Method for determining a tool state

Abstract

In a method for determining a tool state, first a plurality of optical images of a tool surface is recorded in different illumination states. Subsequently, image data of the plurality of optical images is processed to generate an image with surface structure information. This image is then pre-processed to generate one or more pre-processed images. Finally, based on the one or more pre-processed images, the tool state is classified into one of at least two classes by means of a machine learning method.

Description

Technical Field

[0001] This invention relates to a method for determining the state of a tool. The invention also relates to an apparatus for determining the state of a tool and a facility having such an apparatus. Background Technology

[0002] Many tools, such as those used for machining materials like milling cutters and drills, wear down due to their interaction with the workpiece. This can lead to deterioration in machining results, excessive machining times, and / or irreparable damage to the tool (or even the workpiece). This is especially true for tools used in automated machines, so it is crucial to remove the tool from the machine promptly before these effects occur. If possible, the tool should then be refurbished for further use, for example, by replacing the same machining element or by re-refurbishing it (e.g., regrinding). If refurbishment is generally not (or is no longer) possible, or is impossible due to its wear condition, the tool should be sent to a recycling process or disposed of.

[0003] Tools can be removed and replaced after a predetermined duration or a predetermined number of machining cycles, where the duration or number is chosen such that, for the corresponding tool type, no negative impact is expected before this point, even under the "worst-case" scenario. Therefore, tools are often removed, replaced, and refurbished prematurely. The effective lifespan of the tools is correspondingly shortened, and the number of tool replacements and refurbishments exceeds the actual required number.

[0004] Alternatively, the operator evaluates the machining results on the tool and / or workpiece, depending on the tool, using the naked eye or with the aid of auxiliary tools (such as a magnifying glass, microscope, or measuring instrument). The operator then decides whether the tool can be used further.

[0005] This assessment is labor-intensive, often requiring the shutdown of the relevant machines and the removal of tools from them. Inconsistencies can arise, particularly when the assessment is conducted by different personnel.

[0006] A system for automatic wear identification has been proposed. Therefore, CN 108107838 A (Shandong University) relates to wear identification of cutting tools. To this end, a cloud-based knowledge database with wear data was established, and a detection model was trained based on a support vector machine (SVM). The data is continuously updated to improve identification.

[0007] US 7,479,056 B2 (Kycera Tycom) describes an entirely automated system for checking the identity and geometry of a drilling tool, refurbishing the tool, checking based on predetermined tolerances, setting a positioning ring at the tool shank and cleaning and packaging the refurbished tool. The checking of the geometry takes place using optical units. These optical units each comprise a camera on the head side and on the front side for imaging the front surface and the peripheral surface of the tool. The data generated at the time of the tool check can be stored in a controller. Based on the recorded images, predetermined reference points are identified and the distances between these reference points are measured. In addition, a preliminary assessment of the cutting edge condition is also made.

[0008] US 7,479,056 B2 does not disclose more details about the evaluation of the image data. The preliminary assessment is made based on the outer geometry of the tool and the condition of the cutting edges, where it is not clear how this can be determined or classified.

[0009] EP 3 881 968 A1 (Fraisa SA, GF Machining Solutions AG) discloses a method for determining the wear state of a tool, in which the image data of at least one optical image are processed to identify a wear region. Here, the planar and / or spatial extent of the wear region is determined and based on the determined extent, the wear state of the tool is classified.

[0010] This method allows an automatic classification. However, it has been shown, especially for solid material shank tools, that with this method a reliable classification is not achieved in all cases. SUMMARY

[0011] It is the task of the present invention to provide a method which belongs to the technical field mentioned at the outset, which allows an automatic and reliable identification of the state of a tool.

[0012] The solution to this task is defined by the features of claim 1. According to the invention, the method for determining the state of a tool comprises the following steps: a) recording a plurality of optical images of the surface of the tool in different illumination states; b) processing the image data of the plurality of optical images to generate an image with surface structure information; c) pre-processing the image with surface structure information to generate one or more pre-processed images; d) classifying the state of the tool into one of at least two classes by means of a machine learning method based on the one or more pre-processed images.

[0013] The processing of the image data, the pre-processing of the image with surface structure information and the classification of the state take place in a computer-aided manner.

[0014] The different illumination states are distinguished in particular by different illumination directions.

[0015] The optical image is a recording in the visible range of the spectrum or in adjacent wavelength ranges (IR and UV). Typically, the tool or a region of interest thereof is illuminated and the light reflected at the surface is detected with a suitable device (camera; imaging optics with image sensor). The light used for illumination can have a continuous spectrum or a spectrum composed of a plurality of spectral lines or bands, or can be monochromatic. The camera can also detect a broad band, one or more narrow bands or specific frequencies; it can also deliver monochromatic or polychromatic information. The grey-scale image is then preferably further processed.

[0016] The image data preferably comprise a two-dimensional image of the surface, and the image data corresponding to the two-dimensional image are used to generate an image with surface structure information. Effective methods are known for this purpose.

[0017] The optical image does not have to image the entire tool surface. In principle, it is sufficient to image those regions of the tool surface where wear or failure is expected (for example the cutting edge of a rotating tool). For example, the shank of a shank tool or the radially inner region of a grinding or cutting disc can also be excluded from the image. If uniform wear is expected, it can be sufficient to detect only a representative region. If point-like wear (for example chipping) is expected, it is often meaningful to optically detect all potentially affected regions, so that, for example, when the plurality of optical images are combined, the cutting region that interacts with the workpiece is completely detected.

[0018] The image with surface structure information is in particular a two-dimensional grey-scale image.

[0019] The classification into one of at least two classes can comprise the assignment of the tool to a wear level, for example "new (as new)", "low wear", "medium wear" and "high wear". The classification can also be directed at measures to be taken, for example "continue use", "refurbish", "dispose". Further classifications are also possible, for example assignment to one of a plurality of possible refurbishment methods (cleaning, polishing, grinding, recoating, etc.), specification of a point in time for the next inspection or specification of application areas in which the tool can still be used - for example involving less easily processable materials or methods with lower precision requirements.

[0020] The classification can also serve to distinguish between a defect-free new tool and a new tool with production and manufacturing defects (for example coating defects or microchipping). In this case, a division into two classes can be sufficient. Depending on the result, the tool with production or manufacturing defects is directly refurbished or recycled or disposed of. With the aid of the method according to the application, it is thereby possible to systematize and automate quality assurance.

[0021] This classification can be combined with a wear classification, so that the state of the tool can be assessed over the entire lifetime of the tool, starting from the mentioned manufacturing defects, through initial use traces resulting from the first contact with the workpiece, up to the obvious damages affecting the machining quality. These damages are in particular typical wear traces, such as known from free surface wear, bulk wear, chipping and cracks, as well as adhesions in the form of built-up edge.

[0022] The classification can also directly include the generation of parameters for a subsequent tool refurbishment, where in this case the resulting number of classes (set with certain parameter combinations) is much greater. However, alternatively, the parameters for the refurbishment are only generated after the classification based on the original image data and / or the pre-processed images.

