Determination of the degree of wear of a tool

Abstract

This invention relates to a computer-implemented method for more accurately determining the degree of wear of a tool (16). An image data set (10) depicting a wear-related region (11) is provided. Using an artificial neural network (13), a calculation unit (12) assigns each image point (14) from a plurality of image points to a category from a preset category set, wherein the category set contains at least one wear category. Based on the assignment result, the calculation unit (12) determines at least one feature value of the degree of wear.

Description

Technical Field

[0001] This invention relates to a computer-implemented method for determining the degree of tool wear, wherein a set of image data depicting wear-related areas of the tool is provided. The invention also relates to a computer-implemented method for training an artificial neural network to determine the degree of tool wear, an analysis system for determining the degree of tool wear, and a computer program. Background Technology

[0002] In industrial manufacturing, tool wear has a significant impact on production costs. Tool wear depends not only on the tool's past usage time but also on factors such as the material being machined and the machining parameters, and it increases continuously, but typically non-linearly. If tool wear reaches a predetermined maximum allowable wear, the tool is considered worn out. Continued use of a worn tool will result in a significant decline in component quality and machining performance.

[0003] To avoid this, tools are typically replaced before they become necessary, given maximum wear. Consequently, tooling costs increase based on the tool's unused usage time. The relationship between increased wear and the material being processed, as well as batch relationships within the same specified tool, also makes estimating the degree of wear difficult, and thus leads to a tendency to replace tools prematurely.

[0004] In industrial production, tool wear is estimated using computational methods. This involves evaluating machine signals such as vibration, acoustic emission, cutting force, and machine current. However, due to the indirect nature of these methods, high-precision wear estimations are not possible. Furthermore, the methods used are often limited to specific tool types or cutting edge geometries.

[0005] For example, patent document US2016 / 0091393 A1 describes a method in which operating parameters of a machine are determined, particularly position and motion parameters, such as spindle speed or feed rate. Parameters are extracted using cluster analysis, and trends in the parameters are taken into account to estimate the degree of wear.

[0006] In other approaches, laser beams are used to measure tools, which allows for the identification of tool malfunctions, such as cracks. However, this makes it impossible to accurately determine the degree of wear. Summary of the Invention

[0007] Against this background, the technical problem to be solved by the present invention is to describe an improved scheme for determining the wear degree of a tool, which can be flexibly applied to different tool types and can more accurately determine the wear degree.

[0008] The improved approach is based on the idea of ​​analyzing image data from a depiction tool using an artificial neural network. This network assigns a category to multiple image points (or pixels) within a group of image data. The degree of wear is then determined based on the analysis results.

[0009] According to a first independent aspect of the improved scheme, a computer-implemented method for determining the degree of wear of a tool is provided. A set of image data depicting wear-related regions of the tool is provided. By a computing unit, using an artificial neural network, particularly a trained neural network such as a convolutional neural network, each image point in a plurality of image points in the image data set is assigned (or classified, matched, associated) to a category of a predetermined set of categories. The plurality of image points particularly includes three or more image points. The category set contains at least one wear category. Based on the result of assigning each image point in the plurality of image points to a category, at least one feature value of the degree of wear is determined by the computing unit.

[0010] Image data sets are, in particular, images from image sensors, such as CMOS sensors, or digital images from cameras, especially digital camera images. Image data sets have a two-dimensional arrangement of image points, also known as pixels, which are arranged, particularly in columns and rows.

[0011] Multiple image points can, for example, comprise all image points of an image data set or a predefined subset of image points in the image data set. For instance, multiple image points can comprise a predefined portion of the image points in the image data set, such as each second image point, each fourth image point, etc. If multiple image points comprise a subset of image points in the image data set, but not all image points in the image data set, then the multiple image points can, for example, be spatially uniformly distributed or approximately uniformly distributed.

[0012] The tool can be a machining tool for a machine tool, such as a cutting tool, such as a lathe tool, boring tool, milling cutter, or drill bit. It can especially be an indexable insert.

[0013] Wear-related areas of a tool, especially worn areas of the tool, such as surfaces, that change due to the tool’s regular use, and whose changes can limit the tool’s functionality, availability, loadability, or performance, or whose changes are associated with limited functionality, availability, performance, or loadability.

[0014] In particular, the area associated with wear can be the cutting edge or part of the cutting edge of a cutting tool, such as an indexable insert, drill bit, milling tool, or lathe tool.

[0015] The provision of image data sets particularly includes providing image data sets in a computer-readable form, especially enabling the image data sets to be read by a computing unit for processing using a neural network.

