Method for determining a state of a tool

Abstract

The invention relates to a method for determining a state of a tool, wherein a plurality of optical images of a surface of the tool are first recorded in different illumination states. Image data of the plurality of optical images are subsequently processed to generate an image with surface structure information. This is then preprocessed to generate one or more preprocessed images. Finally, the one or more preprocessed images are used to classify the state of the tool into one of at least two classes by means of a machine learning method.

Description

[0001] Method for determining the condition of a tool

[0002] Technical area

[0003] The invention relates to a method for determining the condition of a tool. It further relates to a device for determining the condition of a tool and a system comprising such a device.

[0004] State of the art

[0005] Many tools, e.g., those used for machining materials such as milling cutters, drills, etc., are subject to wear due to their interaction with the workpieces. At a certain extent, this leads to a deteriorated machining result, longer machining times, and / or irreparable damage to the tool (or even the workpiece). Especially in the case of tools used in automated machines, the tools are therefore removed from the machine in good time before the aforementioned effects occur. If possible, the tools are then reconditioned for further use, e.g., by replacing or reconditioning machining elements (e.g., regrinding). If reconditioning is not (or no longer) possible, either generally or due to the degree of wear, the tools are recycled or disposed of.

[0006] The removal and replacement of tools can occur after a specified period of time or a specified number of machining cycles, with the duration or number being selected so that no negative effects are to be expected for the respective tool type up to that point, even in a worst-case scenario. Thus, the tools are generally removed, replaced, and reconditioned too early. Accordingly, the effective service life of the tools is shortened, and the number of tool changes and reconditioning cycles is higher than actually necessary.

[0007] Alternatively, the tools and / or the machining result on the workpiece are inspected by the operating personnel, depending on the tool, either visually or with the aid of aids such as magnifying glasses, microscopes, or measuring devices. The operating personnel then decide whether further use is possible.

[0008] The inspection is labor-intensive, usually requiring the machine to be stopped and often requiring the tool to be removed from the machine. Inconsistencies in the assessment can arise, especially when the inspection is performed by different people.

[0009] Systems for automatic wear detection have already been proposed. For example, CN 108107838 A (Shandong University) concerns wear detection on cutting tools. For this purpose, a cloud-based knowledge database with wear data is created and a detection model based on a support vector machine (SVM) is trained. The data is continuously updated to improve detection.

[0010] US Patent No. 7,479,056 B2 (Kycera Tycom) describes a fully automated system for verifying the identity and geometry of a drilling tool, reconditioning the tool, checking it against specified tolerances, adjusting a positioning ring on the tool shank, and cleaning and packing the reconditioned tool. The geometry is checked using optical units. These include head-mounted and front-mounted cameras for imaging the end and peripheral surfaces of the tool. Data generated during the tool inspection can be stored in the control system. Based on the captured images, specified reference points are identified and distances between them are measured. Furthermore, an initial assessment of the cutting edge condition is performed.

[0011] US Pat. No. 7,479,056 B2 does not disclose any details regarding the evaluation of the image data. The initial assessment is based on the external geometry of the tool and the condition of the cutting edges, although it is unclear how this can be determined or classified. EP 3 881 968 A1 (Fraisa SA, GF Machining Solutions AG) discloses a method for determining the wear condition of a tool, in which image data from at least one optical image is processed to detect a wear zone. The areal and / or spatial extent of the wear zone is determined, and the wear condition of the tool is classified based on the determined extent.

[0012] This method enables automatic classification. However, it has been shown, particularly with solid shank tools, that this method does not always achieve reliable classification.

[0013] Description of the invention

[0014] The object of the invention is to create a method belonging to the technical field mentioned at the beginning, which can automatically and reliably detect the condition of a tool.

[0015] The solution to the problem is defined by the features of claim 1. According to the invention, the method for determining a condition of a tool comprises the following steps: a) capturing a plurality of optical images of a surface of the tool under different illumination conditions; b) processing image data from the plurality of optical images to generate an image with surface structure information; c) preprocessing the image with surface structure information to generate one or more preprocessed images; d) classifying the condition of the tool into one of at least two classes based on the one or more preprocessed images using a machine learning method.

[0016] The processing of the image data, the preprocessing of the image with surface structure information, and the classification of the condition are all carried out computer-aided. The different lighting conditions are characterized in particular by different illumination directions.

[0017] Optical images are recorded in the visible range of the spectrum or in neighboring wavelength ranges (IR and UV). Generally, the tool or its area of ​​interest is illuminated, and light reflected from the surface is captured using a suitable device (camera; imaging optics with image sensor). The illumination light can have a continuous spectrum, a spectrum composed of multiple wave lines or frequency bands, or be monochromatic. The camera can also capture a broad frequency band, one or more narrow bands, or a specific frequency; it can also transmit monochromatic or polychromatic information. However, grayscale images are preferred for further processing.

[0018] Preferably, the image data comprises a two-dimensional image of the surface, and the image data corresponding to the two-dimensional image is used to generate an image with surface structure information. Efficient methods for this purpose are known.

[0019] The optical images do not have to depict the entire tool surface. In principle, it is sufficient to image those areas of the tool surface where wear or defects are expected, for example the cutting edges of a rotary tool. The shank of a shank-type tool or a radially inner area of ​​a grinding or cutting wheel can also be excluded from the image. If uniform wear can be expected, it may be sufficient to capture only representative areas. If localized wear (e.g. chipping) is to be expected, it is generally advisable to optically capture all potentially affected areas so that, for example, when several optical images are combined, the cutting areas that interact with the workpieces are fully captured.

[0020] The image with surface structure information is, in particular, a two-dimensional grayscale image. Classification into one of at least two classes can include assigning the tool to a wear class, e.g., "as good as new," "low wear," "medium wear," and "high wear." Classification can also be directed toward the measures to be taken, e.g., "continue use," "recondition," or "dispose of." Further classifications are possible, e.g., assignment to one of several possible reconditioning processes (cleaning, polishing, grinding, recoating, etc.), the assignment of a time for the next inspection, or an application area in which the tool can still be used—e.g., in processes involving easier-to-machine materials or requiring lower precision.

[0021] Classification can also be used to distinguish between flawless new tools and those with production and manufacturing defects, such as coating defects or microchips. In this case, classification into two classes may be sufficient. Depending on the results, tools with production or manufacturing defects are directly reconditioned, recycled, or disposed of. Using the method according to the invention, quality assurance can thus be systematized and automated.

[0022] This classification and the wear classification can be combined, allowing the condition of the tool to be assessed throughout all phases of its life, from the aforementioned manufacturing defects through initial wear marks from initial contact with the workpiece to manifest damage that affects machining quality. This damage includes, in particular, typical wear marks such as flank wear, crater wear, chipping and cracking, as well as adhesion in the form of built-up edges.

[0023] Classification can also directly include the generation of parameters for subsequent tool reconditioning, in which case the resulting number of classes (sets with a certain parameter combination) is considerably larger. Alternatively, the parameters for reconditioning are generated only after classification, based on the original image data and / or the preprocessed images. It is also possible to perform separate classifications for the condition of different areas of the tool (e.g., the tool face and tool surface), so that, for example, reconditioning is only necessary in certain areas of the tool.

