Method for determining a component quality of a component to be produced by powder bed-based melting

Abstract

The invention relates to a method for determining a component quality of a component (10) to be produced by powder bed-based melting, which component is formed by multiple successive layers (12).

Description

[0001] Description

[0002] Method for determining the component quality of a component to be produced by powder bed-based melting

[0003] The invention relates to a method for determining a component quality of a component to be produced by powder bed-based melting according to patent claim 1.

[0004] To ensure the quality of components manufactured using additive laser powder bed fusion (LPBF), it is necessary to identify relevant anomalies. This typically involves using emission data collected by in-situ melt pool monitoring (MPM) or optical tomography (OT). The collected data is typically processed to allow conclusions to be drawn about anomalies during an LPBF process. For example, the measured emission intensity data can be used to determine the presence of local overheating.

[0005] In a similar way, simulated distributions of process temperatures can be evaluated in order to identify expected process problems before the start of a construction process and, if necessary, to adapt parameters of a tool path or laser path.

[0006] The melt bath describes at least the area in the powder bed in which the powder is melted into a layer using, for example, a laser.

[0007] A common method for detecting anomalies is to set a threshold for monitoring signals and to mark values ​​that deviate from this threshold as anomalies. One problem when assessing an anomaly or process anomaly or predicted local overheating based on the simulation is that not all critical regions identified for a layer or layer lead to disruption of the production process of the component to be manufactured, in particular to the point of process aborts. Typically, too many supposed process problems are identified on the basis of individual layers and, accordingly, too many false alarms are triggered. If the job data is corrected on the basis of the detected alarms in the case of process simulation, this misidentification can lead to unnecessary corrections and thus, for example, to unnecessary process breaks and thus to an extended build time for the component.

[0008] Alternatively, an advanced method for processing in-situ monitoring data is the use of machine learning techniques. In principle, machine learning algorithms for detecting anomalies in LPBF processes offer the advantage of particularly automatable feature detection. Machine learning can thus help identify patterns and irregularities in complex LPBF processes. This requires collecting and preparing a large amount of monitoring data to train a machine learning algorithm. Collecting sufficient data with the right target features is crucial for the quality and accuracy of a machine learning model prediction, as the algorithm only finds patterns whose input data features are covered.

[0009] Nowadays, training data is collected using standard geometries such as cubes, rods or cylinders, whereby process parameters such as laser power, laser speed and hatch spacing can be varied. Based on such parameter variations and collected data, conclusions are drawn about the porosity or crack distribution of a selected process for manufacturing a component. When evaluating threshold values, particularly from monitoring or simulation data, the threshold value is applied to data from one layer to identify critical regions. Using a size threshold, regions deemed too small can then be filtered out, so that only critical regions above a threshold size are marked as relevant. This can at least reduce the number of false positive results.

[0010] Goh, GD, Sing, SL & Yeong, WY present an overview of machine learning in 3D printing. Artif Intell Rev 54, 63-94 (2021). Felix, S., Ray Majumder, S., Mathews, HK et al. also present in-situ process quality monitoring and defect detection for direct metal laser melting. Sci Rep 12, 8503 (2022).

[0011] The object of the present invention is to provide a method by which a component quality is determined or in particular predicted in a particularly advantageous manner, before or during the production of a component by means of powder bed-based melting, and in this case particularly efficient training data are used and / or anomalies not relevant for the component quality can be neglected in an advantageous manner.

[0012] This object is achieved according to the invention by the subject matter of the independent patent claims. Advantageous embodiments and further developments as well as associated advantages of the invention are set forth in the dependent patent claims, the description, and the drawings.

[0013] A first aspect of the invention relates to a method for determining a component quality of a component to be produced by powder bed-based melting, which component can be produced or is formed from a plurality of successive layers or plies. In the method according to the invention, at least one part of the successive layers of the component and / or of a model component is formed on the basis of at least one data set and / or on the basis of a modeling model, which, i.e. the data set and the model, are formed on the basis of layer data of a respective layer, or the layer data serve as input.