[0023] It is also possible to separately classify the state of different regions of the tool, such as the tool nose and the tool peripheral surface, so that for example only certain regions of the tool need to be refurbished.

[0024] Depending on the tool and the application, the wear can manifest itself in different ways. For example, certain dimensions are reduced due to material removal at the tool, deformations occur which lead to a change in the geometry, or individual regions have wear traces at the surface and / or to a certain depth, compared to a usable tool which has not been used.

[0025] The method according to the invention can be used in combination with a large number of tools, in particular with those which mechanically act on the workpiece or are worn by the interaction with the workpiece. These tools include in particular tools for chip-removing material machining, i.e. both rotating tools for milling, drilling or thread cutting, but also stationary or linearly moving tools, such as rotary chisels, punching tools or saws.

[0026] The state determination can be carried out fully automatically, in addition to a manual assessment, it also provides an objective image, since all tools are assessed in the same way. Due to the automation, the state determination can be carried out regularly, so that on the one hand defects in the machining result caused by excessively worn tools can be prevented, and on the other hand tools which are actually still usable are not refurbished, disposed of or recycled too early. The use of machine learning methods allows a simple focus on new tools or new tool types. The state determination can be carried out in a very short time compared to an assessment by means of laboratory equipment, which has a great advantage when used in a refurbishment center, where many tools to be refurbished have to be assessed.

[0027] The image data is preferably processed by means of photometric stereo analysis. Here, this is a well-known shadow-based surface reconstruction method based on two or more images of a surface. Here, the images are generated with the same arrangement of surface and camera and only the illumination state, here the illumination direction, is changed. This makes it possible for a position on the observed surface to correspond simply to an image point in the image. A plurality of irradiance values is thus determined for each image point. From these irradiance values, the surface direction is 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, these images can be further processed by means of an integration method, for example in order 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 state can be based on the images with surface structure information, and the corresponding pre-processing does not include an integration step.

[0028] The photometric stereo analysis method is thus not used for the usual purpose of three-dimensional reproduction of the surface, but rather for generating images containing surface structure information, which enables improved recognition of wear phenomena.

[0029] Preferably, the images with surface structure information are height images and / or curvature images. In a height image, each pixel encodes the relative height of the corresponding point on the surface with respect to a reference plane. In a curvature image, each pixel encodes the local curvature. It has been shown that bulges and depressions occurring in the use of tools, in particular grinding and drilling tools, are particularly easy to see in height images. Curvature images enable particularly good identification of small local defects.

[0030] In principle, the classification can be based on a plurality of images with surface structure information, which can be obtained from a plurality of optical images in different ways, for example height images and curvature images. This makes it possible to combine the advantages of the individual image types.

[0031] Alternatively or additionally, other image types can be used, for example image types which can be generated in the context of photometric stereo analysis, for example oblique images in different directions, which enable good identification of shape deviations, for example of the cutting edge. The use of texture images in addition can help to identify surface rust damage or discoloration, which can thus be distinguished from actual damage.

[0032] Advantageously, the pre-processing step includes cropping to a region of interest, wherein the region of interest includes those regions of the tool surface which are affected by wear or in which a failure is expected. This makes it possible to reduce the processing overhead and to improve the classification quality.

[0033] In the wear determination of solid material shank tools, in particular the end region of the cutting edge and the shank end are cut off at the time of the cut. In another embodiment, regions which are clearly outside the region of the cutting edge in which wear is expected to occur can also be disregarded.

[0034] Alternatively, such a cut is omitted. In principle, the machine learning method can be trained or the features of the cluster can be selected such that regions which are not affected by wear or failure do not lead to a distortion of the classification, i.e. differences in these regions have no influence on the classification of the tool.

[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 so-called support vector machines (SVM) or on artificial neural networks (ANN).

[0036] Within this type of method, in order to generate the one or more preprocessed images, it is advantageous to isolate wear features or failure features in the image with surface structure information. For this purpose, in particular straight lines corresponding to the cutting edges can be identified in the image and subtracted from the image, so that only structures deviating from the straight lines remain in the preprocessed image, which in particular correspond to wear features or failures. Similarly, in particular a distinctive frequency can be identified, for example, due to the constant spacing between adjacent cutting edges, and structures corresponding to the identified frequency can be subtracted. In this way, too, mainly wear features or failure 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, within the Hough transform, as described in detail below, straight lines corresponding, for example, to the tool cutting edges are identified in the image with surface structure information, and then the corresponding image values are subtracted from the image with surface structure information. The image information corresponding to the wear features remains.

[0039] By means of the fast Fourier transform applied to the image with surface structure information, a characteristic frequency is determined due to the repeatedly occurring parallel cutting edges in the image with surface structure information. Features occurring in the image data at this frequency are with a high probability associated with undamaged structures of the tool surface, for example cutting edges. In contrast, features not having this frequency are very likely to be caused by defects. Therefore, in order to obtain a preprocessed image, the regions in the frequency space corresponding to these frequencies can be masked. The subsequent inverse transformation into the position space provides an image of the defects.

[0040] Advantageously, particles having pixels in a predefined brightness range are identified in the pre-processed image and the following plurality of quantities is determined from the identified particles as features of the machine learning method: a) the area of the particles; b) the filled area of the particles; c) the equivalent area diameter of the particles; d) the bounding box area of the particles; e) the major axis length of the particles; f) the maximum Feret diameter of the particles; g) the convex hull area of the particles; h) the minor axis length of the particles; i) the circumference of the particles; j) the mean value of the pixels in the brightness range; k) the standard deviation of the pixels in the brightness range; l) the number of pixels in the brightness range; m) the number of particles.

[0041] Particles are contiguous regions of bright pixels in a binary image generated from the pre-processed image (or similarly, contiguous regions of pixels with a brightness value exceeding a threshold in a grayscale or color image). It has been shown that these features are well suited to characterize wear effects at tool surfaces (especially tool surfaces made of metallic or ceramic materials) based on pre-processed optical images.

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

[0043] These features are suited to represent state-related properties of the pre-processed image.

[0044] The values of the selected features are preferably normalized. Specific weighting can be applied to the individual features in order to consciously take into account different relevance of the features in the classification. This weighting can be performed after normalization or can be obtained by using a criterion that depends on the features.

[0045] In a second group of embodiments, the machine learning method is an unsupervised machine learning method, wherein in a first step a plurality of features is determined based on the pre-processed image or a plurality of pre-processed images and in a second step a clustering is performed.

[0046] Within the scope of such a method, in order to generate a plurality of pre-processed images, the image with surface structure information is divided into a plurality of overlapping tiles, preferably by means of a sliding window method, wherein for each pre-processed image corresponding to a tile a plurality of 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 evaluated. In particular, the image of the tool surface is divided into at least 4 tiles, in particular at least 10 tiles.

[0047] The division into tiles enables a clustering which takes into account local defects. In addition, from the assignment of the individual tiles to different clusters, in addition to the classification, supplementary information about the wear of the tool to be evaluated can be obtained - corresponding to the respective local defects.

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

[0049] Preferably, the values of the selected features are normalized. Individual features can be specifically weighted in order to purposefully take into account different relevance of the features in the classification. This weighting can take place after normalization or can be obtained by using a feature-dependent criterion.