[0024] Depending on the tool and application, wear can manifest itself in various ways. Compared to an unused, serviceable tool, for example, certain dimensions are reduced due to material removal on the tool, deformations occur that lead to a changed geometry, or individual regions exhibit signs of wear on the surface and / or to a certain depth.

[0025] The method according to the invention can be used in conjunction with a variety of tools, particularly those that act mechanically on the workpiece or that are removed due to interaction with the workpiece. These include, in particular, tools for machining materials, namely both rotary tools for milling, drilling, or threading, as well as stationary or linearly moving tools such as turning tools, punching tools, or saws.

[0026] Condition assessment can be carried out fully automatically, and compared to manual assessment, it also provides an objective picture, as all tools are assessed in the same way. Thanks to automation, condition assessment can be performed at regular intervals, thus preventing defects in the machining result caused by excessively worn tools and preventing tools that are still usable from being prematurely reconditioned, disposed of, or recycled. The use of a machine learning process enables easy targeting of new tools or tool types. Compared to an assessment using laboratory equipment, condition assessment can be completed in a very short time, which is a major advantage, especially when used in a reconditioning center where many tools need to be assessed.

[0027] The image data is preferably processed using photometric stereo analysis. This is an established method for shade-based surface reconstruction based on two or more images of the surface. The images are generated using the same surface and camera arrangement, and only the illumination state, specifically the illumination direction, is changed. This results in a simple assignment of the locations on the observed surface to the pixels in the images. For each pixel, several irradiances are determined. From these, surface orientations are reconstructed. This can be done in different ways - as described in more detail below - and, in a first step, provides images with surface structure information. In principle, it is possible to further process these images using an integration method, e.g., to obtain a relative height map.However, in the context of the present method, this second step is not absolutely necessary; the classification of the condition can be carried out based on the images with surface structure information, without the corresponding preprocessing including an integration step.

[0028] The photometric stereo analysis method is therefore not used for the usual purpose of three-dimensional reproduction of a surface, but for generating images that include surface structure information that enable improved detection of wear phenomena.

[0029] The image containing surface structure information is preferably a height image and / or a curvature image. In the height image, each pixel encodes the relative height of the corresponding point on the surface relative to a reference surface. In the curvature image, each pixel encodes the local curvature. It has been shown that elevations and depressions, such as those created by the use of tools, especially grinding and drilling tools, are particularly clearly visible in the height image. The curvature image enables particularly good detection of fine local defects.

[0030] In principle, it is possible to base the classification on multiple images with surface structure information obtained in different ways from the multiple optical images, for example, a height image and a curvature image. This allows the advantages of each image type to be combined.

[0031] Alternatively or additionally, other image types can be used, such as those generated during photometric stereo analysis, such as tilt images in various directions. These allow for the accurate detection of shape deviations, such as a cutting edge. The additional use of a texture image can be helpful in detecting superficial rust damage or discoloration, allowing these to be distinguished from the actual damage.

[0032] The pre-processing step advantageously includes trimming to a region of interest, where the region of interest includes those areas of the tool surface affected by wear or where defects are expected. This reduces processing effort and improves the quality of the classification.

[0033] When determining the wear of a solid shank tool, the cutting edge runout area and the shank end are particularly cut away during trimming. In another embodiment, areas clearly outside the cutting edge area where wear is expected can also be ignored.

[0034] Alternatively, such trimming can be omitted. In principle, the machine learning process can be trained, or the features for clustering can be selected, in such a way that the areas not affected by wear or defects do not lead to a falsification of the classification, i.e., differences in these areas do not influence the classification of the tools.

[0035] In a first group of embodiments, the machine learning method is a supervised machine learning method. Suitable supervised learning methods are based, for example, on a so-called support vector machine (SVM) or an artificial neural network (ANN).

[0036] Within the scope of such methods, wear or defect features are advantageously isolated in the image containing surface structure information to generate one or more preprocessed images. For this purpose, straight lines corresponding to cutting edges can be identified in the image and subtracted from the image, so that only the structures deviating from the straight lines, which correspond in particular to the wear features or defects, remain in the preprocessed image. Similarly, distinctive frequencies, which arise, for example, due to constant distances between adjacent cutting edges, can be identified, and the structures corresponding to the identified frequencies can be subtracted. In this way, too, primarily the wear or defect features remain in the preprocessed image.

[0037] The corresponding preprocessing is carried out in particular by at least one of the following methods: a) applying a Hough transform to the image with surface structure information; b) applying a Fast Fourier transform to the image with surface structure information.

[0038] Finally, as described in detail below, the Hough transformation detects straight lines in the image containing surface structure information, which correspond, for example, to cutting edges of the tool. The corresponding image values ​​are then subtracted from the image containing surface structure information. What remains is the image information corresponding to wear characteristics.

[0039] By applying a Fast Fourier Transformation to the image with surface structure information, the characteristic frequencies are determined based on the parallel cutting edges that repeat in the image with surface structure information. Features that appear at this frequency in the image data are most likely associated with undamaged structures on the tool surface, such as the cutting edges. In contrast, features that do not exhibit this frequency are most likely due to defects. To obtain the preprocessed image, the regions corresponding to the frequencies can be masked in the frequency domain. The back transformation to the spatial domain then provides an image of the defects.

[0040] Advantageously, particles with pixels in a given brightness range are identified in the pre-processed image and, based on the identified particles, several of the following quantities are determined as features for the machine learning process: a) area of ​​the particles; b) filled area of ​​the particles; c) equivalent surface diameter of the particles; d) area of ​​a bounding box of the particles; e) major axis length of the particles; f) maximum Feret diameter of the particles; g) convex area of ​​the particles; h) minor axis length of the particles; i) circumference of the particles; j) mean value of the pixels in the brightness range; k) standard deviation of the pixels in the brightness range; l) number of pixels in the brightness range; m) number of particles.

[0041] A particle is a contiguous region of bright pixels in a binary image generated from the preprocessed image (or, analogously, a contiguous region of pixels in a grayscale or color image whose brightness value exceeds a threshold). These features have been shown to be well-suited for characterizing wear effects on tool surfaces, especially those made of metallic or ceramic materials, using preprocessed optical images.

[0042] The following quantities are preferably determined as features for the machine learning process: a) area of ​​the particles; c) equivalent surface diameter of the particles; e) major axis length of the particles; h) minor axis length of the particles; j) mean value of the pixels in the brightness range and m) number of particles.

[0043] These features are suitable for representing state-relevant properties of the preprocessed images.

[0044] Preferably, the values ​​of the selected features are normalized. Specific weighting of individual features is possible to specifically consider the different relevance of the features during classification. This weighting can be performed after normalization or obtained by using a feature-dependent norm.

[0045] In a second group of embodiments, the machine learning method is an unsupervised machine learning method, wherein in a first step, several features are determined based on the preprocessed image or the plurality of preprocessed images and in a second step, clustering is carried out.