[0014] Powder bed-based melting is an additive manufacturing process in which material is joined layer by layer to create or generate workpieces or components from 3D model data. In powder bed-based melting, selective regions of the powder bed are joined or fused using thermal energy, particularly using a laser beam. The material used for the component is particularly metal and / or plastic. The powder bed itself or the powder generally forms a good thermal insulator. The data set contains, in particular, information about the criticality or anomalies of several combined layers. The modeling model is based in particular on a machine learning method and can make predictions about occurring anomalies based, for example, on the 3D model of the component to be manufactured.The layer data can be obtained, in particular, from measurements such as melt pool monitoring (MPM) or optical tomography (OT). Additionally or alternatively, the layer data can be obtained or generated by simulating the melting process. The model component can have a different shape from the component itself, but can also be the component itself and serves to capture layer data based on measurements.

[0015] In other words, the method according to the invention for determining component quality uses a data set which combines information from a number of successive layers which are to be produced by melting to create the component, in order to make statements about criticality or anomalies which occur during the manufacturing process. The layer data on which the data set is based can be both simulated and measured. Alternatively, the method according to the invention for predicting or determining component quality based on the layer data uses a modeling model which can estimate component quality based on the layer data, the modeling model making particular use of machine learning methods.

[0016] One advantage of the method, particularly in contrast to existing solutions, is that the layer data from several consecutive layers is used to assess the relevance of a critical region or to identify critical vectors, especially exposure vectors. This provides a further advantage that overheating, which leads to surface bulging and, in particular, to so-called melt beads, can be detected particularly reliably.

[0017] A further advantage is that the machine learning model can be particularly compact and requires only a few experimental training setups, while still enabling particularly precise results.

[0018] In an advantageous embodiment of the invention, measurement data for the layer data of the respective layer are recorded by means of a sensor device during production of the model component or the component, wherein the model component can be the component itself, and / or the layer data are calculated by means of a simulation. In other words, a sensor device comprising at least one sensor is provided which records data in particular during production of the component or model component, wherein the sensor data generated by the sensor device can include, for example, emission intensity data or long-term exposures from a camera. Alternatively, the layer data are simulated in particular using simulated distributions. The layer data are then derived from the measurement data or from the simulation.This has the advantage that, for example, layer data can be created in a particularly precise manner, since these are based on measurement data or simulation data.

[0019] In a further advantageous embodiment of the invention, a component model or a 3D model is used for the model component, the geometry or shape of which specifies different properties for a plurality of exposure vectors that can be used or are used in the production of the model component, so that there is a variation in the properties for the plurality of exposure vectors. The properties can in particular comprise a position, a vector length, a minimum mass integral, a vector orientation or further features. In other words, the model component is specified on the basis of a three-dimensional model - the component model - wherein the geometry or the shape of the component model is selected such that in particular for different layers their exposure vectors, which specify the movement of the laser in the layer in the powder bed, differ.For example, a layer of the component can be bounded by a triangular shape on its outer circumference, allowing the vector length of the exposure vector to be varied within the layer to be formed. For example, it is advantageous if different properties can be varied.

[0020] One property of the vector can be the position, in particular the position on the build plate and thus include the xy coordinates of the center of the exposure vector. A second property can be the vector length. Alternatively, the vector time can be a property. Another property is the minimum mass integral or an average mass integral, which is the maximum or average vector temperature, which can be determined by simulation or by a semi-analytical approach. The mass integral determines how much of a hemisphere or ellipse with a certain radius, typically 1 mm, is enclosed in a part. To determine the minimum mass integral, test points can be sampled on an exposure vector, for example, with a regular spacing of typically 0.1 mm. For such samples, mass integrals can be evaluated and their minimum calculated.The mass integral depends on the environment of the exposure point, in particular whether it is surrounded by a lot of molten material or a lot of powder, since the powder is a good thermal insulator compared to the molten material and can therefore lead to local overheating. Another property can be the vector orientation. Additional optional features can be process parameters such as laser power, laser speed, hatch spacing and so on, which can also be taken into account. By generating a model component that covers as many exposure vectors as possible or a large variation in their properties, there is the advantage of achieving a large spread of the model input parameters in order to obtain advantageous training data for the modeling model.

[0021] In an advantageous embodiment of the method, the measurement data recorded during production of the model component and the component model, which in particular can also include information about the exposure vectors or a possible production, are used as training data and / or as validation data for a machine learning method which provides the modeling model or forms the modeling model. In other words, the data of a model component which has a particular geometry is used as training data in order to thus record a large variation in exposure vectors. For this purpose, this model component is produced on the basis of the component model by powder bed-based melting and, in the process, measurements obtained for at least some of the layers and layer data or measurement data are used for the layer data.This data and the information from the component model are combined and can be used as training data. The machine learning method or the modeling model can be, for example, a self-learning algorithm and / or a neural network. This offers the advantage that, based on the shape of the component model and the measurement of the model component produced using the component model, particularly advantageous measurement data can be generated. Thus, the training data can be used particularly advantageously to determine component quality.