[0050] The two-dimensional Gabor filter can filter the image segments on the basis of a predefined kernel. Alternatively, a so-called filter bank is created, in which there are different kernels with different frequencies and different directions. The pre-processed image is filtered with all filters of the filter bank. Finally, the output images of all filters are combined into a single combined image (superposition of the filters).

[0051] Other or further features are also possible. Thus, it can be advantageous if, in addition to the local binary pattern (LBP), in particular, a histogram of oriented gradients (HOG) is also used to obtain features.

[0052] The clustering is preferably divided into 3-8, in particular 4-6, classes. It has been shown that with a suitable number of classes a sufficiently fine division can be achieved in order to distinguish different tools with regard to the measures to be taken (e.g. refurbishment, continued use, disposal). At the same time, the classification is already robust even if the number of tools to be evaluated is relatively small or the training set is relatively small.

[0053] In certain cases it can be advantageous to provide more than 8 classes, for example when specific refurbishment parameters are to be assigned to the tools.

[0054] Preferably, the clustering is carried out by means of a Gaussian Mixture Model (GMM) or a k-means++ method.

[0055] On the basis of the clustering, the entire tool can be assigned a state class corresponding to the worst state class identified in the clustering, i.e. the state class is determined in particular by the state of the most worn region (tile) of the tool.

[0056] Since each tile is evaluated individually, it can be attributed and assigned to the corresponding overall image. From this, information is derived about how often (in percent) which condition class occurs on the corresponding tool. Thus, for example, a percentage of the tool surface with respect to the as-new state can be given. From this, an indication about the overall state of the tool is given. It is also possible to spatially allocate wear or failure points.

[0057] Advantageously, on the basis of the image data of the plurality of optical images and / or on the basis of the images with surface structure information and / or on the basis of the one or more cutouts with surface structure information, a broken cutting edge of the tool is identified, and in the event of an identified broken cutting edge, the tool is directly assigned to the corresponding state class.

[0058] In this case, the tool is thus not fed to the classification by means of the machine learning method. It has been shown that a broken tooth in an optical image of the tool surface leads to features which are significantly different from other optically detectable wear or failure features. Thus, on the one hand, tools with a broken tooth are included which have a negative influence on the classification by means of the machine learning method. But on the other hand, a broken tooth can also be easily identified with conventional image processing methods (or a dedicated upstream machine learning method) and the corresponding tool is directly assigned to the corresponding state class.

[0059] The state determination method can be applied in particular to grinding or drilling tools which cut into a workpiece in a machine tool, in particular to solid material shank tools.

[0060] Advantageously, a first state of the peripheral surface side cutting geometry and a second state of the end surface side cutting geometry of the solid material shank tool are determined individually.

[0061] Different geometrical relationships and different requirements for the integrity of the respective cutting edge can thus be taken into account. For example, for a milling tool, the end face side cutting edges are generally subjected to less load than the peripheral side cutting edges, and vice versa.

[0062] For the classification of the "overall state", the first state and the second state can be obtained differently and / or used differently. In the simplest case, the tool is replaced if at least one of the two states requires it, and the tool is refurbished if at least one of the two states requires it. If the overall performance of the tool does not simply correspond to the weakest link, but is produced by the joint action of the performance of the two tool segments, it can be advantageous to counteract the states to one another in a more complex manner, so that a refurbishment or replacement is only carried out when the overall performance requires it.

[0063] Advantageously, for determining the first state, an algorithm based on a first data set is used, and for determining the second state, an algorithm based on a second data set is used, wherein the first data set and the second data set are different.

[0064] In particular, the two data sets are essentially disjoint. Thus, for example, in a supervised learning method, the first data set comprises image data as training data, which shows peripheral surface regions of a worn tool; the second data set comprises image data, which shows end face regions of a worn tool. An overlap of the first data set and the second data set here exists at most in the transition region (edge or rounding) between the peripheral surface and the end face.

[0065] Not only can the data sets be used differently, but also the algorithms, for example different machine learning algorithms or algorithms with other parameters (for example network topology in neural networks) can be used.

[0066] A method for refurbishing a tool preferably comprises the following steps: a) determining the state of the tool using the method according to the application described; b) if the state meets predetermined conditions, controlling at least one device for refurbishing the tool, in particular by means of a grinding process.

[0067] The predetermined conditions in particular comprise (also) the classification of the tool state. Thus, the classes can be defined from the outset in such a way that they correspond to the measures to be carried out ("continue to use", "refurbish", "recycle") or the measures are directly or indirectly derived from the classification. For example, a wear state can be classified into eight levels 1-8 (1 : new, 8: heavily worn), wherein the conditions are predetermined in such a way that the tool is continued to be used in the case of a classification into levels 1 and 2, refurbished in the case of a classification into levels 3-6, and recycled or disposed of in the case of a classification into levels 7 and 8.

[0068] The status can not only serve as a basis for the decision whether to carry out a refurbishment using the respective device, but also in relation to the measures to be carried out within the scope of the refurbishment. For example, a plurality of refurbishment steps (cleaning, polishing, grinding, a plurality of grinding processes, etc.) can be selected and different selections made depending on the status.

[0069] In particular, parameters of the grinding process and / or machining geometry of the device used for the refurbishment of the tool are determined on the basis of the determined status. The parameters can comprise, for example, a grinding path, one or more feed values, etc. The machining geometry defines the scope, position and type of the tool machining. The determination of the machining geometry can be based directly on the optical images and / or on the processed results, for example the classified processed results.

[0070] This achieves a tool-specific, on-demand refurbishment. In the grinding method, for example, the (additional) material removal can be minimized, thus maximizing the service life of the tool. Furthermore, it is not necessary to carry out a complex examination of the tool (again) before the refurbishment, since the required data already exists.

[0071] In a preferred embodiment, the illumination system for illuminating the tool surface in different illumination states and the camera arrangement for recording a plurality of optical images of the tool surface in different illumination states are arranged at a first place of use. The data obtained at the first place of use is stored in a database. The device for the refurbishment is arranged at a second place of use, and the device for the refurbishment retrieves the data from the database.

[0072] Here, the two places of use are remote from one another and in particular located in another device or another factory. The database can be arranged centrally, such that the data is detected and used in a decentralized manner, but stored centrally. However, the database can also be stored at the first place of use or the second place of use, or the data is kept available at both places and is synchronized periodically.

[0073] This in particular also enables the complete detection and provision of data relating to a specific tool, even if the tool is refurbished by different users or even by different service providers after the respective performance of the refurbishment.

[0074] Preferably, the tool is provided with a unique identifier, and the data assigned to the tool is associated with this unique identifier in the database. The unique identifier is attached to the tool in machine-readable form, for example as an optical marking (bar code, matrix code, in alphanumeric, etc.) or stored on a data carrier (for example RFID). The unique identifier ensures the correct assignment.

[0075] Alternatively, the detected data is stored on a data carrier. Subsequently, the data carrier is transported together with the respective tool from the first place of use to the second place of use. In principle, the data carrier can also be integrated in the tool.