[0046] In such methods, to generate multiple preprocessed images, the image containing surface structure information is divided into multiple overlapping tiles, preferably using a sliding window method. For each preprocessed image corresponding to a tile, multiple predefined features are determined for clustering. The size and overlap of the tiles depend on the tool geometry and the region of interest (ROI) to be assessed. In particular, the image of a tool surface is divided into at least four tiles, especially at least ten tiles.

[0047] The division into tiles enables clustering that takes local defects into account. In addition to classification, assigning individual tiles to different clusters—according to the respective local defects—can provide additional information on the wear of the tool being assessed.

[0048] Preferably, the features from the preprocessed images comprise several from the following groups: a) a mean, a variance and / or a median of the preprocessed image further processed using a two-dimensional Gabor filter; b) a sum of a binary image obtained from the preprocessed image further processed using the two-dimensional Gabor filter; c) features based on a local binary pattern analysis; d) Haralick texture features; e) a mean and / or a standard deviation of pixel values ​​of the preprocessed image.

[0049] Preferably, the values ​​of the selected features are normalized. Specific weighting of individual features is possible to specifically consider the different relevance of the features during classification. This weighting can be performed after normalization or obtained by using a feature-dependent norm.

[0050] The two-dimensional Gabor filter can be based on a predefined kernel with which the image sections are filtered. Alternatively, a so-called filter bank is created, which contains various filter kernels with different frequencies and orientations. The preprocessed image is filtered with all filters in the filter bank. Finally, all output images from the filters are combined into a single combined image (filter superposition).

[0051] Other or additional features are possible. It can be advantageous to use the histogram of gradients (HOG) to extract features, especially in addition to the local binary pattern (LBP).

[0052] Clustering is preferably performed in 3-8, and especially 4-6, classes. It has been shown that with an appropriate number of classes, a sufficiently fine subdivision can be achieved to differentiate between tools with regard to the measures to be taken (e.g., reprocessing, reuse, disposal). At the same time, the classification is already robust even with a comparatively small number of tools to be assessed or a relatively small training set.

[0053] In certain cases, it may be advisable to provide more than 8 classes, e.g., when the tools are to be assigned to specific processing parameters. Clustering is preferably performed using a Gaussian Mixture Model (GMM) or a k-means++ method.

[0054] Based on clustering, the entire tool can be assigned a condition class that corresponds to the worst identified condition class in the cluster, i.e., the condition class is determined in particular by the condition of that area of ​​the tool (that tile) that exhibits the greatest wear.

[0055] Because each tile is evaluated individually, it is possible to trace it back to the overall picture and assign it to it. This provides information about which condition class occurs how often (as a percentage) on each tool. For example, it is possible to specify the condition of the tool surface as a percentage relative to its new condition. This provides information about the overall condition of a tool. It is also possible to spatially assign wear or defect locations.

[0056] Advantageously, a broken cutting edge of the tool is detected based on the image data of the plurality of optical images and / or based on the image with surface structure information and / or based on the one or more cropped images with surface structure information, and if a broken cutting edge is detected, the tool is immediately assigned to a corresponding condition class.

[0057] In this case, the tool is not classified using machine learning. It has been shown that tooth breakage in optical images of the tool surface leads to features that are clearly different from the other optically detectable wear or defect features. Therefore, on the one hand, the inclusion of tools with tooth breakage negatively affects the classification using the machine learning method. On the other hand, it is also possible to easily detect tooth breakage using common image processing methods (or a dedicated upstream machine learning method) and to directly assign the corresponding tools to the appropriate condition class. The method can be used to determine the condition of a grinding or drilling tool for workpiece machining in a machine tool, especially a solid shank tool.

[0058] Advantageously, a first state of a shell-side cutting geometry and a second state of a front-side cutting geometry of the solid material shank tool are determined separately.

[0059] This allows for the different geometric conditions and the different requirements for the integrity of the respective cutting edges to be taken into account. For example, the front-side cutting edges of milling tools are often subjected to less stress than those of drilling tools, while the situation is exactly the opposite for the cutting edges on the outer surface.

[0060] To classify an "overall condition," the first condition and the second condition can be maintained and / or used differently. In the simplest case, a tool should be replaced if at least one of the two conditions requires replacement, and the tool should be reconditioned if at least one of the two conditions requires reconditioning. If the overall performance of the tool does not simply correspond to the weakest link, but results from an interaction between the performance of both tool sections, it may be useful to calculate the conditions in a more complex way, so that reconditioning or replacement only occurs when the overall performance requires it.

[0061] Advantageously, an algorithm based on a first data set is used to determine the first state and an algorithm based on a second data set is used to determine the second state, wherein the first data set and the second data set are different.

[0062] In particular, the two data sets are essentially disjoint. For example, in a supervised learning method, the first data set comprises image data as training data showing the surface area of ​​worn tools, and the second data set comprises image data showing the face area of ​​worn tools. The first and second data sets overlap at most in a transition area (edge ​​or radius) between the surface and face.

[0063] Not only the data volumes can vary, but also the algorithms used, for example, different machine learning algorithms or algorithms with different parameters (e.g. network topologies in neural networks) can be used.

[0064] A method for reconditioning a tool preferably comprises the following steps: a) determining the condition of a tool using the described method according to the invention; b) controlling at least one device for reconditioning the tool, in particular by means of a grinding process, when the condition meets predetermined conditions.

[0065] The specified conditions include, in particular, the classification of the tool's condition. Thus, the classes can be defined from the outset to correspond to the measures to be implemented ("continued use," "to be reconditioned," "to be disposed of"), or the measures can be derived directly or indirectly from a classification. For example, the wear condition can be classified into eight classes 1-8 (1: as good as new, 8: heavily worn). The conditions are specified such that if the tool is classified in classes 1 and 2, it continues to be used, if it is classified in classes 3-6, it is reconditioned, and if it is classified in classes 7 and 8, it is recycled or disposed of.

[0066] The condition can not only serve as a basis for deciding whether to reprocess using the appropriate equipment, but can also be relevant for the steps to be performed during the reprocessing process. For example, several reprocessing steps are available (cleaning, polishing, grinding, multiple grinding processes, etc.), and different selections are made depending on the condition.

[0067] In particular, the parameters of the grinding process and / or the machining geometry of the tool reconditioning device are determined based on the determined condition. These parameters can include, for example, a grinding path, one or more feed values, or similar. The machining geometry defines the extent, location, and type of machining of the tool. The machining geometry can be determined directly from the optical image and / or from machining results, e.g., classification.

[0068] This enables tool-specific, needs-based reconditioning. For example, in grinding processes, the (additional) material removal can be minimized, thus maximizing the tool's service life. Furthermore, it's not necessary to conduct a (repeated) extensive examination of the tool prior to reconditioning because the required data is already available.

[0069] In a preferred embodiment, a lighting system for illuminating a surface of the tool with different lighting conditions and a camera for capturing multiple optical images of the surface of the tool under the different lighting conditions are arranged at a first location. Data obtained at the first location is stored in a database. The processing device is arranged at a second location, and the processing device retrieves data from the database.