[0022] In a further advantageous embodiment of the invention, the measurement data is referred to as an intensity distribution or emission intensity distribution which is determined by means of melt pool monitoring (MPM) - in particular in-situ melt pool monitoring - and / or long-term exposure images for optical tomography. The long-term exposure images correspond in particular to an exposure image which is exposed for each layer or position during formation. In other words, process emissions from the melt pool are monitored, in particular automatically, with at least one sensor detecting the emission from the melt pool, in particular with spatial resolution for each position or layer. If the melt pool is too cold or overheated, this can thus be detected in the measurement data.In optical tomography, an optical image is captured using long-term exposure, for example, using a camera per slice, so that a light distribution for the ensemble of exposure vectors forming the slice can be recorded. For example, using a tomography algorithm, a multi-layer image can be created. This offers the advantage that the measurement data can be used particularly advantageously for carrying out the procedure.

[0023] In a further advantageous embodiment of the invention, the layer data describe a criticality. In other words, a measurement result or the measurement data, such as the intensity distribution, are reinterpreted using a classification in the layer data. The classification is used to assign the criticality so that an assessment of the risk of an anomaly at different positions in the respective layer of the component can be made. For example, a corresponding measuring point in the respective layer can be classified in such a way that it is described or classified as subcritical, semi-critical or critical for the formation of the component and thus its quality. This has the advantage that the layer data can be used particularly advantageously for determining the component quality.

[0024] In a further advantageous embodiment of the invention, for the layer data, several layers, in particular up to 10 and preferably 3 to 5 layers, are arithmetically and / or logically combined and at least one anomaly for the component is derived from this. The anomaly can describe the component quality. In other words, information from several layers is combined either arithmetically, i.e. for example using statistical methods such as mean value, and / or logically, for example based on criticality. If, for example, in three consecutive layers there is a particularly high intensity at an xy coordinate in only one layer and a normal intensity can be found at the same xy coordinate in the other two layers, the arithmetic mean can be used, for example, to conclude at this point that no relevant anomaly is present.The same applies to the logical summary: if, for example, three layers with the same xy coordinate are only shown as critical once and subcritical twice, it can also be concluded that an anomaly is not present. This has the advantage that the process for determining component quality can be carried out with particularly high stability.

[0025] In a further advantageous embodiment of the invention, critical regions are formed from the layer data on the basis of point data and / or critical vectors are formed from point data and / or critical vectors are formed from a vector criticality, and the component quality is derived therefrom. In other words, a critical region is formed in the multiple layers or layers from point data that are distributed along a respective exposure vector, wherein individual layer criticality bitmaps can be created. Additionally or alternatively, critical regions can be determined for multiple layers by omitting a contour search, but calculating the overlap with the critical regions for the exposure vectors of a layer, and marking the vectors or exposure vectors as critical if their overlap threshold is exceeded.Additionally or alternatively, single-layer vector criticality indices are created, whereby pixel values ​​of single-layer bitmaps can be created along the vector. This offers the advantage of being able to answer various questions regarding the formed layers, for example, whether regions or vectors or exposure vectors are critical. This makes the method particularly advantageous for determining component quality.

[0026] In a further advantageous embodiment of the invention, the model component is produced several times to fill the build space for the training data, in particular in different orientations. In other words, several variants of the component model are arranged in the available powder bed in such a way that they are printed individually and free-standing, yet fill the build space or powder bed as completely as possible. With appropriate optical tomography or in-situ melt bed monitoring, measurement data can be recorded for many positions and many vectors or exposure vectors, so that the model or modeling model can be trained particularly advantageously. This has the advantage that the component quality can be estimated particularly advantageously using machine learning.

[0027] In a further advantageous embodiment of the invention, the component model uses scaling factors. In other words, the component model is provided in such a way that different layers or plies are scaled differently, even though they essentially have the same shape. For example, shapes of particular interest for the mass indices can be created at edges, for which measurement data can thus be obtained in a particularly advantageous manner during the manufacture of the model component, in order to enable particularly advantageous predictions about the component quality.