[0076] In another case, the recording of the at least one optical image is carried out directly at the device for the refurbishment.

[0077] The device for determining the tool status according to the application comprises a) an illumination system for illuminating the tool surface in different illumination states; b) a camera for recording a plurality of optical images of the tool surface in different illumination states; c) a first image processing module configured to process the image data of 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 tool status into one of at least two classes by means of a machine learning method based on the one or more preprocessed images.

[0078] The camera is in particular a line camera. With it, the surface of the tool can be detected line by line. In order to detect the peripheral surface of a solid material shank tool, the solid material shank tool is preferably rotated step by step about its longitudinal axis by means of a spindle and the surface is recorded line by line by means of a fixedly arranged line camera. At each rotational position, the tool is illuminated in turn in different illumination states. Thereby, in particular an image data corresponding to an unrolling of the tool periphery is obtained, wherein initially each line occurs multiple times - corresponding to the different illumination states. However, from these image data, a plurality of images showing the unrolling in different illumination states can be simply generated by combining the lines corresponding to the number of illumination states.

[0079] In a further preferred embodiment, the illumination system and the camera are integrated in a machining machine having a tool receptacle, in particular such that the illumination system can illuminate the tool and the camera can record optical images of the tool surface when the tool is accommodated in the receptacle.

[0080] Here, the image processing module and the classification module can be accommodated in a processing device, wherein the processing device is wholly or partially contained in the machining machine or is arranged outside the machining machine and is in signal connection therewith.

[0081] The machining machine can be for example a machine tool or a machining center for drilling or milling. The receptacle can be for example a work spindle, a receptacle in a tool magazine for storing tools for tool change, or a transfer receptacle for transferring tools bi- directionally between a work spindle and a tool magazine.

[0082] It is therefore also advantageous for a machining machine to comprise a camera and to be connected with a processing device or to integrally or partially contain a processing device, wherein the processing device comprises two image processing modules and a classification module.

[0083] The machine tool device according to the application comprises a machine tool, preferably a machining center, a milling or drilling center, and a device for determining a state according to the application. Here, the camera is integrated in the machine tool or arranged thereat. The image processing modules and the classification module are accommodated in a processing device. Here, the processing device is integrally or partially contained in the machine tool or arranged outside of it and in signal connection therewith.

[0084] An advantageous device for refurbishing a tool comprises a) a device for determining a state of a tool according to the application; b) a device for refurbishing a tool, in particular by means of a grinding process; and c) a controller for controlling the device for refurbishing, which is configured in such a way that it receives information about the state from the determining device and controls the device for refurbishing a tool in accordance with the received information.

[0085] Thus, it is preferred to assign a camera to a machining machine, by means of which the wear state of a tool used in the machine can be monitored periodically, for example at each tool change. The image data are further processed directly at the location of the machining machine according to the application, so that it can be decided on the basis of the classification whether the tool can continue to be used. If this is the case, it is stored in the tool magazine. Otherwise, it is discarded, and the data about the wear state, and optionally the image data or further information obtained therefrom, are stored in a database. The discarded tool is then physically transported to a refurbishing device. The refurbishing device reads the data assigned to the tool from the database and controls the refurbishing device on the basis of said data. The refurbished tool is then transported to the same machining machine or to another machining machine and used there again.

[0086] Further data can also be used in the refurbishing, for example data about the history of the tool (number of use cycles, previous refurbishing, etc.) or data about the requirements of the customer for its particular machining process.

[0087] Further advantageous embodiments and combinations of features of the application result from the following detailed description and the claims in their entirety. BRIEF DESCRIPTION OF DRAWINGS

[0088] The drawings used to explain the embodiments show: Figure 1A , B: schematic front view and schematic side view of one embodiment of an image recording device for use in a method for determining the wear state of a shank tool according to the application; Figure 2 : Synthetic image of the flank of a tool; Figure 3 : Height image of the flank of an end mill with different wear; Figure 4 : Schematic illustration of the cropping of the height image to a region of interest (ROI); Figure 5 : Illustration of the pre-processing of the height image by means of the Hough transformation; Figure 6 : Corresponding pre-processed image as basis for supervised machine learning classification; Figure 7 : Pre-processed height image by means of the Fast Fourier Transformation as basis for supervised machine learning classification; Figure 8A E: Result of the clustering by means of the k-means++ algorithm; and Figure 9 : Schematic block diagram of a facility for determining the wear state and refurbished tools according to the invention.

[0089] In the figures, in principle, identical parts are provided with identical reference signs. DETAILED DESCRIPTION

[0090] The implementation of the method according to the invention is explained below on the basis of an example of determining the wear of the main cutting edge of a solid material end mill.

[0091] In a solid material end mill, in particular, two types of wear are to be distinguished: - Free surface wear: This is the wear in the area of the free surface adjacent to the cutting edge. Excessive cutting speeds, low wear resistance of the cutting edge or insufficient supply of cooling fluid promote this type of wear.

[0092] - Cutting edge chipping: This can occur at the tool tip and at the free surface. It is promoted if the tool is used further with severe wear or if the tool is used to machine hard workpieces.

[0093] In principle, the method according to the invention detects both types of wear. However, as explained in more detail below, the result can be improved if, in an upstream step, the cutting edge chipping is identified, for example by means of image processing methods known per se, and then tools without cutting edge chipping are analyzed using the method for the classification of the free surface wear.

[0094] Figure 1A 、 1BA schematic front view and a schematic side view showing one embodiment of an image recording device 1 for use in a method for determining the wear state of a shank tool according to the present application. The respective device is known in principle and commercially available, for example from the DOME series of devices from SAC Sirius Advanced Cybernetics GmbH, Karlsruhe, Germany.

[0095] With the image recording device 1, first an optical recording of the peripheral surface of the end mill 2 is generated, in particular of the peripheral surface working area (without shank). For this purpose, the end mill 2 is clamped with its shank in the chuck 3 of a vertically arranged main spindle 4. The mill is illuminated with a dome illumination 5. The dome illumination comprises a dome 6 with a diffusely reflecting concave inner surface facing the main spindle with the end mill 2. The dome 6 is surrounded by a black base plate 7. A LED light strip 8 is arranged at the base plate and extends around the edge of the dome 6, which comprises 8 segment-like illumination elements 8.1...8.8, which can be selectively actuated to produce different illumination directions. The dome 6 is additionally arranged on a straight linear guide 9 extending radially with respect to the end mill 2, so that the focal point can be set for the respective measurement object.

[0096] A linear array camera 10 is arranged behind the dome 6 and can optically detect the end mill 2 peripheral surface area facing the dome 6 through a vertical slit in the dome. In the context of the present embodiment, this is a commercially available 8K monochrome linear array camera. This 8K monochrome linear array camera records a line of 8192 grey scale pixels each time it records. The objective used is a precision objective with a focal length of 105 mm and an aperture of f / 5.6.

[0097] The main spindle 4 with the chuck 3 can be precisely rotated around its vertical axis by means of an electric motor with an encoder, so that the entire peripheral surface can be detected with the linear array camera 10 in one full rotation. At each rotational position, the linear array camera 10 records one pixel line each. For each image recording, the linear array camera 10 and the illumination segments 8.1...8 are triggered jointly.