[0070] The two deployment locations are remote from each other and are typically located in a different facility or plant. The database can be centrally located, so that the data is collected and used decentrally but stored centrally. However, the database can also be stored at the first deployment location or the second deployment location, or the data can be maintained at both locations and regularly synchronized.

[0071] This also makes it possible, in particular, to fully record and retain data relevant for a specific tool even if the tool is used by different users (after preparation) or even prepared by different service providers.

[0072] Preferably, the tool is provided with a unique identifier, and the data assigned to the tool is linked to the unique identifier in the database. The unique identifier is preferably attached to the tool in a machine-readable form, e.g., as an optical marking (barcode, matrix code, alphanumeric, etc.) or stored on a data carrier (e.g., RFID). The unique identifier ensures correct assignment.

[0073] Alternatively, the recorded data is stored on a data storage device. This device is then transported from the first to the second location along with the corresponding tool. In principle, the data storage device can also be integrated into the tool.

[0074] In another case, the recording of at least one optical image takes place directly at the processing facility.

[0075] A device according to the invention for determining the condition of a tool comprises a) an illumination system for illuminating a surface of the tool with different illumination conditions; b) a camera for recording multiple optical images of the surface of the tool under the different illumination conditions; c) a first image processing module configured to process image data from the multiple optical images to generate an image with surface structure information; d) a second image processing module configured to further process the image with surface structure information to generate one or more preprocessed images; and d) a classification module configured to classify the condition of the tool into one of at least two classes using a machine learning method based on the one or more preprocessed images.

[0076] The camera is, in particular, a line scan camera. With this, the surface of the tool can be recorded line by line. To record the surface of the casing of a solid-material shank tool, the tool is preferably rotated step by step around its longitudinal axis using a spindle, and the surface is recorded line by line using a fixed line scan camera. In each rotational position, the tool is illuminated successively with the different lighting conditions. This results in image data that correspond to a developed view of the tool casing, with each line initially appearing multiple times - corresponding to the different lighting conditions. However, from this image data, several images can be easily generated by combining lines whose spacing corresponds to the number of lighting conditions, showing the developed view under the different lighting conditions.

[0077] In a further preferred embodiment, the illumination system and the camera are integrated into a processing machine with a holder for the tool, in particular such that the illumination system can illuminate the tool and the camera can record the optical images of the surface of the tool when the tool is held in the holder.

[0078] The image processing modules and the classification module can be accommodated in a processing device, wherein the processing device is wholly or partially contained in the processing machine or is arranged externally thereto and connected to it in terms of signals.

[0079] The processing machine can be, for example, a machine tool for drilling or milling, or a machining center. The holder can be, for example, the work spindle, a holder in a magazine for holding tools for tool changes, or a transport holder for transferring the tool between the work spindle and the magazine and vice versa.

[0080] A processing machine which comprises a camera and which is connected to a processing device or which contains the latter in whole or in part is therefore particularly advantageous, the processing device comprising the two image processing modules and the classification module.

[0081] A machine tool arrangement according to the invention comprises a machine tool, preferably a machining center, a milling center, or a drilling center, and a device according to the invention for determining the status. The camera is integrated into the machine tool or arranged on it. The image processing modules and the classification module are accommodated in a processing device. The processing device is entirely or partially contained in the machine tool or arranged externally thereto and connected to it via signals.

[0082] An advantageous arrangement for reconditioning a tool comprises a) a device according to the invention for determining the condition of a tool; b) a device for reconditioning the tool, in particular by means of a grinding process; and c) a controller for controlling the reconditioning device, which is configured to receive information about the condition from the device for determining and to control the device for reconditioning the tool depending on the information received.

[0083] For example, the processing machine is preferably assigned a camera, by means of which the wear status of the tools used in the machine can be regularly monitored, e.g., during each tool change. According to the invention, the image data is further processed directly at the location of the processing machine, so that a decision can be made based on the classification whether the tool can still be used. If this is the case, it is stored in the tool magazine. Otherwise, it is discarded, and data on the wear status (and, if applicable, the image data or other information obtained from it) are stored in a database. The discarded tool is then physically transported to the reconditioning facility. This reads the data assigned to the tool from the database and controls the reconditioning facility based on this data.The reconditioned tool is then transported to the same or another processing machine and used again there.

[0084] During reconditioning, additional data can be used, e.g., regarding the tool's history (number of usage cycles, previous reconditionings, etc.) or regarding the customer's requirements for their specific machining processes. Further advantageous embodiments and combinations of features of the invention emerge from the following detailed description and the entirety of the patent claims.

[0085] Short description of the drawings

[0086] The drawings used to explain the embodiment show:

[0087] Fig. 1A, B a schematic front view and a schematic side view of an embodiment of an image recording device for use in a method according to the invention for determining a wear condition of a shank tool;

[0088] Fig. 2 synthetic images of the lateral surface of the shank tool;

[0089] Fig. 3 Height images of the surface of end mills with different

[0090] Wear;

[0091] Fig. 4 is a schematic representation of the cropping of the height image to a

[0092] Region of Interest (ROI);

[0093] Fig. 5 shows the preprocessing of the height image using the Hough

[0094] Transformation;

[0095] Fig. 6 the corresponding preprocessed image as a basis for the supervised machine

[0096] Learning classification;

[0097] Fig. 7 shows the elevation image preprocessed using Fast Fourier Transform as a basis for supervised machine learning classification;

[0098] Fig. 8A-E Results of clustering using the k-means++ algorithm; and

[0099] Fig. 9 is a schematic block diagram of a system according to the invention for

[0100] Determining the wear condition and preparing a tool.

[0101] In principle, identical parts are provided with identical reference numerals in the figures. Ways of carrying out the invention

[0102] An embodiment of the method according to the invention is explained below using the example of determining the wear of the main cutting edges of a solid material end mill.

[0103] For solid end mills, two types of wear must be distinguished:

[0104] Flank wear: This refers to wear in the flank area adjacent to the cutting edge. This is caused by excessively high cutting speeds, low wear resistance of the cutting edge, or insufficient coolant supply.

[0105] Cutting edge chipping: These can occur at the cutting tip as well as on the flank. Their occurrence is favored when a heavily worn tool is continued to be used or when the tool is used to machine excessively hard workpieces.

[0106] In principle, the method according to the invention detects both types of wear. However, as explained in more detail below, the results can be improved if cutting edge chipping is detected in a prior step—e.g., using conventional image processing methods—and the tools without cutting edge chipping are then analyzed using the method to classify flank wear.

[0107] Figures 1A and 1B show a schematic front view and a schematic side view of an embodiment of an image recording device 1 for use in a method according to the invention for determining the wear condition of a shank tool. Corresponding devices are generally known and commercially available, e.g., devices of the trevista® DOME series from SAC Sirius Advanced Cybernetics GmbH, Karlsruhe, Germany.