[0016] Providing image data sets particularly includes storing image data sets on storage media or storage units, and providing storage media or storage units.

[0017] For example, an image point can correspond to a single sampled value of an image sensor, or, in the case of a CMOS sensor, a value measured by a photodetector or photodiode. However, the image data set can also be preprocessed to, for example, combine the individual sampled values ​​to form a valid image point, and an image point in this basic understanding can also correspond to a valid image point.

[0018] By assigning a category to each of multiple image points, specifically assigning one category to each individual image point among several of those points. For example, for each assignment of an image point to a category, a feature value that determines the probability of a correct assignment can be determined by a computational unit.

[0019] A neural network can be understood, for example, as software code or a combination of software code components, wherein the software code or software code components are stored in a computer-readable form, such as stored on a storage medium or storage unit. A neural network can be understood in particular as a software module, or may contain one or more software modules. A computing unit can read or execute a neural network, that is, execute the software code or software code components.

[0020] Neural networks are specifically designed as Convolutional Neural Networks (CNNs). This means that a neural network has at least one convolutional layer or layer. In particular, neural networks can also have regular, especially non-convolutional, layers or layers, such as fully connected hidden planes, or one or more fully connected planes. In these cases, the neural network will still be referred to as a Convolutional Neural Network or CNN.

[0021] The allocation results specifically include information related to the allocation or the assignment results (or classification results, matching results) of each image point to its corresponding category. For example, the allocation results may include information for each of multiple image points, indicating which category from a predefined set of categories has been assigned to that image point. The results may also include corresponding feature values ​​indicating the probability of a correct allocation.

[0022] Unlike existing solutions, the improved approach avoids using indirect extrapolation based on other machine parameters to estimate wear levels. Instead, it directly analyzes wear-related areas of the tool visually. Therefore, the complex nonlinear relationship between the tool's past usage time and wear increase, as well as the responsibility relationships, are irrelevant to the analysis based on the improved approach. Consequently, wear levels can be analyzed or determined significantly more accurately.

[0023] This, in particular, avoids unnecessary premature tool replacements. This leads to cost savings, which can be especially significant in industrial manufacturing environments. For example, modern machine tool cutting heads can have up to 50 cutting edges, which are replaced at regular intervals, ranging from minutes to hours depending on the tool material and raw materials.

[0024] The improved scheme also allows for the determination of wear levels independently of external disturbances or environmental influences. Furthermore, different wear types can be identified using the improved scheme.

[0025] By using artificial neural networks, direct visual analysis of wear-related areas can be performed, and this is provided in principle without being limited to specific tool types or blade geometries. This allows for highly flexible use of improved solutions.

[0026] The improved approach does not assign the entire image, or the entire image dataset, to a single category. Instead, each image point is assigned its own category independently. This pixel-by-pixel analysis reduces computational workload, especially when training neural networks. Furthermore, this method allows for particularly precise analysis and thus more accurate determination of wear levels.

[0027] According to at least one embodiment, a set of image data is provided having a resolution in the range of 1 image point / mm to 500 image points / mm, for example, from 50 image points / mm to 150 image points / mm, preferably about 100 pixels / mm.

[0028] The degree of wear can be accurately and reliably determined with a correspondingly high resolution.

[0029] At least one wear category may have exactly one wear category or more than one wear category, wherein different wear categories may correspond to different types of wear indices, different wear mechanisms or different wear types.

[0030] In addition to at least one wear category, the preset category set may include, for example, a background category and / or a category for undamaged tool surfaces.

[0031] If the corresponding image point is not on the tool, then the background category is assigned to that image point, for example.

[0032] If the corresponding image point is located on the tool and does not show wear, then the category of undamaged tool surface is assigned to that image point, for example.

[0033] If an image point is located on a tool and there is corresponding wear at that point on the tool, then the corresponding wear category of at least one wear category is assigned to the image point.

[0034] According to at least one embodiment, the tool has a machining tool for a machine tool, especially a cutting tool for a lathe or milling machine, such as an indexable insert.

[0035] According to at least one embodiment, the neural network applies a computational unit to the environmental region of each of a plurality of image points so as to assign the corresponding image point to one of the categories in a preset category set.

[0036] The environmental region, especially the subset of multiple image points, is selected or defined according to preset rules.

[0037] Because the defined environment of an image point is used as input to a neural network, the image information of that point can be effectively compared and evaluated with its surrounding points. The relative image information in the surrounding environment is used to characterize or classify the image point.

[0038] According to at least one implementation, the corresponding subset of image points is an associated subset. This means that each image point in the subset has at least one neighboring image point among a plurality of image points, which is also part of the subset.