[0098] The end mill 2 to be inspected is clamped in the chuck 3 of the above-described image recording device 1. The distance between the dome 6 and the end mill 2 is set so that the best possible illumination of the peripheral surface area of the tool to be inspected is achieved and so that the linear array camera 10 arranged behind the dome 6 can be focused on this area. The tool is rotated stepwise around its longitudinal axis and at each rotational position four images of different illumination directions are recorded each. In order to change the illumination direction, four of the eight illumination elements 8.1...8.8 of the LED light strip 8 arranged uniformly on the circumference of the dome 6 are selectively actuated: image lighting section Lighting section angle (measured from directly above) direction 1 8.8 292.5 - 337.5° Top left 2 8.2 22.5 - 67.5° Top right 3 8.4 112.5 - 157.5° Bottom right 4 8.6 202.5 - 247.5° Bottom left .

[0099] The device makes it possible to record the entire circumference in a few seconds. Thereby a high throughput of tools to be evaluated is achieved.

[0100] Since the relative arrangement between camera and tool does not change between the image recordings in different illumination directions, the image points of the individual images of the same rotational position can uniquely correspond to each other and a plurality of irradiance values can easily be assigned to each image point of the detected line. The line-by-line recording is carried out until the entire circumference is scanned, i.e. the end mill 2 has rotated 360° in the chuck 3. The individual pixel rows corresponding to the same illumination are then respectively stitched into one image. Thereby four images of the circumference of the tool are obtained. From these images a composite image is now derived: a) a tilt image in the X direction; b) a tilt image in the Y direction; c) a curvature image; d) a height image; e) a texture image.

[0101] The above-mentioned commercially available image processing software Coake® of SAC Sirius Advanced Cybernetics GmbH makes it possible to generate such composite images from the images generated with the image recording device 1.

[0102] In principle, the surface normal can be determined for each point (pixel) of the surface from the recorded image intensity, the known characteristics of the illumination and the surface emission characteristics in a known manner, see for example B. K. P. Horn, M. J. Brooks: "Shape from Shading", The MIT Press, Cambridge MA, 1989.

[0103] The tilt images in the X or Y direction can now be determined from the surface normal. They result from the brightness change in the respective direction. The respective derivative can be calculated by finite differences, Sobel operators or other derivative kernels.

[0104] The height map can be determined by integration on the basis of the two tilt images. For this purpose various methods are known which inter alia avoid errors due to bias and ambiguity from the integral accumulation. The height image corresponds to the height map.

[0105] The height map The curvature image is calculated as follows : This method takes into account the curvature in both directions, but is challenging in application due to the interference of image noise. Techniques and filters are known for solving this problem.

[0106] Further explanation of the computation of the corresponding images can be found in the literature, such as 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 / ROBOT.1989.99964.

[0107] Texture images are essentially optical images of a surface.

[0108] Different images in Figure 2 The text shows: a) A tilted image in the x-direction; b) Tilt image in the y-direction; c) Curvature image; d) Height image; e) Texture image.

[0109] It is clear that defects in the cutting edge region stand out prominently from the image background in the curvature and height images. Due to the large difference in the front and rear tilt of the cutting edge, the tilted images are more difficult to interpret, and the contrast in the texture images is significantly lower.

[0110] In the described embodiments, height images are primarily used for further processing because they provide good results in defect identification. Additionally or alternatively, other synthetic images may be used, such as x and / or y tilt images, which are well-suited for the analysis of forming tools and threading cutters, for example; or curvature images, which enable the identification of minute defects in the micrometer range. In contrast, texture images are particularly suitable for identifying defects on coatings.

[0111] Now Figure 3 The image shows height images of the surfaces of shank end mills, where these shank end mills have different wear levels. Three images are shown for each level: a) Excellent condition, new product; b) In excellent condition, like new, immediately after renovation; c) Good condition, with no signs of use; d) Moderate condition, showing signs of use and damage; e) Poor condition, with major damage such as chipping or broken teeth.

[0112] As can be clearly seen from the images, the wear and tear is easily identifiable in the high-resolution images.

[0113] Next, the height image is cropped to the region of interest (ROI) 15. This allows subsequent processing to be limited to the area on the tool surface where wear is expected to occur. Figure 4 The image shows the cropping process, where the shank end 11 is on the left and the front end is on the right in the displayed height image. On the right side, i.e., the end face side, the outermost white pixel is located, and the image is cropped to that pixel (step 12). On the left side, the image is also first cropped to the outermost white pixel (step 13), and then a constant width is cropped to reach the cutting end (step 14). Since the spacing is always approximately the same, a width that is essentially the same for all shank end mills can be selected.

[0114] The cropped height image now forms the basis for subsequent classification using machine learning methods. Two method variations are shown below: first, a supervised machine learning method, and then an unsupervised machine learning method.

[0115] Supervised machine learning Within the scope of supervised machine learning methods, and within the scope of the embodiments shown, it is intended to assign a wear level to the tool based on a height image of the tool surface, which corresponds as well as possible to a wear level determined by an experienced expert.

[0116] In practice, experts assess tools by identifying the most significant wear points. These points determine the overall wear level of the tool. That is, the classification assigned by the expert depends on a specific area on the tool. This area determines the severity of the wear, rather than the overall picture of the tool. However, due to this classification, the tool as a whole falls into that category. Supervised machine learning methods address this problem through specific data preparation.

[0117] Within the scope of data preparation, the cropped height image is simplified as follows: only those worn areas that are still visible and ultimately important for classification are shown. Two data preparation methods are described below: 1. Hough Transform Straight lines are detected in the image using the so-called Hough transform (see US3,069,654, PVC Hough). The straight lines detected in the height image that correspond to the cutting edge are subtracted from the cropped height image, resulting in a pre-processed image that essentially only displays wear features. This pre-processed image can then be reliably classified because almost all the features it contains are relevant to classification.

[0118] Before performing the actual Hough transform, the cutting edge is cut off: since wear is mainly located on the cutting edge, other areas in the image can be eliminated. This reduces the amount of data and computation time. After cutting off, the image is obtained... Figure 5 The situation shown in a)

[0119] The image is now converted to a binary image and further processed into an edge image using the Canny edge detector. Figure 5 b). Then, the Hough transform is used to detect straight edges and transform it back into the image. The image now still contains the portion where the straight edges extend ( Figure 5 c). Therefore, if this image is subtracted from the edge image, only the wear features are retained. This preprocessed image with wear features is then used as the basis for supervised machine learning classification. The image in Figure 6 It is shown again in the text.

[0120] 2. Fast Fourier Transform During preprocessing using Fast Fourier Transform (FFT), prominent frequencies in the tool surface image are removed, leaving behind defects that do not naturally possess these frequencies. These frequencies are generated by parallel cutting edges and grinding paths. It is well known that these frequencies can be identified using FFT in the image data.

