Computer-implemented method for determining defects in an object manufactured by an additive manufacturing process

Description

[0001] The invention relates to a computer-implemented method (claim 1) and a computer program product (claim 13) for determining defects of an object produced by means of an additive manufacturing process, wherein volume elements or spatial coordinates of the object to be produced are processed during the additive manufacturing process.

[0002] In additive manufacturing, an object is typically built up layer by layer, for example, on a substrate. Several processes are known to achieve this. These include selective laser melting (SLM), electron beam melting (EBM), selective laser sintering (SLS), binder jetting, fused deposition modeling (FDM), and stereolithography.

[0003] In the laser melting process, for example, a powder bed of fine metal particles with a flat surface is prepared and melted at defined points using a laser beam, after which it re-solidifies. This process creates a bond with adjacent powder material. The laser can trace desired paths on the surface, which can be arbitrarily complex. A layer of the desired object is created in the areas traversed by the laser. Then, another layer of powder is applied, and the desired areas within this additional layer are melted. This creates a bond with the layer below. Consequently, the object is built up layer by layer with further layers of powder.

[0004] In this process, information about the geometry to be manufactured, derived from a CAD model, is used to determine the relevant areas to be melted for each layer. Since this is performed across numerous layers, it is done for 3D coordinates (or, in other words, volume elements or voxels). After the final layer has been manufactured, the component can be removed from the powder bed. The unmelted powder and any support structures are removed, revealing the desired geometry, which can then be used as a finished component or subjected to further manufacturing steps (e.g., surface finishing).

[0005] The additive manufacturing process is comparatively complex. A wide variety of parameters directly influence the geometry of the manufactured component, including grain size, shape and surface properties of the powder used, layer thickness, laser power, laser spot diameter and path, powder bed preheating temperature, geometry orientation within the powder bed, support and auxiliary structures used (e.g., for geometric stability and heat transfer), and many more. Therefore, the manufactured geometry often does not correspond to the desired geometry of the specified 3D coordinates at which the powder was nominally melted. In particular, thermal effects can lead to deformations of the overall geometry or local surface deviations. These effects, or rather,Deformations can be partially simulated in advance or determined by measuring the geometry of manufactured prototypes. This information can then be used to correct the manufacturing process. One possibility is to adjust the geometry of the coordinates or volume elements to be melted according to the deviations from the target geometry, so that these deviations are minimized in the next build process or iteration. Various defects, such as pores, cavities, microstructure disruptions, porosities, cracks, delaminations, or inclusions, can also occur within the material. These defects are usually undesirable because they negatively affect the mechanical properties.

[0006] Against this background, the present invention aims to overcome the aforementioned disadvantages of the prior art and to provide an improved method for determining defects in an object.

[0007] The main features of the invention are specified in claims 1 and 13. Embodiments are the subject of claims 2 to 12.

[0008] This creates a method that allows potential defects in an object to be detected directly from process data generated during additive manufacturing. In a preparatory step, the positions of defects in one or more objects are identified in the measurement data of the objects. These positions can be individual coordinates, but also clearly defined volume areas or geometries. The measurement data can be generated using standard imaging techniques and subsequent analysis. The resulting process data of the objects being examined using measurement data is referred to here as initial process data. The positions of the defects, defined by the defect coordinates, are correlated with the project data coordinate system of the initial process data. Subsequently, the initial process data, along with the correlated coordinates, is stored as training data.The training data enables the machine learning algorithm to learn and recognize characteristic features present in the initial process data. These features correspond to the identified defects or defect artifacts, as well as to defect-free areas or material regions. Subsequently, the machine learning algorithm can identify defects in the process data generated during the production of further objects. The process data from these subsequent objects is referred to here as the second set of process data.

[0009] In a first process step according to the invention, spatially resolved first process data of n objects are determined, wherein the first process data are acquired during an additive manufacturing process for the production of the n objects, and wherein the first process data define a process data coordinate system for each of the n objects.