[0028] For use cases or application situations that may arise during the method and which are not explicitly described here, it may be provided that, in accordance with the method, an error message and / or a request to enter user feedback is issued and / or a standard setting and / or a predetermined initial state is set.

[0029] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identity are included.

[0030] It shows: FIG 1 schematic side view of a device for powder bed-based melting for the production of a component;

[0031] FIG 2 schematic perspective view of a component model which can be produced as a model component by the method;

[0032] FIG 3 perspective view of a powder bed of the device with several copies of the component model according to FIG 2 arranged in the powder bed of the device according to FIG 1;

[0033] FIG 4 top views of two layers with associated exposure vectors of the component according to FIG 2 ;

[0034] FIG 5 Top view of a layer of another component model; and

[0035] FIG 6 perspective view of a component model which has a layer according to FIG 5 .

[0036] With reference to the figures, a method for determining or assessing a component quality of a component 10 to be produced by powder bed-based melting, which component can be produced or is formed from several successive layers 12, is to be presented.

[0037] To produce the component 10, a device 16 having a powder bed 14 is used, which device has a light source 18, which can in particular provide a laser beam 20 that is moved along an exposure vector 22 in order to selectively melt the powder 24 contained in the powder bed 14 at a specific position in the powder bed 14, whereby after the laser beam 20 has traversed the exposure vector 22, a layer 12 or layer of the component or of a model component 26 is produced or formed. For a further layer 12, the build platform with the already produced part of the component 10 is then lowered and a new layer is applied or melted thereon using new powder 24. The base of the powder bed can in particular be formed by a build plate 32.

[0038] The powder 24 can in particular be a metallic material and / or plastic, wherein it can be mixed with other materials, such as a ceramic.

[0039] In the method for determining the component quality of the component 10, the component quality is determined as a function of at least one data set and / or as a function of a modeling model, which is formed on the basis of layer data of a respective layer 12 of at least a part of the successive layers 12 of the component 10 and / or the model component 26.

[0040] In other words, the data set is provided in particular by the layer data. Additionally or alternatively, the layer data can be used as input for the modeling model, which is provided in particular by means of a machine learning method, such as a self-learning algorithm and / or a neural network, wherein additional layer data can be created as training data for the model or modeling model. Thus, on the basis of the training using the training data, which is based on layer data, for example, of the model component 26, the modeling model can make predictions about the component quality of the component 10. For this purpose, the trained model receives a three-dimensional model of the component 10 as input.

[0041] Advantageously, the layer data of the respective layer 12 can be formed during the production of the model component 26 and / or the component 10 from measurement data recorded by a sensor device 28 for the respective layer 12. Additionally or alternatively, the layer data can be calculated or simulated by a simulation.

[0042] Measurement data used in the method can advantageously be an intensity distribution of a melt pool monitoring (MPM) and / or long-exposure images for optical tomography.

[0043] Based on the data set, which in particular comprises the measured or simulated layer data, the component quality can be estimated by, for example, evaluating the relevance of critical regions contained in the sensor and / or simulation data or layer data. The evaluation is carried out by combining a configurable number of consecutive layers 12 or their layer data. For this purpose, the measured values ​​or measured data and / or criticalities of typically fewer than 10 layers or layers 12, preferably 3 to 5 individual layers or layers 12, can be combined. With this combination, an arithmetic operation can be carried out, such as averaging the measured values ​​of the individual layers 12 or the individual layer measured values, or even a logical combination of individual layer criticality indices.Thus, the layer data of the respective layer 12 can describe a criticality. The criticalities can be represented, for example, by a two-dimensional distribution, in particular in the form of a bitmap with a configurable list, so that the corresponding layer data for each layer 12 is available as this bitmap. Thus, each pixel can correspond to the criticality at a point in a 2D area (typically lateral dimensions of the component 10).

[0044] Preferably, a multi-layer criticality is calculated from the mean of the individual layer measurements. For example, a multi-layer criticality bitmap can be generated or calculated based on a threshold comparison of a mean distribution. If the individual layer criticalities are used as input data, a count of the individual layer criticalities can be used. For example, a value in the spatial multi-layer criticality distribution is set to one if at least two of the individual layer criticalities considered are set as critical.