[0108] The image recording device 1 first creates optical images of the lateral surface of the end mill 2, or more precisely, of the working area of ​​the lateral surface (without the shank). For this purpose, the end mill 2 is clamped with its shank in a vertical chuck 3 of a spindle 4. The milling cutter is illuminated by a dome light 5. This comprises a dome 6 with a diffusely reflecting, concave inner surface facing the spindle with the end mill 2. The dome 6 is surrounded by a black base plate 7. An LED strip 8 is arranged on this, surrounding the edge of the dome 6. The strip comprises 8 segment-shaped lighting elements 8.1...8.8, which can be selectively controlled to generate different illumination directions. The dome 6 is also arranged on a linear guide 9 that runs radially with respect to the end mill 2, so that the focus can be adjusted to the respective measuring object.

[0109] A line scan camera 10 is arranged behind the dome 6 and can optically capture the area of ​​the lateral surface of the end mill 2 facing the dome 6 through a vertical slot in the dome. In the exemplary embodiment, this is a commercially available 8K monochrome line scan camera. This captures a pixel line with 8192 pixels in grayscale per image. The lens used is a precision lens with a focal length of 105 mm and an aperture of f / 5.6.

[0110] The spindle 4 with the chuck 3 can be precisely rotated around its vertical axis by means of a motor with an encoder, so that the entire surface can be captured by the line scan camera 10 in one complete rotation. In each rotational position, the line scan camera 10 records one pixel line at a time. The line scan camera 10 and the illumination segments 8.1...8 are triggered together for each image acquisition.

[0111] The end mill 2 to be examined is clamped into the chuck 3 of the image recording device 1 described above. The distance between the mandrel 6 and the end mill 2 is adjusted to achieve the best possible illumination of the area of ​​the tool casing to be examined and to enable the line scan camera 10 located behind the mandrel 6 to focus on this area. The tool is rotated step by step around its longitudinal axis, with four images with different illumination directions being recorded at each rotational position. To change the illumination direction, four of the light elements 8.1... 8.8 of the LED strip 8, which are evenly arranged in segments around the circumference of the mandrel 6, are selectively controlled:

[0112] The described device enables the entire surface to be scanned within a few seconds, enabling a high throughput of tools to be assessed.

[0113] Since the relative arrangement between camera and tool does not change between the acquisition of images with different illumination directions, pixels of the individual images with the same rotational position can be clearly assigned to one another, and multiple irradiances can easily be assigned to each pixel of the acquired line. The line-by-line acquisition takes place until the entire peripheral surface has been scanned, i.e. the end mill 2 has been rotated 360° in the chuck 3. The individual rows of pixels corresponding to the same illumination are then each cut together to form a single image. This results in four images of the peripheral surface of the tool. Synthetic images are then derived from these images: a) an inclination image in the X direction; b) an inclination image in the Y direction; c) a curvature image; d) a height image; e) a texture image.

[0114] The commercially available image processing software Coake® from the aforementioned SAC Sirius Advanced Cybernetics GmbH enables the generation of such synthetic images from the images generated with the image acquisition device 1. In principle, a surface normal can be determined for each point on the surface (pixel) in a conventional manner from the recorded image intensities, the known illumination properties, and the surface radiation properties; see, for example, BBKP Horn, MJ Brooks: "Shape from Shading," The MIT Press, Cambridge MA, 1989.

[0115] From the surface normals, the inclination images in the X and Y directions can now be determined. These result from the brightness changes depending on the respective direction. The corresponding derivatives can be calculated using finite differences, Sobel operators, or other derivative kernels.

[0116] Based on the two slope images, a height map can be determined by integration. Various methods are known for this purpose, which in particular prevent deviations and ambiguities from accumulating due to integration and thus leading to distorted results. The height image corresponds to the height map.

[0117] From the height map Z(x,y) the curvature map K(x,y) can be calculated as follows:

[0118] This approach takes curvatures in both directions into account, but is challenging to apply due to disturbing image noise. Techniques and filters are known to address this problem.

[0119] Further information on calculating the corresponding images can be found in the literature, e.g., in A. Distante, C. Distante, Handbook of Image Processing and Computer Vision, Vol. 3, Chapter 5: Shape from Shading, Springer Nature Switzerland AG, 2020; RJ Woodham, "Determining surface curvature with photometric stereo," Proceedings, 1989 International Conference on Robotics and Automation, Scottsdale, AZ, USA, 1989, pp. 36-42, vol. 1, doi: 10.1109 / RCBCT.1989.99964.

[0120] The texture image essentially corresponds to the optical representation of the surface.

[0121] The different images are shown in Figure 2: a) inclination image in x-direction; b) inclination image in y-direction; c) curvature image; d) height image; e) texture image.

[0122] It is clearly visible that the defects in the area of ​​the cutting edge stand out particularly well from the surrounding image on the curvature and height images. The inclination images are more difficult to interpret due to the large differences in inclination in front of and behind the cutting edge, and the contrast in the texture image is significantly lower.

[0123] In the described embodiment, the height image is primarily used for further processing, as it provides good results in terms of defect detection. Additional or alternatively, other synthetic images can also be used, such as the x- and / or y-tilt image, which are well suited for analyzing profiled tools and thread cutters, or the curvature image, which enables the detection of fine and small defects in the micrometer range. The texture image, on the other hand, is particularly well suited for detecting defects in a coating.

[0124] Figure 3 shows elevation views of the surface of end mills, each showing varying degrees of wear. Three images are shown for each class: a) very good condition, new; b) very good condition, as good as new, immediately after reconditioning; c) good condition, signs of wear; d) moderate condition, signs of wear and damage; e) poor condition, major damage such as chipping or tooth breakage.

[0125] It is clearly visible from the images that the relevant signs of wear are clearly visible in the elevation image.

[0126] Next, the height image is cropped to a region of interest (ROI) 15. This means that subsequent processing can be limited to the area of ​​the tool surface where wear is expected. Figure 4 shows the cropping process, with the shank end 11 positioned on the left in the height image shown and the face on the right. On the right-hand side, i.e. the face, the outermost white pixel is searched for and cropped up to this point (step 12). On the left-hand side, cropping is also carried out up to the outermost white pixel (step 13), and a constant width is additionally selected to reach the cutting end (step 14). The width can essentially be selected to be the same for all end mills, since the distance is always approximately the same.

[0127] The cropped elevation image now forms the basis for subsequent classification using a machine learning method. Two process variants are presented below: first, a supervised machine learning method, followed by an unsupervised machine learning method.

[0128] Supervised Machine Learning

[0129] Within the framework of the supervised machine learning process, in the context of the illustrated embodiment, a tool is to be assigned a wear class based on the height image of its surface, which corresponds as closely as possible to a wear class as would be determined by an experienced expert.

[0130] In practice, experts evaluate such a tool by looking for the largest wear points. These points determine how worn the entire tool is. This means that the classification assigned by the experts depends on a specific area on the tool. This area determines the severity of the wear, not the overall appearance of the tool. However, the classification assigns the tool as a whole to this class. This problem is addressed with specific data preparation in the context of the supervised machine learning method.