[0039] According to at least one implementation, the corresponding image points to which the neural network is applied in its environment region are either surrounded or enclosed by a subset of the remaining image points, or the corresponding image points are edge points of multiple image points.

[0040] In this example, an image point in an image data set that corresponds to the first or last image point in a row or column can be understood as an edge point.

[0041] According to at least one embodiment, the environment region includes corresponding image points to which the neural network is applied to its surrounding environment, as well as all image points, particularly directly adjacent to the corresponding image points. If the image points in the plurality of image points are arranged in rows and columns as described, then the corresponding image point has eight image points directly adjacent to each other, such that in this case, the subset contains, for example, nine image points.

[0042] According to at least one embodiment, in addition to the adjacent image points of the corresponding image point, the environment region also includes all the adjacent image points of the adjacent image point.

[0043] In other words, the environment region contains the corresponding image point and the next and the next adjacent image points. If the image points are arranged in rows and columns, the environment region may contain, for example, 25 image points.

[0044] According to at least one embodiment, the environmental region includes corresponding image points and the remaining image points among a plurality of image points, which have a certain distance from the corresponding image points, the distance being less than a preset maximum distance.

[0045] The environment region can be viewed as a sliding window that moves progressively across the entire image data set, where a category is assigned to the corresponding image fragment for each position of the window.

[0046] According to at least one embodiment, the environmental region includes a corresponding image point and all remaining image points among the plurality of image points, wherein the row spacing between the remaining image points and the corresponding image points is less than a preset maximum row spacing, and the column spacing between the remaining image points and the corresponding image points is less than a preset maximum column spacing.

[0047] In such an implementation, the environmental area has, for example, a rectangular shape.

[0048] According to at least one embodiment, the result of the allocation is processed by a calculation unit according to the morphological image processing operation, and at least one feature value of the degree of wear is determined by the calculation unit based on the processed result.

[0049] According to at least one embodiment, morphological image processing operations include erosion operations and / or dilation operations.

[0050] According to at least one embodiment, the morphological image processing operation includes an opening operation, i.e., in particular an erosion operation, followed by a dilation operation.

[0051] By applying morphological image processing operations multiple times when necessary, misclassifications can be corrected and noise in the assignment results can be reduced. Therefore, higher classification accuracy can be achieved, and the degree of wear can be determined accordingly.

[0052] According to at least one implementation, a category image is generated based on the allocation of image points among a plurality of image points according to a category, i.e., based on the result of the allocation, wherein the category image corresponds to a copy of the image data set, in which each image point is marked or emphasized according to the category assigned to it, for example, by color.

[0053] In order to process the assigned results according to morphological image processing operations, for example, categorical images can be processed using morphological image processing operations.

[0054] According to at least one embodiment, category images and / or category images processed according to morphological image processing operations can be output to a user, for example, through a user interface, especially a display or a display unit of a user interface. The user can understand from the output what results have been achieved according to the improved scheme or method, and, if necessary, what conclusions or recommendations can be drawn accordingly.

[0055] According to at least one embodiment, a calculation unit determines the proportion of image points from a plurality of image points that have been assigned to at least one wear category, i.e., in particular, one wear category of at least one wear category. Based on this proportion, the calculation unit determines the wear surface, such as an effective wear surface, as a feature value of the degree of wear, i.e., a feature value as at least one feature value.

[0056] In other words, the allocation result includes the share of image points that have been assigned to at least one wear category, and at least one feature value includes the wear surface.

[0057] In implementations of image processing operations that apply morphology, the wear surface or its determination should be understood as such that the wear surface is determined based on the proportion of image points among a plurality of image points that have been assigned to at least one wear category and are always assigned to one of at least one wear category after processing by the morphological image processing operation.

[0058] This also applies to other characteristic values ​​used to determine the degree of wear, especially the width of the wear mark.

[0059] For example, this share can be calculated as the number of points assigned to at least one wear category divided by the total number of image points in a plurality of image points. Wear surfaces are particularly proportional to the share multiplied by the total number of image points in a plurality of image points divided by the total number of tool image points. Here, the number of tool image points corresponds to the number of image points assigned to at least one wear category or a category for undamaged tool surfaces.

[0060] In an alternative implementation, the share can also be calculated as the number of image points assigned to at least one wear category divided by the number of tool image points. Area is directly proportional to the share.

[0061] For each tool, the wear surface directly provides information about the remaining lifespan of the tool and / or can be an important indicator of the current degree of wear on the tool.