[0121] Specifically, an FFT is first applied to the image to be preprocessed. This yields the dominant frequencies. The strongest signals will originate from the angle of the cutting edge. Therefore, this angle can be determined based on the transformed image. Now, the original image to be preprocessed is rotated by the determined angle so that the cutting edge extends horizontally in the rotated image. The rotated image is then subjected to another FFT and transformed to the frequency space. Undesired frequencies (corresponding to straight, intact cutting edges) are then eliminated using a mask. The resulting new image in the frequency space is then transformed back by an FFT, rotating back the initially determined angle. This yields an image where only irregular structures, particularly defects, are present. A threshold can now be applied to process the generated image into a preprocessed image with wear characteristics, intended as a basis for supervised machine learning classification. This preprocessed image... Figure 7 As shown in the image.

[0122] Next, features for the machine learning classifier are generated. In this case, features that are as meaningful as possible to the wear level of the tool to be evaluated are sought. These features are based on the shape descriptor of the particles. Particles are continuous regions of bright pixels (1) in a binary image. They can be identified by conventional image processing methods. Within the scope of this embodiment, the following 13 features (or a selection thereof) are determined: Serial Number feature definition a） Particle area Number of white pixels in the grain b） Filling area of ​​particles If all holes are filled, the number of white pixels in the particle c） equivalent area diameter of particles The diameter of a circle with the same area as the particle d） Area of ​​the bounding box of the particle Number of pixels within the (rectangular) bounding box e） The principal axis length of the particle The principal axis length of the ellipse having the same normalized second central moment as the particle f） Maximum Ferete diameter of the particle Longest distance between two pixels of a particle g） Convex hull area of ​​particles Number of pixels in the convex hull image of the particle h） Secondary axis length of the particle The length of the minor axis of an ellipse having the same normalized second central moment as the particle i） circumference of the particle Length of the line passing through the center of the edge pixel j） average value Average pixel value within the brightness range k） Standard deviation Standard deviation of pixels within the brightness range l） Number of pixels Number of pixels within the brightness range m） Particle count Total number of particles identified .

[0123] To enable comparison of features despite their different value ranges, they are normalized. This is done using the so-called L2 standard (L2-Norm), which is calculated as follows: .

[0124] For example, in solid material shank end mills, a simplified vector containing only the following features can be used as the feature vector: a) The area of ​​the particle; c) The equivalent area diameter of the particle; e) Principal axis length of the particle; h) The length of the particle's secondary axis; j) The average pixel value within the brightness range; and m) number of particles.

[0125] Depending on the type of tool, other combinations may be advantageous. These can be determined based on correlation analysis, where for features whose cross-correlation exceeds a certain threshold (e.g., 0.95), only one of them is included in the vector.

[0126] The vectors are then classified using a Support Vector Machine (SVM). The results are discussed below.

[0127] Unsupervised machine learning Within the scope of unsupervised machine learning methods, related groups (hereinafter referred to as "clustering") are identified in the examined data. To this end, within the scope of this embodiment, individual segments of the cropped height image are observed and clustered.

[0128] The fragments are generated using a sliding window method with overlapping frames. Here, the input image is iterated over using a predefined window and stride. Features are computed from the generated image fragments. These fragments can then be represented in feature space as a point cloud, where each point represents an image fragment. For example, a window size of 400×400 pixels and a stride of 220 pixels could be used.

[0129] For unsupervised machine learning, different features are used, especially since features applicable to single, small image fragments are now needed. Emphasis is placed on texture features because impairments at the tool location significantly affect texture. Normalization is performed using the L2 criterion as described above. A certain number of features for further processing are determined for each image fragment using the following method: - Two-dimensional Gabor filter; - Local Binary Pattern (LBP); - Haralick characteristics; - Features based on histograms.

[0130] These methods are described below.

[0131] Irregular structures can be detected using a two-dimensional Gabor filter. This two-dimensional Gabor filter consists of a Gaussian filter kernel, which is applied by convolution and image shifting. Three parameters can be set for the filter kernel: Theta parameter, frequency, and standard deviation. Theta, in degrees, determines the kernel's orientation. All features aligned with the kernel's orientation are emphasized. That is, if the value of Theta corresponds to the angle of the cutting edge, the cutting edge and grinding trajectory are emphasized in the filtered image.

[0132] The ideal frequency is determined experimentally. In principle, a smaller frequency corresponds to a higher resolution. The appropriate standard deviation is also determined empirically based on the given tool geometry.

[0133] Then, for example, the following features can be formed from the aggregate of the corresponding pixel values ​​of each transformed image segment: a) The average value of the grayscale image; b) Variance of the grayscale image; c) The median of the grayscale image; d) The sum of binary images generated from grayscale images using a threshold.

[0134] Local patterns in an image can be analyzed using so-called Local Binary Patterns (LBP). The number of patterns appearing is recorded in a histogram. Each pattern corresponds to a binary code, generated by comparing the center pixel with its neighboring pixels. If the brightness value of a neighboring pixel corresponds to or exceeds that of the center pixel, the neighboring pixel is assigned a value of 1; otherwise, it is assigned a value of 0. The binary code is then derived from an ordered sequence of these bit values.

[0135] In the case of LBP, there are three predefined parameters: radius, number of profile points, and method. The radius determines which adjacent pixels are analyzed, and the number of profile points corresponds to the number of adjacent pixels at that radius (i.e., 8 profile points for radius 1 or 16 profile points for radius 2). It has been shown that radius 1 provides good results in the wear zone of solid material shank end mills.

[0136] As a method for generating binary code, only uniform patterns are detected individually, and all non-uniform patterns are assigned to another container (Bin). A uniform pattern here refers to a pattern with at most two 0-1 or 1-0 transitions. With 8 contour points (radius 1), this results in 58 different uniform patterns (corresponding to 58 containers) and another container for all non-uniform patterns. These 59 features are represented by the corresponding values ​​in the histogram.

[0137] Haralick features are derived from a grayscale co-occurrence matrix (GLCM), which combines which pixel pairs appear equally frequently, whether 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, 1 is chosen as the distance between pixels in a pixel pair (adjacent pixels). The following 14 features are calculated by GLCM in all four possible pixel pair directions, for a total of 56 features: feature describe Second moment of angle If a few large values ​​exist in the GLCM, the metric becomes larger. If all values ​​are similar, the metric becomes smaller. Inverse moment / homogeneity This value is high if most items in the GLCM are located on the diagonal. This is the case when checking for local homogeneity of objects. Correlation A measure of the linear correlation between the grayscale values ​​of adjacent pixels. This value increases if the texture scale is much larger than the distance between adjacent pixels. However, it decreases if the texture scale is on the same order of magnitude as the distance. Contrast If most items in a GLCM are far from the diagonal, it indicates that there are large local variations in grayscale values. entropy A measure of the degree of order or disorder within an object. A high GLCM value indicates relatively similar pixel neighborhoods. A low GLCM value indicates uneven pixel distribution. variance This metric provides information about the deviation of each item in the GLCM from the mean; relatively large differences in gray values ​​are weighted more strongly. and average The average value of the sum of gray levels and variance variance of the sum of gray levels Entropy A uniform (flat) distribution of the sum of gray levels has the maximum entropy. variance Variance of grayscale difference Difference Entropy A uniform (flat) distribution of grayscale difference has the maximum entropy. Information metric of relevance 1 Normalized information content of common entropy Information metric of relevance 2 The difference between common entropy and common entropy under the independence assumption Maximum correlation coefficient A measure of the convergence speed of a low-level Markov chain .