[0131] During data preparation, the cropped elevation image is simplified so that only the wear surfaces, which are ultimately important for classification, are visible. Two methods for data preparation are presented below: 1. Hough transformation

[0132] Using a so-called Hough transformation, straight lines are detected in an image (see US Pat. No. 3,069,654, PVC Hough). The straight lines detected in the height image, corresponding to the cutting edges, are subtracted from the cropped height image, essentially generating a preprocessed image that shows only the wear characteristics. This preprocessed image can then be reliably classified because essentially all of the features it contains are relevant for classification.

[0133] Before the actual Hough transformation is performed, the cutting edges are cut free. Since the wear is primarily located on the cutting edge, the other areas in the image can be eliminated. This allows for a reduction in the amount of data and computation time. After cutting free, the situation is shown in Fig. 5a.

[0134] The image is then further processed into an edge image by converting it to a binary image and using a Canny edge detector (Figure 5b). A Hough transform is then used to detect the straight edges and transform them back into an image. This image now contains the portion of the edges that are straight (Figure 5c). If this image is subtracted from the edge image, only the wear features remain. This preprocessed image with the wear features is then used as the basis for the supervised machine learning classification. It is shown in Figure 6.

[0135] 2. Fast Fourier Transform

[0136] During preprocessing using Fast Fourier Transformation (FFT), the pronounced frequencies in the tool surface image are removed, leaving only the defects that naturally do not exhibit these frequencies. These frequencies result from the parallel cutting edges and grinding marks. As is well known, they can be identified in image data using an FFT.

[0137] Specifically, an FFT is first applied to the image to be preprocessed. This yields the dominant frequencies. The strongest signal will come from the angle of the cutting edge. This angle can then be determined based on the transformed image. The original image to be preprocessed is then rotated by the specified angle so that the cutting edges run horizontally in the rotated image. The rotated image is then subjected to an FFT and thereby transformed into the frequency domain. The unwanted frequencies (corresponding to the straight, intact cutting edges) are then eliminated using a mask. The new image thus generated in the frequency domain is then inversely transformed using an FFT and rotated back by the initially determined angle. This produces an image in which only the irregular structures, in particular the defects, are present.A threshold can now be applied to convert this resulting image into a preprocessed image with wear characteristics, which will serve as the basis for supervised machine learning classification. This preprocessed image is shown in Figure 7.

[0138] Next, the features for the machine learning classifier are generated. This searches for features that are as informative as possible about the wear class of the tools to be assessed. The features are based on particle shape descriptors. A particle is a contiguous region of bright pixels ( 1) in the binary image. They can be identified using common image processing methods. In the example, the following 13 features (or a selection of them) are determined:

[0139] To compare features despite different value ranges, they are normalized. This is done using the so-called L2 norm, which is calculated as follows:

[0140] For solid end mills, for example, a reduced vector can be used as the feature vector, which includes only the following features: a) particle area; c) equivalent surface diameter of the particles; e) major axis length of the particles; h) minor axis length of the particles; j) mean value of the pixels in the brightness range; and m) number of particles. Depending on the tool type, other combinations may be useful. These can be determined using correlation analyses, whereby only one of the features that exhibit a mutual correlation above a certain threshold (e.g., 0.95) is included in the vector.

[0141] The vectors are then classified using a support vector machine (SVM). The results are discussed below.

[0142] Unsupervised Machine Learning

[0143] The unsupervised machine learning method identifies related groups (hereinafter "clusters") in the data under analysis. For this purpose, individual sections of the cropped elevation image are examined and clustered in the present example.

[0144] The sections are generated using a sliding-window approach with overlap. This involves iterating through the input image with a predefined window and a step size. The features are calculated from the generated image sections. These can then be represented as a point cloud in feature space, with each point representing an image section. For example, window sizes of 400 x 400 px and a step size of 220 px can be used.

[0145] For unsupervised machine learning, different features are used than those mentioned above, particularly because suitable features are now required for individual, small image sections. The focus is on texture features, since tool damage has a significant impact on the texture. Normalization was performed using the L2 norm as described above. Using the following methods, a certain number of features were determined for each image section for further processing:

[0146] 2d Gabor filter;

[0147] Local Binary Patterns (LBP)

[0148] Haralick features; histogram-based features. The methods are described below.

[0149] The 2D Gabor filter makes it possible to detect irregular structures. It consists of a Gaussian filter kernel, which is processed with the image using convolution. Three parameters can be adjusted for the filter kernel: the theta parameter, the frequency, and the standard deviation. Theta determines the orientation of the kernel in degrees. All features aligned with the kernel are highlighted. This means that if the theta value corresponds to the angle of the cutting edges, these and the grinding marks are highlighted in the filtered image.

[0150] The ideal frequency was determined through experiments. Generally, a lower frequency corresponds to higher resolution. The appropriate standard deviation is also determined empirically based on the given tool geometry.

[0151] For each of the transformed image sections, the following features can then be formed from the totality of the respective pixel values: a) the mean of the grayscale image; b) the variance of the grayscale image; c) the median of the grayscale image; d) the sum of the binary image generated from the grayscale image using a threshold.

[0152] Local binary patterns (LBPs) can be used to analyze local patterns in an image. The number of patterns occurring is recorded in a histogram. Each pattern corresponds to a binary code generated by comparing a central pixel with neighboring pixels. Each neighboring pixel is assigned the value 1 if its brightness value equals or exceeds that of the central pixel, and the value 0 otherwise. The ordered sequence of these bit values ​​then results in the binary code.

[0153] LBP has three preset parameters: the radius, the number of contour points, and the method. The radius determines which neighboring pixels are analyzed, and the number of contour points corresponds to the number of neighboring pixels on this radius (e.g., 8 contour points for radius 1 or 16 contour points for radius 2). A radius of 1 has been shown to produce good results for wear surfaces on solid end mills.

[0154] As a method for generating the binary code, only uniform patterns were recorded individually, and all non-uniform patterns were assigned to a further bin. Uniform patterns are those that have a maximum of two 0-1 or 1-0 transitions. With 8 contour points (radius 1), this results in 58 different uniform patterns (corresponding to 58 bins) and a further bin for all non-uniform patterns. The 59 features are formed from the corresponding values ​​in the histogram. The Haralick features are derived from the grayscale co-occurrence matrix (GLCM), which summarizes which pixel pairs occur how often, with the pixel pairs viewed horizontally, vertically, or diagonally (Haralick RM, Shanmugam K. & Dinstein I. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics 3, 610-621 (1973)). In this case, the distance between the pixels of a pixel pair is chosen to be 1 (neighboring pixel).The following 14 features were calculated from the GLCM in all four possible pixel pair directions, for a total of 56 features:.

[0155] The histogram-based features are the mean and standard deviation of the pixel values ​​of an image section.

[0156] The set of total (texture) features can be reduced to the essential and meaningful features by appropriate measures, such as correlation analysis.