[0045] In order to be able to estimate the component quality based on anomalies, which may be caused, for example, by an excessively high melting temperature caused by an exposure vector 22, critical exposure vectors 22 are identified—and may need to be corrected. For example, an overlap of the exposure vectors 22 of a layer 12 or position with the critical regions can be determined. A contour search with optional size filtering can be performed to identify critical regions.

[0046] Overall, several cases can be distinguished for the detection of a multi-layer critical region:

[0047] In the first case, detection of multi-layer critical regions can be performed from point data.

[0048] Based on data points distributed along the exposure vectors 22, for example based on MPM intensity values ​​or simulated temperatures, individual layer criticality bitmaps are created by setting pixel values ​​of a bitmap according to the associated measured values. Raw values ​​of the measured data can be used directly, and criticality indices, such as subcritical, semi-critical, or critical, can be determined, which can then be used as logical data generated from the pipe data via a threshold comparison. During bitmap generation, care should be taken to ensure that no undefined pixels remain within exposed areas (particularly per layer 12), since otherwise holey areas could arise when the individual layer bitmaps are subsequently linked.This can be achieved, for example, by a sufficiently low bitmap resolution and thus by larger pixels or a bitmap dilation process, for example, with a morphological close. The multi-layer criticality bitmap is then created from the individual layer bitmaps using an arithmetic or logical operation. A subsequent contour search, which describes, for example, a length or circumnavigation of an area, and optional size filtering can then describe critical regions or anomalies in the multi-layer data relating to several layers.

[0049] In the second case, multi-layer critical vectors are detected from point data.

[0050] In this case, multi-layer critical regions are determined analogously to the first case, although the contour search is not required. The overlap with the critical regions is then calculated for the exposure vectors 22 for a layer or slice 12, and the exposure vectors 22 are marked as critical if the overlap threshold is exceeded. For performance reasons, the overlap can be calculated using an electronic computing device by counting critical, equidistant sampling points on the vectors or exposure vectors 22. Additionally or alternatively, the length of a vector polygon overlap can also be determined if the critical regions are previously determined by a contour search.

[0051] Thus, performing the method according to the first case can determine the regions and according to the second case, for example, answer the question of how long the corresponding exposure vector 22 runs on critical points, and thereby, for example, determine whether the vector itself is critical.

[0052] Additionally or alternatively, according to a third case, a detection of multi-layer critical vectors from vector criticalities can be carried out.

[0053] In this case, single-layer vector criticality indices (e.g., uncritical and critical) can be used as input data.

[0054] Pixel values ​​of the single layer bitmaps are set accordingly along the vector or exposure vector.

[0055] After an optional dilation step, the bitmaps are combined analogously and / or logically to form a multi-layer criticality bitmap. Subsequently, as in the second case, the overlap in the exposure vector of a layer is determined. The vector multi-layer criticality is determined analogously to case 2 by comparing the threshold values ​​of the overlaps.

[0056] In the cases described, the corresponding layer data or the corresponding individual multilayer bitmaps are used as the data set, depending on which the component quality can be determined, since the data sets mentioned can be used to determine, for example, critical regions or vectors.

[0057] To evaluate the relevance of a critical region or to identify critical vectors of measurement simulation data or data from a simulated measurement (e.g., temperature), several consecutive layers 12 are used. This allows particularly reliable detection of overheating that leads to bulging of the surface or melt beads of the component 10, in particular, more reliably than is possible with separate evaluation of the data from individual layers. Thus, the presented method is particularly advantageous for determining component quality.

[0058] Additionally or alternatively, the modeling model can be used, which is provided in particular by or by means of a machine learning method. In this case, a component model 30 can be used for the model component 26, the geometry or shape of which can provide for a variation of the exposure vectors 22 used in the production of the component model 26, so that different properties are present for the model component 26 with multiple exposure vectors 22.

[0059] The model component 26 or the component model 30 serves in particular as a reference component and, due to its shape, is designed so that feature values ​​for the machine learning algorithm and / or the neural network are distributed as evenly as possible. Thus, a feature list used for the machine learning algorithm can first be described. The model component 26 can be manufactured to fill the installation space, for example, using copies of itself, particularly for the training data.

[0060] Each of the layers 12 of the model component 26 is produced by an exposure vector or a set of exposure vectors 22. FIG. 4 shows, for example, the exposure vectors 22 of two layers 12. The features or properties of the exposure vectors 22 should advantageously vary, in particular, between the layers 12.