[0062] According to at least one embodiment, for a column or row of image points among a plurality of image points, a calculation unit determines an additional share of image points already assigned to at least one wear category. The wear mark width of the column or row is determined based on this additional share. The calculation unit then determines additional feature values ​​for at least one feature value representing the degree of wear based on the wear mark width of the row or column.

[0063] In other words, the assignment result includes the number of image points in a column or row that have been assigned to at least one of the wear categories, and at least one feature value includes the wear mark width of that column or row. The wear mark width can be described, for example, in units of length, i.e., meters or millimeters, or in units of image points.

[0064] The width of the wear mark is directly proportional to the other proportions.

[0065] Depending on the tool, the width of the wear marks, especially together with the corresponding calculated width of the wear marks in other rows or columns, can be a meaningful indicator of the degree of tool wear.

[0066] According to at least one embodiment, for at least one additional column or row of image points among a plurality of image points, a calculation unit determines a corresponding additional share of image points already assigned to at least one wear category. The calculation unit then determines a corresponding additional wear mark width for the corresponding additional column or row based on the corresponding additional share. Finally, the calculation unit determines additional feature values ​​based on the wear mark width and the additional wear mark width.

[0067] By considering the additional wear mark width, the reliability and effectiveness of the additional feature values ​​are improved.

[0068] In particular, the wear mark width is determined for each column of image points or for each row of image points in a plurality of image points, and additional feature values ​​are determined based on all these wear mark widths.

[0069] Whether to consider using rows or columns of image points to calculate the width of the wear mark depends on the tool's orientation within the image and / or the preprocessing of the image data set.

[0070] According to at least one embodiment, at least one characteristic value of the degree of wear, and in particular another characteristic value, includes statistical characteristic parameters of all specific wear mark widths, such as maximum, average or median.

[0071] According to at least one implementation, the result of the allocation includes all wear categories or parameters derived from the distribution of image points from a plurality of image points to at least one wear category.

[0072] According to at least one embodiment, the wear-related area is imaged by a camera, particularly a microscope camera device, in order to generate and provide a set of image data.

[0073] Microscope camera equipment specifically includes a microscope camera or microscope and a camera coupled to the microscope.

[0074] This enables precise and detailed mapping of wear-related areas, and consequently, high accuracy in determining the degree of wear.

[0075] According to at least one embodiment, a calculation unit compares at least one feature value with at least one preset limit value. The calculation unit then determines a value related to the remaining usage time of the tool based on the comparison result.

[0076] The remaining usage time of a tool specifically corresponds to the remaining time until the tool is expected to reach its predefined maximum wear.

[0077] Values ​​related to the remaining lifespan of a tool can, for example, correspond to binary values, indicating whether maximum wear has been reached.

[0078] Values ​​related to remaining usage time may also include time (e.g., in hours, minutes, and / or seconds), and / or include a percentage of the total usage time of the tool.

[0079] By comparing at least one feature value with at least one limit value, in particular by comparing each feature value of at least one feature value with the associated limit value of at least one preset limit value.

[0080] Different limit values ​​of at least one limit value can also depend on or define each other.

[0081] By determining the value related to the remaining usage time, it can be determined whether the tool must be replaced, whether it can continue to be used, or when the tool needs to be replaced.

[0082] According to at least one embodiment, a tool replacement suggestion is proposed by a calculation unit based on the result of comparing at least one feature value with at least one preset limit value.

[0083] In particular, replacement suggestions can be output as visual signals on the user interface.

[0084] If one or more of the eigenvalues ​​are greater than or equal to the associated limit value, then a tool replacement can be suggested, for example, based on the replacement recommendation.

[0085] According to another independent aspect of the improved scheme, a computer-implemented method for training an artificial neural network to determine the degree of tool wear is described. A training image dataset is provided, depicting wear-related regions of a reference tool. Each image point in the training image dataset is provided with a reference assignment to a category in a predefined category set, wherein the category set contains at least one wear category. A computational unit is trained to compute the output of the neural network for each image point and compares the output with the reference assignment. The computational unit is trained to adjust the neural network based on the comparison results.

[0086] The training computation unit can be a computation unit or a separate, independent computation unit.

[0087] By using a reference assignment, which can be stored as a file on a storage medium or storage unit, a predefined category from a set of categories is assigned to each of a plurality of image points. Therefore, the corresponding reference assignment of an image point can be understood as a label or target used to train a neural network.

[0088] For each image point, an output neural network corresponds to a distinct category. The network assigns that category to the corresponding image point in an untrained or partially trained state. This output is compared to a reference assignment by comparing the distinct category of the corresponding image point with the category assigned according to the reference assignment.