[0010] Process data, as defined in the invention, refers to data that can be directly acquired during the additive manufacturing process and provides information about various parameters. Due to the successive material accumulation or solidification along geometrically defined paths or layers following a known temporal progression, the process data is already spatially resolved when regularly or continuously recorded during the manufacturing process. In particular, the process data could be present in a regular grid, such as a raster of individual volume elements (voxels), or be transformed into such a raster. The process data can represent a wide variety of physical properties, the nature of which depends primarily on the acquisition method.

[0011] Process data can include, for example, an optical measurement parameter. If a laser-based manufacturing process is used, such as SLM or SLS, a camera or photodiode could be positioned in the laser beam path to detect light reflected back through the material. It could be useful to examine each individual volume element irradiated with the laser and detect the light power reflected back by the material. Suitable optics could ensure that only the emissions from the heated material are measured, not reflections from the laser itself. Such a measurement is advantageous because the optical properties of a material section can depend on its quality. For example, if fluctuations in the reflected light power are detected in a particular area, this could indicate that defects have occurred there.

[0012] Simultaneous optical acquisition of the entire layer during the build process would also be conceivable, for example, using a camera. Additionally, laser-induced ultrasound could be used for acquisition. Alternatively or additionally, after a layer of the object has been manufactured, its geometry can be measured, for example, by structured light projection, or an optical image of the layer can be taken. However, other variations for performing accompanying measurements and acquiring process data are also conceivable. It should be noted, however, that the precise accuracy of the process data acquisition and the specific physical quantities recorded are irrelevant. What is important for the method according to the invention is that spatially resolved process data is available.

[0013] It is important to note that the process data initially exists in the target geometry of the component to be manufactured or the coordinates to be melted, as it is assigned to the respective current position in the manufacturing process. Deformations have not yet occurred at the time the individual measurements are taken. Consequently, the initial process data, or the process data in general, do not represent the exact actual geometry of the manufactured object.

[0014] To prepare for training the machine learning algorithm, the aforementioned training data is created. This begins with the additive manufacturing of a specific number n of objects. The number n can be chosen as needed, particularly based on the reliability and repeatability of the manufacturing process.

[0015] If a defect occurs only in exceptional cases at any point on an object, the number n can be chosen to be correspondingly high, so that a representative number of defects can be correlated with process data. If, due to particularly high demands on component quality, especially fine structures, and a special material, local defects are to be expected more frequently, the number n can also be chosen to be lower. The goal is to produce a representative number n of objects that are tested for defects non-destructively using known methods.

[0016] For the purposes of the invention, a defect is understood as a condition in which the manufactured object does not exhibit the desired property intended for the object at a specific location. This could, for example, be a local deviation in the external geometry of a spatial section of the manufactured object. A typical example of this would be a local deviation of the surface of the spatial section from the intended geometry. If the manufacturing process is based on the selective melting of material, for example, too much material might have been melted in the area of ​​the surface of a section. In particular, this could include relatively small-scale deformations, such as dents or burrs.

[0017] In another example, the deviations can also include local variations in the internal properties of the spatial section. This can refer to existing pores, cavities, outgassing, structural loosening, porosity, cracks, delaminations, or inclusions of air, another gas, raw material, or a vacuum.

[0018] The examination of the n objects to determine measurement data includes the use of a computed tomography scanner.

[0019] This is followed by a direct, pixel- or voxel-wise identification of areas affected by defects in order to determine the positions of the defects. The positions of the defects are represented by the defect coordinates. Furthermore, the determination of the edge lines between defects and the material is conceivable. The investigation can, in particular, include an analysis of gray value, gradient, and shape properties. The defect coordinates can therefore contain information about all volume elements of the manufactured object that exhibit defects, whereby defect attributes can be assigned to individual volume elements or groups of volume elements.

[0020] The findings from the examination of the n objects can be used effectively when correlated with the initial process data generated during their manufacture. This initial process data can be quite complex, and defects might manifest themselves in various ways as a likely very small subset. Irregularities in the initial process data do not necessarily correspond to a defect in the manufactured object, as, for example, a defect occurring in one layer could be eliminated by melted material from an overlying layer. Consequently, correlating the defect coordinates with the initial process data can be used to collect training data for a machine learning algorithm. This algorithm can then be used to examine future (second) process data.