[0061] A first of the properties to be varied can describe a position on a build plate 32 which forms the bottom of the powder bed 14. The build plate 32 can, for example, be assigned an xy coordinate system, and a property of the respective exposure vector 22 can represent a position or path on the build plate 32, wherein the center of the exposure vector 22 or, alternatively, coordinates of the vector start and end can be used to describe the "position" property. In the case of a round build plate 32, the position can be used, for example, in polar coordinates instead of xy coordinates.

[0062] A second property can be the vector length or, alternatively, a vector time or a minimum or maximum repetition time. The vector time is the time required to expose an exposure vector 22. The minimum repetition time is the time between the vector start and the previous exposure time of the point closest to the vector beam. The maximum repetition time can be the time between the vector end and the previous exposure time of the point closest to the vector end.

[0063] A third property can be the minimum mass integral, alternatively a maximum or average mass integral or an average vector temperature, which can be determined in particular by simulation and using a semi-analytical approach. The mass integral can determine how much of a hemisphere or ellipse with a certain radius, typically approx. 1 mm, is enclosed in the already exposed component. To determine the minimum mass integral, for example, an exposure vector 22 can be sampled at test points with a regular spacing of typically 0.1 mm. For these samples, the integrals are evaluated in each case and their minimum is calculated.

[0064] A further property could be a vector orientation or, alternatively, a stripe orientation. The vector orientation is the angle between an exposure or illumination vector 22 and the abscissa. The stripe orientation can represent the direction of an exposure vector sequence. Furthermore, there are other features, such as process parameters, for example, a laser power, a laser speed, or a hatch spacing. In particular, optional features can be used if different sets of process parameters need to be evaluated and / or combined.

[0065] Process parameter sets can, for example, include measurement data recorded by the sensor device 28.

[0066] FIG 2 shows a model component 26 or the three-dimensional component model 30 on which the model component 26 is based, which can have a particular variation in the just outlined properties of the exposure vectors 22 that can be used for its production due to its advantageous geometry.

[0067] The component model 30 of FIG. 2 thus has an advantageous geometry for machine learning training; for this purpose, it is divided into seven sections 34 along its height. The number of sections 34 is arbitrary and merely exemplary. Thus, the model component 26 to be produced from it can be manufactured with more or fewer sections 34.

[0068] The height of each section 34 is in particular equal to a layer thickness of a respective layer 12 to be produced by the device 16 multiplied by the number of different layer orientations to be examined or used for the layer data. -sect -layer ' ^orient

[0069] The order of the strip orientations can be identical in each section 34. This order can be chosen so that the orientations are evenly distributed within the sections 34. A practical choice of the number of different layers and orientations can be, for example, n or ient=36 or n or ient=72 or another sufficient size (>18) integer divisor of 360. The order of the stripe orientation in each section can be constructed as follows:

[0070] Alternatively, a different regular distribution of the orientation angles can be used. One condition for the distribution should be that two consecutive layers 12 differ sufficiently in orientation and, in particular, are shifted or rotated relative to each other by more than 45°, for example.

[0071] For the training geometries of the model component 26 for the training data, in particular, the same hatching pattern that is to be used for the component 10 can be used. Depending on the component size, stripes 5 to 10 mm wide or checkerboard patterns of a similar size are generally used. In the example of FIGS. 2 to 4, for example, a stripe pattern 10 mm wide can be used.

[0072] The lower section 34 of the geometry of the component model 30 can be designed as a cylinder. Advantageously, the cylinder can have a diameter of, for example, 12 to 16 mm. Cylinder bases that are too large lead to long exposure vectors 22 and to the generation of too many stripes of similar size.

[0073] Conversely, if the cylinders are too small, no long exposure vectors 22 are covered during training. The goal of the advantageous geometry of the component model 30 is to cover the aforementioned features or properties as evenly distributed as possible across the layers 12 and across the entire geometry.

[0074] In the example shown in FIG 2, the diameter of the cylinder is 15 mm.

[0075] The second section 34, i.e., the first one above the cylinder, and the subsequent sections 34 are formed, in particular, from triangular prisms rotated relative to one another. The base of each triangular prism is a regular triangle inscribed in the lower circle (circumference of the cylinder). Alternatively, other regular polygons or other shapes can be used instead of the triangle, as long as a wide spread of the lengths of the exposure vectors 22 is achieved. The triangle does not have to be inscribed exactly in the circle. Slightly smaller triangles are also suitable.