[0089] In particular, the neural network is adjusted by changing the weights of each neuron in order to reduce the deviation between the output and the reference assignment of the corresponding image points.

[0090] To calculate the output of the neural network, the neural network is applied to the environmental region of the corresponding image point, particularly as described above regarding the computer implementation for determining the degree of wear.

[0091] According to at least one embodiment of the method for determining the wear level of a tool, the neural network is trained using a method for training an artificial neural network according to an improved scheme. The method steps for training the neural network are also method steps for determining the wear level.

[0092] In particular, the artificial neural network is trained before assigning image points from a set of image data to the corresponding categories. However, additionally, training can be repeated at later time points to optimize the neural network model.

[0093] According to at least one embodiment of the method for determining the wear level of a tool, the neural network is trained using a method for training an artificial neural network according to an improved scheme. The method steps for training the neural network precede the method steps for determining the wear level, and are not themselves part of the method for determining the wear level.

[0094] Another implementation of the method for training an artificial neural network according to the improved scheme is directly derived from different implementations of the method for determining the wear level of a tool according to the improved scheme, and vice versa.

[0095] According to another independent aspect of the improved scheme, an analysis system for determining the degree of tool wear is provided. The analysis system has a computational unit and a storage unit. An image data set, depicting wear-related areas of the tool, is stored in the storage unit. The computational unit is designed to, using, in particular, a trained artificial neural network, especially a convolutional neural network, assign (or classify, match, associate) each of multiple image points in the image data set to a category of a predetermined set of categories, wherein the set of categories contains at least one wear category. The computational unit is further designed to determine at least one feature value of the degree of wear based on the result of assigning the image points from the multiple image points to their respective categories.

[0096] In particular, a storage unit may contain one or more storage media.

[0097] Neural networks are stored, in particular, on memory units.

[0098] For example, the categories of a predefined category set are also stored in storage units.

[0099] According to at least one embodiment, the analysis system includes an output unit, such as an image output unit, specifically for outputting category images or category images processed according to morphological image processing operations to a user of the analysis system.

[0100] According to at least one embodiment, the analysis system has a microscope camera device designed to image wear-related areas of the tool in order to generate a set of image data.

[0101] Microscope camera devices or computing units are particularly designed to store groups of image data on storage units.

[0102] According to at least one embodiment of the analysis system based on the improved scheme, the neural network is trained by the method for training the artificial neural network based on the improved scheme.

[0103] Further implementations of the analysis system based on the improved scheme are directly derived from different designs of the methods for determining wear levels and for training artificial neural networks based on the improved scheme, and vice versa. In particular, the analysis system is designed or programmed to execute the methods based on the improved scheme, or the analysis system executes the methods based on the improved scheme.

[0104] According to another independent aspect of the improved scheme, a computer program with instructions is described. When the computer program is executed by a computer system, especially by an analysis system according to the improved scheme, for example by the computing unit of the analysis system, the instructions cause the computer system to execute a method according to the improved scheme for determining the degree of tool wear.

[0105] According to another independent aspect of the improved scheme, an additional computer program with additional commands is described. When the additional computer program is executed by a computer system, especially an analysis system according to the improved scheme, such as the processing unit of the analysis system, the additional commands cause the computer system to execute the method for training artificial neural networks according to the improved scheme.

[0106] According to another independent aspect of the improved scheme, a computer-readable storage medium is described, on which a computer program and / or other computer programs according to the improved scheme are stored.

[0107] The features and combinations of features mentioned above in the specification, as well as those mentioned below in the description of the drawings and / or shown separately in the drawings, may be used not only in the combinations described separately, but also in other combinations, without departing from the scope of protection of this invention. All embodiments and combinations of features that do not have the features of the originally drafted independent claims and / or exceed or deviate from the feature combinations described in the reference relationships of the claims are also considered disclosed. Attached Figure Description

[0108] The present invention will then be described in detail with reference to specific embodiments and related schematic diagrams. In the drawings, identical or functionally identical elements may be given the same reference numerals. It is not necessary to repeat the description of identical or functionally identical elements for different drawings if necessary.

[0109] In the attached diagram:

[0110] Figure 1 A schematic diagram illustrating an exemplary implementation of the analysis system according to the improved scheme is shown; and

[0111] Figure 2Flowcharts are shown of exemplary embodiments of a method for training a neural network according to an improved scheme and of exemplary embodiments of a method for determining the wear level of a tool according to an improved scheme. Detailed Implementation

[0112] Figure 1 A schematic diagram of an analysis system 20 according to an improved scheme is shown.

[0113] The analysis system 20 has a computing unit 12 and a storage unit 19 coupled to the computing unit 12.

