaluating the image BNS of the first cutting edge type.

[0045] The transfer function TF operates in Figure 3 In more detail:

[0046] By means of the transfer function TF, the image BHS of the cutting edge of the second cutting edge type is transformed into an image BNS of the cutting edge 50 of the first cutting edge type.

[0047] The transfer function TF firstly transforms the color of the cutting edge of the second cutting edge type so that the resulting image BNS is adapted in color to the cutting edge 50 of the first cutting edge type.

[0048] Furthermore, the shape of the cutting edge of the second cutting edge type can be transformed into the shape of the first cutting edge type by means of a transfer function TF: for this purpose, the image BHS of the second cutting edge type is divided into small segments SEG, which are appropriately shifted along the direction R, so that the modified image results in an image BNS of the shape NF of the first cutting edge type. Alternatively or additionally, a compression and / or expansion of the image information can be performed, so that the shape AF of the cutting edge of the second cutting edge type is transformed into the shape NF of the cutting edge 50 of the first cutting edge type. The color of the segments SEG is symbolized in the figures by means of colored dots or by means of hatched dots.

[0049] With the aid of the transfer function TF determined in this way, all images of the cutting edge of the second cutting edge type can now be converted into images of the first cutting edge type by applying the transfer function TF to all images BHS of the cutting edge of the second tool type in a further method step AHB. In this way, a new data set KADS of images BNS of the cutting edge of the first cutting edge type is generated. These newly generated images BNS of the data set KADS are now assigned their corresponding categories of wear state MHS. In the illustrated embodiment, the corresponding categories MHS of the wear state correspond to the corresponding categories of the wear state of the cutting edge of the second cutting edge type. The categories are therefore simply transferred to the newly generated images BNS of the cutting edge 50 of the first cutting edge type and are simply assigned image by image in this way. The corresponding categories MHS of the wear state are transmitted to the neural network with the aid of the data transmission TAM, so that the neural network NN not only receives the categories MHS of the wear state but also receives the corresponding newly generated images BNS of the cutting edge of the first cutting edge type.

[0050] In a training step MT, the neural network is trained using this data set KADS of new images BNS of cutting edges 50 of the first cutting edge type and the wear categories MHS respectively assigned to the images. After the neural network NN has been trained and thus forms a trained model TNM, the model is stored and can be used operatively by the evaluation unit AUSW during the machining of the workpiece 30. With the aid of the trained neural network NN, the wear state of the cutting edge 50 can be reliably determined in the estimation process PROG, and the time point for the replacement of the cutting edge 50 can be determined. In addition, the machining parameters of the device 10 for machining the workpiece 30 can be adapted in an adaptation step SEGM as a function of the wear state of the cutting edge 50 in order to compensate for the progressive wear of the cutting edge 50.

Claims

1. A method for training a neural network (NN) based on image data (BNS) for identifying a tool state (MHS), wherein the neural network (NN) is trained to identify a tool state (MHS) of a first tool type (50) and wherein image data (BHS) of a second tool type are used which are subjected to image processing (TF), the image data (BHS) of the second tool type being converted into image data (BNS) of the first tool type (50) by means of the image processing, wherein the image data (BHS) are converted such that characteristic features (CMHS) of the second tool type are transformed into characteristic features (CMNS) of the first tool type (50), wherein the neural network (NN) is trained based on the converted image data (BNS).

2. The method according to claim 1, wherein the image data (BHS) are converted such that a shape (AF) of the second tool type is transformed into a shape (NF) of the first tool type (50).

3. The method according to any of the preceding claims 1 to 2, wherein the image data (BHS) are converted such that the color of the second tool type is transformed into the color of the first tool type (50).

4. The method according to any of the preceding claims 1 to 2, wherein the first tool type (50) and / or the second tool type is designed for subtractive machining of a workpiece (30).

5. Method according to any of the preceding claims 1 - 2, wherein the first tool type (50) and / or the second tool type form a cutting and / or milling tool.

6. The method according to any of the preceding claims 1 to 2, wherein the image data originate from optical imaging (BEE) by means of a camera (60).

7. The method according to any one of the preceding claims 1 to 2, wherein the image data originate from optical imaging (BEE) by means of a microscopic camera.

8. The method according to any one of the preceding claims 1 to 2, wherein a tool state (MHS) is respectively assigned to the image data (BNS), and the tool state is respectively used together with the image data (BNS) for training the neural network (NN).

9. Method according to any of the preceding claims 1 - 2, wherein the tool state (MHS) is a wear state.

10. The method according to any one of the preceding claims 1 to 2, wherein the image data (BHS) are converted by: changing color channels, and / or compressing and / or expanding the image data (BHS), and / or dividing and changing, combining the image data (BHS).

11. The method according to claim 1, wherein the image data (BHS) are transformed by positioning and / or orienting and / or compressing and / or expanding the image data (BHS) in a changed manner.

12. A method for machining and / or manufacturing, which is performed with the aid of a first tool type (50), wherein a tool state (MHS) of the first tool type (50) is identified with the aid of a neural network (NN), which is trained according to a method (NMG) according to any one of the preceding claims 1 to 11.

13. A device for processing and / or manufacturing with the aid of a first tool type (50), comprising a neural network (NN) which is trained (NMG) according to a method according to any one of claims 1 to 11 or which is designed to implement a method according to claim 12.

Images