[0076] Each section thus comprises twisted prisms or triangular prisms, which are stacked on top of each other in descending order of twist angle. The twists ensure that the feature of the minimum mass integral can be uniformly covered in the training data. For example, with nickel-based alloys, overhang angles of up to 40° can be produced without overheating. Therefore, the twist angles of the upper six sections 34 in FIG. 2 are chosen as follows: 90, 80, 70, ..., up to 40°.

[0077] To adjust process parameters or materials that are less prone to overheating, profiles with even lower minimum overhang angles can be used.

[0078] As mentioned above, the number of sections 34 is arbitrary. Other descending sequences up to 40° can also be used. The order of the twist angles is descending because, in the event of overheating, only the upper sections 34 are affected and can therefore be easily excluded from the training dataset. In general, materials prone to overheating require fewer sections 34 of twisted prisms (the upper section 34 has a larger twist angle).

[0079] Alternatively, instead of sections 34, a continuous function can be used which represents the angle of rotation.

[0080] FIG. 3 shows a distribution of several model components 26 on the build plate 32. One possibility for the advantageous distribution is to additionally rotate every second component model 30 around its center point, i.e., neighboring geometries have different orientations. In the case of triangles, a rotation of 60° can be advantageous.

[0081] FIG 3 shows the rotation of nine adjacent model components 26, whereby more than two different rotations or orientations can be used in order to achieve a favorable distribution over the building plate 32.

[0082] Figure 4 shows an example of a stripe pattern in two adjacent parts with a 60° rotation, where the lines represent the exposure vectors 22. A rotation of 60° serves to better distribute the vector set features.

[0083] The model components 26 are distributed as evenly as possible, in particular such that the largest possible area on the build plate 32 is covered with the model components 26, thus allowing a particularly small distance to be maintained between the adjacent model components 26. The typical small distance between the parts in a powder-bed-based melting process can be between 5 and 15 mm.

[0084] In addition, the edges and corners of a rectangular building panel 32 should be covered whenever possible. However, some corners cannot always be covered, for example, if the building panel 32's fasteners are nearby.

[0085] For a round building plate 32, it is recommended to cover the area around the plate boundary as close as possible in order to obtain advantageous training data.

[0086] For the training data, it can be particularly advantageous to produce the model component 26 several times, in particular in different orientations and filling the installation space.

[0087] Depending on the fill level of the build plate 32, areas where no model component 26 is formed in the build space or powder bed 14 may need to be extrapolated for the training data. Due to the mandatory gaps between the model components 26, there will always be some uncovered xy positions. If the data collected from the measurements cannot be interpreted well enough, the build plate 32 can be refilled with model components 26, which are built in the previously existing gaps in order to compensate for the training data for each xy position.

[0088] If multiple process parameters, such as laser power, laser speed, and / or hatch spacing, are to be investigated, the build or manufacturing process should be repeated with different parameter sets. In contrast, downskin / upskin parameters and the corresponding exposure strategies can be integrated into the same build or manufacturing process. Figures 5 and 6 show a further advantageous geometry of a component model 30, with Figure 5 specifically illustrating a layer 12 that is used in multiple levels or layers 12 of the component model 30 by means of a scaling factor.

[0089] The shape or geometry of the embodiment of the model component 26 of FIGS. 5 and 6 can be particularly useful for determining the correctness of the prediction of the machine learning algorithm on which the modeling model is based, for example.

[0090] The component model 30 is constructed similarly to that of FIGS. 2 to 4, except that the lower cylinder is formed with a larger diameter, for example, 23 mm, to achieve the star-shaped structure. Alternatively, the cylinder can be omitted, and the lower part begins directly with a star-shaped section 34, of which FIG. 5 shows a layer 12.

[0091] The star-shaped sections can be rotated analogously to the triangular prisms of FIGS. 2 to 4. Furthermore, a scaling factor can be applied from bottom to top. The scaling factor, or growth scaling factor, can correspond to the same angle chosen for the rotation in each section 34, i.e., mathematically, this factor can be calculated as follows:

[0092] Alternatively, the scaling factor can also be fgrowth =1 . Or only large scaling factors can be used without rotation, so that the component simply has a straight star-shaped geometry .