[0138] The features based on histograms are the mean and standard deviation of pixel values ​​in an image segment.

[0139] The total number of (texture) features can be reduced to substantial and meaningful features through appropriate measures (e.g., correlation analysis). The reduced number can be approximated by further dimensionality reduction, such as principal component analysis (PCA) to, for example, three principal axes. Clustering is then performed using a suitable clustering algorithm. For this purpose, for example, the k-means++ algorithm or a Gaussian mixture model (GMM) can be used.

[0140] Within the scope of this embodiment, each image is intended to be assigned to one of the following four wear levels: - New grade (new knife) - Wear level 1 (low wear) - Wear level 2 (medium wear) - Wear level 3 (severe wear).

[0141] The k-means++ algorithm is applied to the data to cluster it into a predefined number of categories. Within the scope of this algorithm, a corresponding number of cluster centers are defined and placed in the point cloud according to specific criteria. Each data point is assigned to the nearest cluster center. For each cluster center, the average value is calculated from the assigned data points, and the center is moved accordingly. This process is continued until no more cluster centers are moved.

[0142] Analysis of the results using the so-called elbow method shows that 4, 5, or 6 clusters should well reflect the structure of the data. According to profile analysis, the best profile coefficient was obtained for a cluster number of 2. However, this number does not allow for sufficiently fine differentiation of tool wear. The second best profile coefficient was obtained for a cluster number of 4. Therefore, based on this consideration, a number of four clusters also seems reasonable.

[0143] The number of clusters is also predefined in the GMM clustering algorithm. This algorithm assumes that the clusters are Gaussian distributed, so a Gaussian distribution is fitted to the data for each cluster. Profilometry shows that—except for the less useful number of clusters 2—good coefficients are obtained for cluster numbers 4 and 5. Therefore, it seems reasonable to predefine four clusters here.

[0144] The results of clustering using k-means++ are as follows: Figure 8A As shown in -D. The point cloud is divided into four groups, with their cluster centers marked by white numbered circles. Figure 8A In -D, a set of tools is emphasized using a black cross: Figure 8A New cutting tools; The new tool occupies cluster 0 very well. Therefore, the new tool can be reliably classified in cluster 0.

[0145] Figure 8B Refurbished knives; For tools that were refurbished directly, there was slight scattering in cluster 3. However, most tools—similar to new knives—were reassigned to cluster 0, which seems reasonable for tools that had just been refurbished.

[0146] Figure 8C Wear level 1 according to expert classification; Wear level 3 occupies cluster 3 very well. However, scattering can also be seen in cluster 1, with a line even extending into cluster 2: analysis shows that the corresponding data points all originate from the same tool, which has very severe contamination in the chip space.

[0147] Figure 8D Wear level 2 according to expert classification; Wear level 2 occupies a good portion of cluster 1, and there are also elements in cluster 3, corresponding to areas of less wear on the surface. There are no defects in cluster 2.

[0148] Figure 8E Wear level 3 according to expert classification.

[0149] The pattern is scattered throughout all clusters 1, 2, and 3. This indicates areas of severe wear.

[0150] From a qualitative perspective, the results of clustering using GMM are very similar to those of k-means++ clustering.

[0151] To check whether the assignment of wear levels to tools aligns with the expert classification, each tool must be analyzed individually. It should be noted that even heavily worn tools often have areas of less wear—therefore, those image segments with high wear are decisive in their assignment to wear levels. Within the framework of unsupervised machine learning, the approach is that the determined wear level of a tool always corresponds to the most severe wear point appearing on the tool.

[0152] To validate these methods, 200 used shank end mills were used. Each shank end mill was classified into one of three wear levels by three experts. Due to discrepancies in the experts' classifications, only the 82 tools that were unanimously evaluated by the experts were further considered. In addition, 20 new and refurbished shank end mills were added.

[0153] Of the final 122 tools, a shell recording for machine learning was performed on each tool using the aforementioned image recording device.

[0154] For supervised learning, Hough transform (HT) is used once and fast Fourier transform (FFT) is used once.

[0155] In the unsupervised concept, the interpreted clustering algorithm k-means++ and Gaussian mixture model (GMM) are compared.

[0156] For each solution approach (supervised and unsupervised), two models are trained separately. One model uses a tool with wear level 3, and the other does not. The reason for omitting wear level 3 is that this level contains many tools with large fractures, which correspond to high wear, but this is not reflected in the surface texture (or is only reflected very locally).

[0157] The data records were manually subdivided into training data records and test data records. Therefore, each classifier was based on the same test and training data. This allows supervised and unsupervised learning classifiers to be directly compared to each other in every case. An 80% to 20% split was used. Here, in each wear level, 1 / 5 of the available data was moved into the test data record. This resulted in the following split: New / Refurbished VK 1 VK 2 VK 3 total Training dataset 32 24 26 15 97 Test dataset 8 5 6 3 22 .

[0158] For evaluation, in particular, the metric of accuracy is used, which examines the ratio of correctly classified tools to the total number of predictions. Its definition is as follows: , The parameters have the following meanings: TP: True positive; TN: True negative; FP: False positive; FN: False negative.

[0159] Overall, the accuracy results relative to the test dataset are as follows: .

[0160] Overall, unsupervised learning using GMM clustering yields the best results, with significantly better accuracy when wear level 3 is omitted. Therefore, in extending this method, it is appropriate to identify large local defects (e.g., broken teeth) through other methods (e.g., appropriate image processing) and directly assign these tools to wear level 3. These additional tools can then be further analyzed using the described method.

[0161] The described classification method using unsupervised machine learning also offers the advantage of obtaining a more detailed image of tool wear beyond just wear levels due to the local classification of many individual image fragments. For example, a histogram based on the assignment of individual image fragments to wear levels can easily identify, for instance, the size of the portion with the highest wear level. Another interesting dimension is the sum of the wear levels of the tool data points.

[0162] Figure 9 This is a schematic block diagram of a tooling facility for determining wear condition and refurbishing tools according to the present invention.

[0163] The facility includes a machining machine 101, such as a milling machine, arranged in the first factory 100. The machining machine includes, in a manner known per se, at least a spindle 102, a tool magazine 103, and a transfer device 104 with a tool holder, by means of which the tool 2 can be changed between the spindle 102 and the tool magazine 103. In the described example, the tool is a solid material shank end mill with a helical primary cutting edge on the circumferential surface and a straight secondary cutting edge on the end face. The transfer device 104 further allows the tool 2 to be fed to a removal position 105 for disposal. The tool 2 can also be transferred to the image recording device 1, which is further described above.

[0164] Data recorded by the image recording device 1 is transmitted to the 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 data from the image recording device 1 and processes it into an image with surface structure information as described above.

[0165] The image is fed to the second image processing module 112 and further processed there, namely, cropped to a certain size and filtered as described above. The corresponding preprocessed image is fed to the classification module 113, which assigns a wear level ("continue to use", "refurbish", "dispose of") to the depicted tool.