[0021] Machine learning algorithms are able to recognize or determine patterns or rules from multiple examples of a problem, which together form a generalization. Suitable machine learning algorithms can include, for example, artificial neural networks, deep learning or machine learning in general, or template matching. Conventional image processing or pattern recognition algorithms can also be used, where certain parameters can be adjusted to optimize them for the best possible results with the training data.

[0022] In addition to the algorithms mentioned above, methods such as "Support Vector Machine", "Random Forest", or simple regression can also be used as machine learning algorithms. Training the machine learning algorithm using the training data can then be considered largely complete when, for example, the algorithm reliably detects in a test run, using the initial process data, whether or not defects are present in a given area.

[0023] It is important to note that the training data must be reusable for a specific configuration of the object to be manufactured under consistent manufacturing conditions. If fundamental properties of the object or the manufacturing conditions are changed, new training data should be collected.

[0024] For the training process, training data from another source of process data, for which suitable defect detection is already possible, could also be used. This information could then be used to train the defect detection of the initial process data.

[0025] In summary, the method according to the invention allows a self-learning algorithm to be trained to perform non-destructive testing of additively manufactured objects for potential defects during or after their production. Directly after manufacturing or even during the manufacturing process, the self-learning algorithm can assess whether critical defects are present in the object.

[0026] According to an example of the invention, the step of analyzing the second process data for defects using the machine learning algorithm can be performed during the acquisition of the second process data. The machine learning algorithm can analyze the second process data as it is being acquired during the manufacturing of the object using the additive manufacturing process. This allows defects generated at the beginning of the object's production to be detected promptly during the further manufacturing process using the machine learning algorithm. If the criticality of the defect requires it, the object can be rejected during production, i.e., the production process can be prematurely terminated. This saves time and material, since objects with critical defects detected during production can be rejected even in an incomplete state.Furthermore, it is conceivable to use the information about the defects to adjust the manufacturing parameters for the current or subsequent manufacturing processes in such a way that further defects can be avoided or existing defects can be repaired.

[0027] According to an example of the invention, during the manufacturing of the object, a physical measurement for spatially resolved process data is acquired for each volume element or spatial coordinate processed during the manufacturing process. This applies to both the first and second process data. The process data are therefore essentially generated continuously during the manufacturing of the object, and the determination of the first and second process data can thus be optimally integrated into the manufacturing process. As mentioned above, the laser can be used to perform a desired measurement, particularly in laser-based manufacturing processes.

[0028] During the manufacture of the object, at least two different physical measurements can be recorded for each volume element or spatial coordinate processed during the manufacturing process, representing the spatially resolved first and second process data.

[0029] By using different measurement parameters, a measurement of one parameter can be supported by the measurement of a second. Sometimes, different types of defects may occur, which are particularly well detected with one parameter but not with the other. Combining the acquisition of both parameters can improve defect identification. For example, a photodiode could be used to detect laser beams reflected from molten material. This allows for the determination of brightness, or at least the reflectance, during a specific time period. Simultaneously, a camera can be used, which captures a significantly larger amount of information and provides at least a spatially resolved gradation of brightness as process data for a specific section of the object.It is not necessary for the first and second process data to have the same measured quantities. The measured quantities of the first and second process data can differ. For example, the first process data can be based on data from a photodiode, and the second process data on data from a camera.

[0030] The spatially resolved first and second process data can contain information about the sequence in which volume elements or spatial coordinates are processed during manufacturing. This allows for time-dependent statements, such as whether the production of one volume element influences the production of another, or whether a defect in one material layer can be repaired or corrected by a subsequent layer. The (first or second) process data can therefore consist of tuples containing multiple data fields with timestamps and measured values. Furthermore, the data can be arranged in a spatial grid, with each grid point containing a data record with at least one measured value.