[0114] Image data set 10 is stored in storage unit 19, which depicts the wear-related area 11 of tool 16.

[0115] Furthermore, a software module with an artificial neural network 13 is stored on storage unit 19, wherein the neural network 13 specifically includes a convolutional neural network 25. Therefore, the neural network 13 itself is also referred to as a convolutional neural network (CNN).

[0116] Optionally, the analysis system 20 may have a microscope camera device 17, through which the tool 16, and in particular the wear-related area 11 of the tool 16, can be imaged to generate a set of image data 10.

[0117] Optionally, the analysis system 20 may also have an image output unit 34 to provide visual output to the user of the analysis system 20.

[0118] Tool 16 can be, in particular, a cutting tool for a machine tool. For example, tool 16 can be an indexable insert for a machine tool, such as a lathe or milling machine. The wear-related area 11 of tool 16 corresponds in particular to the cutting edge or a portion thereof of tool 16.

[0119] However, the improved approach, and especially the analysis system or method based on the improved approach, is not limited to cutting tools or even indexable inserts. The described steps and implementation methods are similarly applicable to other tool types.

[0120] Subsequently, exemplary architectures and topologies of the neural network 13, respectively applicable to use in the analysis system 20 or in methods according to the improved scheme, are described. However, other architectures may also be used if necessary.

[0121] Figure 1The neural network 13 shown has three consecutive convolutional planes 26, 27, and 28. Each convolutional plane 26, 27, and 28 contains, for example, a convolutional layer in a practical sense, followed by an activation function, specifically a rectified or rectifier activation function, which can also be called a rectified linear unit activation function (RELU activation function for short), and further followed by a pooling layer. For example, the pooling layer can be designed to perform 2×2 max pooling or 2×2 average pooling.

[0122] For example, each convolutional plane 26, 27, 28, or convolutional layer in its practical sense, can implement a different number of filters. Here, the number of filters can increase, for example, from the first convolutional plane 26 through the second convolutional plane 27 toward the third convolutional plane 28. In particular, the number of filters in the second convolutional plane 27 can be twice the number of filters in the first convolutional plane 26 and half the number of filters in the third convolutional plane 28.

[0123] The output of the third convolutional plane 28, or CNN 25, is typically multidimensional. In the flattened layer 29 following the CNN, the output of CNN 25 is transformed into a one-dimensional feature vector.

[0124] For example, a conventional neural network 30 is followed by a flat layer 29. The neural network 30 includes, for example, a classifier with, for example, two hidden planes 31, 32, and in particular, dense layers 31, 32. The dense layers 31, 32 can in particular form a multilayer perceptron.

[0125] Furthermore, the neural network 30 may have, for example, an output layer 33 placed after the dense layers 31 and 32. For example, the softmax function may be used for partitioning.

[0126] The neural network 13 thus constructed can assign categories from a preset set of categories to the input, which, for example, corresponds to a portion of the image data set 10.

[0127] The artificial neurons for each category of the preset category set are located in the output layer 33, with five neurons corresponding to the five categories.

[0128] The functions of analysis system 20 will be discussed in the following reference. Figure 2 The method described therein, based on the improved scheme, will be elaborated in more detail.

[0129] Figure 2A flowchart of a method for training a neural network 13 to determine the wear level of a tool 16 according to an improved scheme is shown. In particular, method steps 1 to 4 are shown, which represent the method for training the artificial neural network 13. Furthermore, a flowchart of the method for determining the wear level of the tool 16 according to the improved scheme is shown, particularly through method steps 5 to 9. It is then assumed that the analysis system 20 is used both in the training phase, i.e., for executing the method for training the neural network 13, and in the operation phase, i.e., for executing the method for determining the wear level of the tool 16. However, this is not necessary, as the training phase is independent of the operation phase.

[0130] In step 1 of the method for training neural network 13, a training image data set 18 is provided, which depicts wear-related regions of a reference tool. The reference tool is, for example, a tool of the same or similar type as tool 16 whose wear level should be determined during operation.

[0131] The training image data set 18 is shown in particular at high resolution. For example, a resolution of at least 100 image points per millimeter can be provided for the edge of the reference tool. In step 2 of the method, a reference assignment of each image point in the training image data set 18 to a category of a preset category set is provided. In other words, the data set is provided by pairs of wear images and the associated category of each image point or pixel.

[0132] An untrained or partially trained neural network 13 is sequentially applied to each image point (including the predefined environment of the corresponding image point) via computation unit 12. Computation unit 12 compares the corresponding output of neural network 13 with the corresponding reference assignment.