[0166] In this way, each tool is inspected for its wear condition after being removed from the spindle. It may be advantageous to perform a cleaning step before inspection to prevent measurements from being affected by adhering dust or chips. For this purpose, a cleaning device, such as one with a liquid or air nozzle, can be used. If the wear condition allows for further use, the tool is stored in tool magazine 103. If the tool needs to be refurbished or disposed of or recycled, it is moved to removal position 105. The classification results are also indicated. Data related to the tool 2 to be refurbished, along with the tool's unique identifier, is stored in a central database 120. The central identifier is also recorded on the tool 2, for example, optically or electronically.

[0167] If tool 2 needs to be refurbished, it is sent to refurbishment equipment 150 in the usual manner. There, the identifier is first read by a reading device 151, such as with the aid of a camera or RFID reader and downstream electronics. Then, controller 152 retrieves data related to tool 2 from database 120 based on the identifier. Subsequently, the refurbishment machine, such as grinding machine 153, is controlled based on the retrieved data. The data includes, for example, information related to the area to be refurbished (end face, circumferential face; specific indication of the cutting edge or cutting edge area) and / or information related to the current geometry of the tool. Thus, refurbishment can be performed efficiently and specifically without further data acquisition. The information related to the refurbished work is again stored in database 120 in association with the tool identifier.

[0168] After the refurbishment is performed, tool 2 is sent back to factory 100 (or another factory). It can continue to be used there.

[0169] This invention is not limited to the embodiments shown. Another source of surface images can be used, and preprocessing and / or classification can be performed in a different manner. Furthermore, the image recording equipment and processing unit can be arranged and operated independently of a specific processing machine.

[0170] In summary, it should be noted that the present invention provides a method for determining the state of a tool, which can automatically and reliably detect the state of the tool.

Claims

1. A method for determining the state of an instrument, comprising the following steps: a) Record multiple optical images of the tool surface under different lighting conditions; b) Process the image data of the plurality of optical images to generate an image with surface structure information; c) Preprocess the image containing surface structure information to generate one or more preprocessed images; d) Based on the one or more preprocessed images, classify the tool state into at least one of two categories using machine learning methods.

2. The method according to claim 1, characterized in that, The image data processing is performed using photometric stereo analysis.

3. The method according to claim 2, characterized in that, Images containing surface structure information are height images and / or curvature images.

4. The method according to any one of claims 1 to 3, characterized in that, The preprocessing step includes cutting to the region of interest, which includes those areas of the tool surface that are affected by wear or are expected to fail.

5. The method according to any one of claims 1 to 4, characterized in that, The machine learning method described is a supervised machine learning method.

6. The method according to claim 5, characterized in that, To generate the one or more preprocessed images, wear features or fault features are isolated from the images containing surface structure information, particularly by at least one of the following methods: a) Apply the Hough transform to images containing surface structure information; b) Apply Fast Fourier Transform to images with surface structure information.

7. The method according to claim 5 or 6, characterized in that, In a preprocessed image, particles with pixels within a pre-given brightness range are identified, and based on the identified particles, several of the following parameters are determined as features for a machine learning method: a) The area of ​​the particle; b) The filling area of ​​the particles; c) The equivalent area diameter of the particle; d) Area of ​​the particle's bounding box; e) Principal axis length of the particle; f) Maximum Freette diameter of the particle; g) The convex hull area of ​​the particle; h) The length of the particle's secondary axis; i) The perimeter of the particle; j) The average value of pixels within the brightness range; k) Standard deviation of pixels within the brightness range; l) Number of pixels within the brightness range; m) number of particles.

8. The method according to claim 7, characterized in that, The following parameters were identified as features of the machine learning method: a) The area of ​​the particle; c) The equivalent area diameter of the particle; e) Principal axis length of the particle; h) The length of the particle's secondary axis; j) The average pixel value within the brightness range; and m) number of particles.

9. The method according to any one of claims 1 to 8, characterized in that, The machine learning method is an unsupervised machine learning method, wherein multiple features are determined based on a preprocessed image or multiple preprocessed images in the first step, and clustering is performed in the second step.

10. The method according to claim 9, characterized in that, To generate multiple preprocessed images, the image with surface structure information is divided into multiple overlapping patches using a sliding window method. For each preprocessed image corresponding to a patch, multiple pre-given features are determined for clustering.

11. The method according to claim 10, characterized in that, Features from the preprocessed image include multiples of the following groups: a) The mean, variance, and / or median of the preprocessed image further processed using a two-dimensional Gabor filter; b) The sum of binary images obtained from the preprocessed image further processed with a two-dimensional Gabor filter; c) Features based on local binary pattern analysis; d) Haralick texture features; e) The mean and / or standard deviation of the pixel values ​​of the preprocessed image.

12. The method according to any one of claims 9 to 11, characterized in that, Clustering is divided into 3-8 categories, especially 4-6 categories.

13. The method according to any one of claims 9 to 12, characterized in that, Clustering is performed using a Gaussian mixture model or the k-means++ method.

14. The method according to any one of claims 1 to 13, characterized in that, Based on image data from the plurality of optical images and / or based on images with surface structure information and / or based on one or more cut images with surface structure information, the broken cutting edge of the tool is identified, and when the broken cutting edge is identified, the tool is directly assigned to the corresponding state category.

15. The method according to any one of claims 1 to 14, characterized in that, The method for determining the state is applied to grinding or drilling tools used in machine tools to process workpieces, especially to shank tools made of solid materials.

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

17. The method according to claim 16, characterized in that, To determine the first state, an algorithm based on the first dataset is used, and to determine the second state, an algorithm based on the second dataset is used, where the first and second datasets are different.

18. A method for refurbishing tools, comprising the following steps: a) Determine the state of the tool using the method according to any one of claims 1 to 17; b) If the state meets the pre-given conditions, control at least one device for refurbishing tools, especially by means of a grinding process.

19. The method according to claim 18, characterized in that, An illumination system for illuminating the tool surface under different lighting conditions and a camera for recording multiple optical images of the tool surface under different lighting conditions are arranged at a first point of use. The data obtained at the first point of use is stored in a database. A device for refurbishment is arranged at a second point of use, and the device for refurbishment retrieves data from the database.

20. An apparatus for determining the state of a tool, comprising: a) A lighting system for illuminating tool surfaces in different lighting conditions; b) A camera used to record multiple optical images of the tool surface under different lighting conditions; c) A first image processing module configured to process image data of 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; as well as d) A classification module configured to classify the tool state into at least two categories based on the one or more preprocessed images using machine learning methods.

21. The device according to claim 20, characterized in that, The lighting system and camera are integrated into a machining machine with a tool housing, in particular, so that when the tool is housed in the housing, the lighting system can illuminate the tool and the camera can record an optical image of the tool surface.

22. An apparatus comprising: a) The apparatus for determining the state of a tool as described in claim 20 or 21; b) Apparatus for refurbishing tools, especially by means of a grinding process; c) A controller for controlling the refurbishing apparatus, configured to receive state-related information from the device for determination, and to control the apparatus for the refurbishing tools based on the received information.

Images