[0093] A circled octagon can be chosen as the basic geometry or shape in the lower cylinder, as shown in FIGS. 5 and 6. Alternatively, another angular polygon could have been chosen.

[0094] The polygon or octagon forms a hole in the center, which represents an area of ​​the respective layer 12 that is generally of secondary relevance for assessing component quality. In particular, the hole serves to save material.

[0095] Since the information obtained from the melt pool monitoring from the center of the structure is not essential, the model component 26 serves to detect the edges and any anomalies occurring there in the outer region. In another variant of the model component 26, only one or a few beams of the star shape can be produced in order to also save material, particularly in the form of the powder 24.

[0096] The number of sections can also be chosen arbitrarily for this form of the component model 30. The exemplary embodiment of FIGS. 5 and 6 has five sections 34 with a twisted star shape and growth scaling factor.

[0097] In contrast to the embodiment of the model component 26 according to FIGS. 2 to 4, the angles for the rotation are chosen more aggressively.

[0098] The twist angles at the top of the structure are therefore less than 40°. A nickel-based alloy can be used for the material, for which the following twist angles can be selected: 80°, 60°, 45°, 35°, and 25°.

[0099] The resulting geometry serves as a good control for the output of the machine learning algorithm and / or the neural network , i.e. the modeling model , since it contains both (lower) sections 34 that can be built without anomalies but especially also (upper) section 34 where overheating anomalies can occur.

[0100] Here, too, it may be practical to print several model components 26 distributed across the build plate 32. The rules for covering the build plate 32 may be less strict than in a training phase in which the model component 26 and component model 30 are used according to FIGS. 2 to 4. Thus, the model components 26 can be placed virtually arbitrarily on the build plate 32.

[0101] Thus, the determination of component quality is shown here using the method when the modeling model is used. Using the method presented, the component quality can be advantageously determined using at least one data set that includes layer data, i.e., also using a modeling model that also includes layer data. This advantageously prevents, for example, faulty manufacturing of the component 10.

[0102] The method is used both for multi-layer criticality assessment for quality assurance of LPBF processes and for the design of reference models for the detection of anomalies with in-situ monitoring in laser powder bed fusion processes.

[0103] Reference symbol list

[0104] 10 components

[0105] 12 layer 14 powder bed

[0106] 16 Device

[0107] 18 Light source

[0108] 20 laser beam

[0109] 22 Exposure vector 24 Powder

[0110] 26 model components

[0111] 28 Sensor device

[0112] 30 component model

[0113] 32 Component plate 34 Section

Claims

Patent claims 1. Method for determining a component quality of a component (10) to be produced by powder bed-based melting, which component is formed from several successive layers (12), depending on at least one data set and / or depending on a modeling model, which are formed on the basis of layer data of a respective layer (12) of at least a part of the successive layers (12) of the component (10) and / or of a model component (26).

2. Method according to claim 1, characterized in that for the layer data of the respective layer (12) during the production of the model component (26) measurement data are recorded by means of a sensor device (28) for the respective layer (12) and / or the layer data are calculated by a simulation.

3. Method according to claim 2, characterized in that a component model (30) is used for the model component (26), the shape of which component model predetermines different properties for a plurality of exposure vectors (22) used in the production of the model component (26), so that a variation of the properties is present for the plurality of exposure vectors (22).

4. The method according to claim 3, characterized in that the measurement data acquired during the production of the model component (26) and the component model (30) are used as training data and / or as validation data for a method of the machine learner which provides the modeling model.

5. Method according to one of claims 2 to 4, characterized in that an intensity distribution of a melt pool monitoring and / or long-term exposure images for an optical tomography are used as the measurement data.

6. Method according to one of the preceding claims, characterized in that the layer data describe a criticality.

7. Method according to one of the preceding claims, characterized in that for the layer data, several, in particular up to 10 and preferably 3-5, layers (12) are arithmetically and / or logically combined and from this at least one anomaly for the component (10) is derived, which describes the component quality.

8. Method according to one of the preceding claims, characterized in that critical regions are formed from the layer data on the basis of point data and / or a critical vector is formed on the basis of point data and / or a critical vector is formed on the basis of a vector criticality and the component quality is derived therefrom.

9. Method according to one of claims 2 to 8, characterized in that for the training data the model component (26) is produced several times in particular in a different orientation to fill the installation space.

10. Method according to one of claims 3 to 9, characterized in that the component model uses scaling factors.

Images