[0133] To learn the relationship between the input data (i.e., training image data set 18) and the output values ​​(i.e., the associated categories), the neural network 13 is adapted by the computation unit 12 in step 4 of the method. In particular, the corresponding weights of the neural network, especially the CNN 25, are adapted to reduce the error between the input data and the output values. The described steps are repeated for a large number of additional training image data sets 18', 18'' until the neural network 13 has been sufficiently trained, i.e., the error in the assignment is within a preset acceptable tolerance range.

[0134] In an exemplary embodiment, the preset set of categories may include, for example, five categories, three of which are wear categories. Wear categories here specifically represent typical wear types of tool 16. In the case of a cutting tool, the first wear category may be, for example, a surface wear category, representing surface wear on tool 16. The second wear category may be a grooving category (also referred to as a "groove category"), representing cavity-shaped, slotted, or trench-shaped wear features. The third type of wear may be a built-up edge category (also referred to as a "build-up-edge-Klasse"), representing wear features caused by the accumulation of workpiece material on the tool, i.e., a so-called built-up edge.

[0135] The background category can be used as an additional category, for example, when an image point is not located on the tool surface, the background category is assigned to that image point.

[0136] In other implementations, other category types, especially other wear categories in other tool types, may be advantageous.

[0137] A diverse dataset consisting of tool images and their corresponding categories can be advantageously used during the training phase. For example, it can be used to analyze images of different tools, particularly those of different shapes and materials, under varying degrees of wear and different blade types. Consequently, the trained CNN is particularly stable against perturbations such as deviated blades.

[0138] In step 5 of the method for determining the degree of wear, an image data set 10 is provided, in particular stored on a storage unit 19, the image data set 10 depicting the wear-related area 11 of the tool 16, that is, in particular the cutting edge of the tool 16.

[0139] Here, image data set 10 can be generated, in particular, by microscope camera device 17.

[0140] Then, in step 6 of the method, the image data set 10 is analyzed pixel-by-pixel, i.e., image point-by-point, by dividing the image data set 10 into individual image regions 15. Here, each image region 15 may contain an image point 14 (which should be assigned a category by the neural network 13) and other image points surrounding that image point 14, i.e., other image points directly adjacent to that image point 14. In the case of a grid arrangement consisting of rows and columns, the image data set 10 is, for example, divided into several image regions, each having 3×3, i.e., nine image points.

[0141] In step 7 of the method, the computing unit 12 applies the trained neural network 13 to each image point 14 by using the corresponding image region 15 as the input to the neural network 13, and assigning one of the categories to the corresponding image region 15 and the image point 14 being considered accordingly.

[0142] After evaluating all image points in this manner, optional reprocessing is performed in step 7 to reduce signal noise. Here, the classified image can be subjected to an opening operation, i.e., an erosion operation, followed by a dilation operation, by the calculation unit 12. The opening operation can also be performed multiple times consecutively, for example, three times.

[0143] In an exemplary implementation, a kernel, for example, a rectangle with a size of 12×8 image points, is used for the opening operation.

[0144] In step 8, the cleaned category image 21 is schematically shown. Image points assigned to the background category after processing are represented by 22. Image points assigned to the category for undamaged tool surfaces after reprocessing are represented by 23. Image points assigned to one of the wear categories are represented by 24. The cleaned category image 21 can be displayed to the user, for example, on the display unit 34 for better understanding.

[0145] In step 9 of the method, in the thus cleaned image, at least one characteristic value of the degree of wear of the tool 16 can be determined by means of the calculation unit 12. For this purpose, for example, the entire worn surface can be determined as the number or share of those image points that have been assigned to one of the wear categories.

[0146] Optionally, the maximum and / or average wear mark width and / or the median wear mark width can also be determined. For this purpose, the cleaned category image can, for example, be oriented such that the upper side of the wear-related region 11 is horizontally oriented. Then, for each column, the number of image points already assigned to the wear category can be determined as the wear mark width by the calculation unit 12. Then, for example, the maximum, average, or median wear mark width can be calculated.

[0147] The conversion from image points to SI units can be performed using a one-time calibration. To do this, images with traces at fixed intervals, such as a grid with 1mm row spacing, can be recorded, and the intervals between the traces can be determined within the image points, thus establishing a relationship between SI units and image point sizes.

[0148] The characteristic values ​​thus determined by the degree of wear, namely the worn surface, the maximum or average wear mark width, or the median wear mark width, can be compared with relevant predefined limit values ​​(which, for example, correspond to the maximum wear). If the limit value is exceeded, then tool 16 can be replaced. For this purpose, for example, calculation unit 12 can generate corresponding visual output on display unit 34. Otherwise, the current state of tool 16 or the remaining usage time can be output.