[0157] The reduced set can be approximated by further dimensionality reduction, for example, using principal component analysis (PCA) to, say, three principal axes. Clustering is then performed using an appropriate clustering algorithm. For this purpose, the k-means++ algorithm or a Gaussian Mixture Model (GMM) can be used.

[0158] In the context of the example, each of the images should be assigned to one of the following four wear classes: Class NEW (new tool)

[0159] Wear class 1 (low wear)

[0160] Wear class 2 (medium wear)

[0161] Wear class 3 (heavy wear)

[0162] The k-means++ algorithm is applied to the data with the goal of clustering into a specified number of classes. Within the algorithm, a corresponding number of cluster centers are defined and placed in the point cloud according to specific criteria. The data points are assigned to the nearest cluster center. The mean value of each cluster center is calculated from the assigned data points, and the center is shifted accordingly. This process is repeated until the cluster centers no longer shift.

[0163] An analysis of the results using the so-called elbow method revealed that a class number of 4, 5, or 6 should accurately represent the data structure. According to a silhouette analysis, the best silhouette coefficient was obtained for cluster number 2. However, this number would not allow for a sufficiently fine differentiation of tool wear. Cluster number 4 yielded the second-best silhouette coefficient. Based on this analysis, the number of four classes also seems reasonable.

[0164] The GMM clustering algorithm also specifies the number of classes. This algorithm assumes that the clusters are Gaussian distributed, so a Gaussian distribution is fitted to the data for each class. Silhouette analysis has shown that—apart from the unhelpful class number of 2—good coefficients are obtained for cluster numbers 4 and 5. Thus, specifying four classes seems reasonable here as well.

[0165] The results of clustering using k-means++ are shown in Fig. 8A-D. The point cloud was divided into four groups, with the cluster centers marked with white numbered circles. In Fig. 8A-D, each group of tools is highlighted with black crosses:

[0166] Fig. 8A New tools; The new tools fill cluster 0 very well. New tools can therefore be reliably classified in cluster 0.

[0167] Fig. 8B reconditioned tools;

[0168] There is a slight scatter in cluster 3 for the tools immediately after reconditioning. However, most tools – similar to new tools – were again assigned to cluster 0, which seems reasonable for freshly reconditioned tools.

[0169] Fig. 8C Wear class 1 according to expert classification;

[0170] Wear class 3 occupies cluster 3 very well. However, a scattering is also evident in cluster 1, and one strand even extends into cluster 2: The analysis showed that the corresponding data points all originate from the same tool, which exhibited very heavy contamination in the chip space.

[0171] Fig. 8D Wear class 2 according to expert classification;

[0172] Wear class 2 fills cluster 1 well, with elements in cluster 3 corresponding to the less worn areas of the surface. There are no outliers in cluster 2.

[0173] Fig. 8E Wear class 3 according to expert classification.

[0174] The scatter here extends through all clusters 1, 2 and 3. This indicates areas with high levels of wear.

[0175] From a qualitative point of view, the results of clustering using GMM are very similar to those of kmeans++ clustering.

[0176] To verify whether the wear class assignments to the tools match the experts' classifications, the tools must be analyzed individually. It should be noted that even heavily worn tools usually have regions where wear is low—thus, the image sections with high levels of wear are crucial for assigning a tool to a wear class. Unsupervised machine learning used the approach whereby the determined wear class of a tool always corresponds to the largest wear point on the tool.

[0177] To validate the approaches, 200 used end mills were used. Each was classified into one of three wear classes by three experts. Since there were differences between the individual expert classifications, only the 82 tools that received unanimous ratings from the experts were further examined. In addition, 20 new and 20 freshly reconditioned end mills were included.

[0178] Of the final 122 tools, a shell image was taken for each of them using the image recording device described above, which was used for machine learning.

[0179] For supervised learning, the Hough Transformation (HT) and the Fast Fourier Transformation (FFT) were used.

[0180] For the unsupervised concepts, the explained clustering algorithms k-Means++ and Gaussian Mixture Models (GMM) were compared.

[0181] Two models were trained for each approach (supervised and unsupervised): one with tools of wear class 3 and one without. The omission of wear class 3 arises due to the problem that this class contains many tools with large breakages, which corresponds to high wear that is not reflected in the surface texture (or only very locally).

[0182] The dataset was manually divided into training and test datasets. Each classifier was thus based on identical test and training data. This allowed the supervised and unsupervised learning classifiers to be directly compared. An 80% to 20% split was used. For each wear class, 1 / 5 of the available data was shifted to the test dataset. This resulted in the following split:

[0183] For the evaluation, the metric Accuracy was used in particular, which looks at the ratio of correctly classified tools to the total number of predictions. It is defined as follows: where the parameters mean the following:

[0184] TP true positives;

[0185] TN true negatives;

[0186] FP false positives; FN false negatives.

[0187] Overall, the accuracy of the test data sets is as follows: Overall, unsupervised learning with GMM clustering yields the best results, with the accuracy being significantly better when omitting wear class 3. An extension of the method thus makes it possible to detect large local defects, such as tooth breakage, using other methods (e.g., appropriate image processing) and directly assign these tools to wear class 3. The remaining defects can then be further analyzed using the described method.

[0188] The described method for classification using unsupervised machine learning offers the further advantage that, due to the local classification of many individual image sections along with the wear class, a more differentiated picture of the tool's wear can be obtained. For example, a histogram of the assignment of individual image sections to wear classes clearly shows the proportion of the highest wear class. Another interesting metric is the sum of the wear classes of the data points of a tool.

[0189] Figure 9 is a schematic block diagram of a system according to the invention for determining the wear condition and for preparing a tool.

[0190] The system comprises a processing machine 101, e.g., a milling machine, which is arranged in a first work station 100. The processing machine comprises, in a manner known per se, (at least) one work spindle 102, a tool magazine 103, and a transfer device 104 with a tool holder, by means of which tools 2 can be exchanged between the work spindle 102 and the tool magazine 103. In the example described, the tools are solid-material end mills with helical main cutting edges on the casing and straight secondary cutting edges on the face of the tool 2. The transfer device 104 also enables the removal of a tool 2, wherein the tool 2 is moved to a removal position 105. The tool 2 can also be transferred to the image recording device 1 described above.

[0191] The data recorded by the image recording device 1 are transmitted to a processing unit 110. This is a computer on which a first image processing module 111, a second image processing module 112, and a classification module 113 are implemented in software. The image processing module 111 receives the data from the image recording device 1 and processes it, as described above, into an image with surface structure information.

[0192] This image is fed to the second image processing module 112, where it is further processed, namely cropped and filtered as described above. The correspondingly preprocessed images are fed to the classification module 113, which assigns a wear class ("reusable," "to be reconditioned," "to be disposed of") to the imaged tool.

[0193] In this way, each tool is checked for its wear status after being removed from the work spindle. It may be useful to perform a cleaning step before the test so that the measurements are not affected by adhering dust or chips. For this purpose, a cleaning device, e.g., with a liquid or air nozzle, can be used. If the wear status allows further use, the tool is stored in the tool magazine 103. If reconditioning is necessary or the tool is to be disposed of or recycled, it is moved to the removal position 105. At the same time, the classification result is displayed. Data on the tool 2 to be reconditioned, along with a unique identifier for the tool, is stored in a central database 120. The central identifier is also noted on the tool 2, e.g., visually or electronically.

[0194] If tool 2 is to be reconditioned, it is sent to a reconditioning facility 150 in the usual way. There, the identifier is first read out using a reader 151, e.g., using a camera or an RFID reader and downstream electronics. A controller 152 then retrieves the data on tool 2 from database 120 based on the identifier. The reconditioning machine, e.g., a grinding machine 153, is then controlled based on the retrieved data. The data includes, for example, information on the areas to be reconditioned (face, surface; specific information on the cutting edges or cutting regions) and / or information on the current geometry of the tool. This allows reconditioning to be carried out efficiently and effectively without further data acquisition. Information on the completed reconditioning is, in turn, stored in database 120, assigned to the tool identifier.

[0195] After reconditioning, the tool 2 is returned to the factory 100 (or to another factory). There, it can be used again. The invention is not limited to the illustrated embodiments. Specifically, a different source of surface images can be used, and the preprocessing and / or classification can be performed in a different way. Furthermore, the image recording device and the processing unit can be arranged and operated independently of a specific processing machine. In summary, the invention provides a method for determining the condition of a tool, which can automatically and reliably detect the condition of a tool.

Claims

Patent claims 1. A method for determining a condition of a tool, comprising the following steps: a) taking a plurality of optical images of a surface of the tool under different illumination conditions; b) processing image data from the plurality of optical images to generate an image with surface structure information; c) preprocessing the image with surface structure information to generate one or more preprocessed images; d) classifying the condition of the tool into one of at least two classes based on the one or more preprocessed images by means of a machine learning method.

2. Method according to claim 1, characterized in that the processing of the image data is carried out by means of photometric stereo analysis.

3. Method according to claim 2, characterized in that the image with surface structure information is a height image and / or a curvature image.

4. Method according to one of claims 1 to 3, characterized in that the pre-processing step comprises cutting to a region of interest, the region of interest comprising those regions of the surface of the tool which are affected by wear or where defects are to be expected.

5. Method according to one of claims 1 to 4, characterized in that the machine learning method is a supervised machine learning method.

6. Method according to claim 5, characterized in that for generating the one or more preprocessed images in the image with surface structure information Wear or defect features are isolated, in particular by at least one of the following methods: a) applying a Hough transform to the image with surface structure information; b) applying a Fast Fourier transform to the image with surface structure information.

7. The method according to claim 5 or 6, characterized in that particles with pixels in a predetermined brightness range are identified in the pre-processed image and, based on the identified particles, several of the following quantities are determined as features for the machine learning method: a) area of the particles; b) filled area of the particles; c) equivalent surface diameter of the particles; d) area of a bounding box of the particles; e) major axis length of the particles; f) maximum Feret diameter of the particles; g) convex area of the particles; h) minor axis length of the particles; i) circumferential radius of the particles; j) mean value of the pixels in the brightness range; k) standard deviation of the pixels in the brightness range; l) number of pixels in the brightness range; m) number of particles.

8. Method according to claim 7, characterized in that the following variables are determined as features for the machine learning method: a) area of the particles; c) equivalent surface diameter of the particles; e) major axis length of the particles; h) minor axis length of the particles; j) mean value of the pixels in the brightness range and m) number of particles.

9. Method according to one of claims 1 to 8, characterized in that the machine learning method is an unsupervised machine learning method, wherein in a first step a plurality of features is determined on the basis of the preprocessed image or the plurality of preprocessed images and in a second step a clustering is carried out.

10. The method according to claim 9, characterized in that, in order to generate the plurality of preprocessed images, the image with surface structure information is divided into a plurality of overlapping tiles using a sliding window method, wherein a plurality of predetermined features for clustering are determined for each preprocessed image corresponding to a tile. 1 1. Method according to claim 10, characterized in that the features from the preprocessed images comprise several from the following groups: a) a mean, a variance and / or a median of the preprocessed image further processed by means of a two-dimensional Gabor filter; b) a sum of a binary image obtained from the preprocessed image further processed by means of the two-dimensional Gabor filter; c) features based on a local binary pattern analysis; d) Haralick texture features; e) a mean and / or standard deviation of pixel values of the preprocessed image.

12. Method according to one of claims 9 to 11, characterized in that the clustering is carried out in 3-8, in particular in 4-6, classes.

13. Method according to one of claims 9 to 12, characterized in that the clustering is carried out by means of a Gaussian mixture model or a k-means++ method.

14. Method according to one of claims 1 to 13, characterized in that a broken cutting edge of the tool is detected on the basis of the image data of the plurality of optical images and / or on the basis of the image with surface structure information and / or on the basis of the one or more cropped images with surface structure information, and that if a broken cutting edge is detected, the tool is immediately assigned to a corresponding condition class.

15. Method according to one of claims 1 to 14, characterized in that the method for determining the condition is applied to a grinding or drilling tool for workpiece machining in a machine tool, in particular to a solid material shank tool.

16. The method according to claim 15, characterized in that a first state of a shell-side cutting geometry and a second state of a front-side cutting geometry of the solid material shank tool are determined separately.

17. The method according to claim 16, characterized in that an algorithm based on a first data set is used to determine the first state and that an algorithm based on a second data set is used to determine the second state, wherein the first data set and the second data set are different.

18. Method for reconditioning a tool, comprising the following steps: a) Determining the condition of a tool using a method according to one of claims 1 to 17; b) Controlling at least one device for preparing the tool, in particular by means of a grinding process, when the condition meets predetermined conditions.

19. The method according to claim 18, characterized in that an illumination system for illuminating a surface of the tool with different illumination states and a camera for recording a plurality of optical images of the surface of the tool in the different illumination states are arranged at a first location, that data obtained at the first location are stored in a database, that the device for processing is arranged at a second location and that the device for processing retrieves data from the database.

20. A device for determining the condition of a tool, comprising: a) an illumination system for illuminating a surface of the tool with different illumination conditions; b) a camera for recording a plurality of optical images of the surface of the tool under the different illumination conditions; c) a first image processing module configured to process image data from the plurality of optical images to generate an image with surface structure information; d) a second image processing module configured to further process the image with surface structure information to generate one or more preprocessed images; and d) a classification module configured to classify the condition of the tool into one of at least two classes using a machine learning method based on the one or more preprocessed images.

21. Device according to claim 20, characterized in that the lighting system and the camera are integrated into a processing machine with a holder for the tool, in particular such that the lighting system can illuminate the tool and the camera can record the optical images of the surface of the tool when the tool is held in the holder.

22. An arrangement comprising: a) a device for determining the condition of a tool according to claim 20 or 21; b) a device for reconditioning the tool, in particular by means of a grinding process; c) a controller for controlling the reconditioning device, which is configured to receive information about the condition from the device for determining and to control the device for reconditioning the tool depending on the information received.

Images