[0149] For example, on the one hand, relevant information can be fed back to the machine tool's controller. On the other hand, information can be sent to local edge devices and / or the cloud, where the current process status can be observed, and thus continuous analysis can be performed, for example, regarding the impact of material batches or machine parameters on the final wear.

[0150] The improved approach also offers the potential to increase the acceptance of machine learning and artificial intelligence methods in industrial environments because it not only identifies abstract values ​​but also establishes classified images as an intermediate step. To understand the decisions made, the classified images can be overlaid with camera images and the identified worn surfaces, and then displayed to the machine operator, such as an external tablet, via an output interface.

[0151] Compared to indirect measurement methods, direct wear measurement based on image data is less susceptible to signal noise and environmental interference. The deep learning scheme using CNNs is characterized by its particularly stable properties. The method employed can identify wear itself in the image, regardless of the blade geometry, tool material, or tool coating.

[0152] The improved scheme can also be used to evaluate images recorded under different blade conditions.

[0153] Another advantage of the improved solution is its consistency, which is not only proven on the machine tool itself, but also enables connectivity to the cloud and thus global analysis.

[0154] The information about tool wear obtained through the improved scheme allows for a more accurate determination of the remaining tool life, and thus allows for optimization of product quality, or for reducing manufacturing costs by using the tools more effectively.

Claims

1. A computer-implemented method for determining the degree of wear of a tool (16), wherein, An image data set (10) providing a wear-related region (11) for a drawing tool (16) is characterized in that, by means of a calculation unit (12), in the case of using an artificial neural network (13), each image point (14) of a plurality of image points in the image data set (10) is assigned a category of a preset category set, wherein the category set contains at least one wear category; and at least one feature value of the degree of wear is determined based on the result of the assignment, wherein, for a column or row of image points among the plurality of image points, the calculation unit (12) determines an additional share of image points already assigned to at least one wear category; the wear mark width of the column or row is determined based on the additional share; and an additional feature value is determined based on the wear mark width.

2. The method according to claim 1, characterized in that, The neural network (13) is applied to the environmental region (15) of the corresponding image point (14) for each of the plurality of image points (14) by the computing unit (12) so as to assign the corresponding image point (14) to one of the categories.

3. The method according to claim 2, characterized in that, The environmental region (15) includes the corresponding image point (14) and all image points adjacent to the corresponding image point (14); or The environmental region (15) includes the corresponding image point (14) and all remaining image points, wherein the distance between the remaining image points and the corresponding image point (14) is less than a preset maximum distance; or The environment area (15) includes the corresponding image point (14) and all remaining image points. The row spacing between the remaining image points and the corresponding image point (14) is less than the preset maximum row spacing, and the column spacing between the remaining image points and the corresponding image point (14) is less than the preset maximum column spacing.

4. The method according to claim 1, characterized in that, The calculation unit (12) processes the allocated results according to the morphological image processing operation; and determines at least one feature value of the wear degree based on the processed results.

5. The method according to claim 1, characterized in that, The calculation unit (12) determines the share of image points assigned to at least one wear category; and determines the wear surface as a feature value of the wear degree based on the share.

6. The method according to claim 1, characterized in that, For at least one additional column or row of image points among a plurality of image points, the calculation unit (12) determines a corresponding additional share of image points assigned to at least one wear category; determines a corresponding additional wear mark width of the corresponding additional column or row based on the corresponding additional share; and determines additional feature values ​​based on the wear mark width and the additional wear mark width.

7. The method according to claim 1, characterized in that, The wear-related area (11) is imaged by a camera (17) to generate and provide a set of image data (10).

8. The method according to claim 1, characterized in that, The calculation unit (12) compares at least one feature value with at least one preset limit value; and determines a value related to the remaining usage time of the tool (16) based on the comparison result.

9. The method according to claim 1, characterized in that, The neural network (13) is trained or has been trained by a method in which a training image data set (18) is provided, the training image data set depicting a wear-related region (11) of a reference tool, characterized in that each image point (14) of a plurality of image points in the training image data set (18) is provided with a reference assignment to a category of a preset category set, wherein the category set contains at least one wear category; and the output of the neural network (13) is calculated for each image point by a training computation unit, and the output is compared with the reference assignment; and the neural network (13) is adjusted by the training computation unit according to the comparison result in order to train the neural network (13).

10. A computer program product having commands that, when executed by a computer system, cause the computer program to perform the method according to any one of claims 1 to 9.

Images