[0031] In this context, it is conceivable to store further global or locally resolved parameters in the process data as a direct data stream or in the form of metadata, such as real and / or nominal values ​​for laser power, laser spot size, layer thickness, powder bed temperature, powder properties, laser spot velocity and trajectory, or others. These parameters, in conjunction with the other process data, allow for the evaluation of directly dependent measured quantities. This could, for example, include brightness values ​​detected by photodiodes, which may depend directly on the laser power and / or the size of the laser spot. Furthermore, the second set of process data can be analyzed for deformations in the form of deviations in the surface shape of at least one volume element from the target geometry of the object, thus identifying defects.This can include specific measurement parameters that go beyond mere brightness information from a photodiode. The surface in this context is, in particular, the free surface that points in the processing direction and is covered by a subsequent layer, unless it is an outer surface.

[0032] In laser melting processes or other methods, the material undergoes at least one phase or state transition during the production of the corresponding volume element. For example, in laser melting, a metallic powder can be melted and then revert to a solid state. An underlying defect or incompletely melted material could result in an indentation or bulge that affects the surface of the volume element. This defect, in the form of a deformed surface, could be detected by image acquisition, possibly using fringe projection with a pattern projector, and incorporated into the secondary process data.

[0033] The second set of process data can be further examined for open or closed cavities containing raw material or fluid as defects. This could indicate incomplete material curing or faulty overhang production.

[0034] The examination of the n objects is performed using computed tomography (CT). This technique is capable of capturing even complex internal geometries and can provide measurement data that are highly informative regarding geometric deviations and material defects. For the n objects, initial process data can therefore be determined first. Once the objects are completed, computed tomography scans can be generated and analyzed to determine the measurement data. The spatially resolved initial process data is structurally similar to the data from a computed tomography scan, allowing for good correlation between the two. The CT scan data can then be subjected to conventional defect analysis, from which a defect mask can be derived for each of the n objects.This should be understood as a representation that contains information about the areas where defects are present in the component.

[0035] Synthetic CT data can be generated from the process data. This involves using a prediction of the local X-ray absorption or material density. This prediction can then be analyzed for defects using conventional algorithms or algorithms typically used for CT data analysis. In this case, the training data consists of the grayscale values ​​of the CT data itself.

[0036] In this context, it's worth mentioning that algorithms for analyzing CT data could also be machine learning algorithms. An example of this would be an algorithm that segments the object's volume into material and background, or defect and defect-free areas. Simulated measurements can be used to train an algorithm that is based, for example, on field comparison.

[0037] The examination of the measurement data includes the substep of segmenting defects in the acquired measurement data to identify volume elements exhibiting defects. This provides spatially resolved information about which volume elements describe the geometry of defects. Individual volume elements exhibiting defects can be identified. The remaining volume elements can be marked as defect-free or considered defect-free simply because they are not marked. It can be advantageous to remove artifacts before segmentation, as these commonly occur in CT data and would complicate the analysis. The chosen size of the individual volume elements in the CT data can depend on the manufacturing process of the object. The segmentation could be generated using established analysis methods proven effective for CT data.The machine learning algorithm can be trained by correlating the segmented defect data with the initial process data. This information can then be directly derived from subsequently acquired second-generation process data using the machine learning algorithm.

[0038] Furthermore, the segmentation of defects in the acquired measurement data can be performed into at least two defect classes. In the correlation step, these at least two defect classes are linked to the corresponding correlated coordinates of the process data coordinate system. The machine learning algorithm is thus further trained to recognize different defect classes in the first and second process data sets. The at least two defect classes can include the previously mentioned different defect types, such as pores, cavities, microstructure loosening, porosities, cracks, delaminations, inclusions, or deviations in surface shape. The inclusions could be further differentiated into air-filled cavities, cavities filled with powder or other raw material, or cavities filled with foreign material. This can facilitate a user's decision as to whether the manufactured object is still usable.

[0039] Furthermore, it can be advantageous to assign probability information about the presence of a defect to individual volume elements during segmentation. Instead of a purely binary segmentation into defective or defect-free volume elements, this allows the machine learning algorithm to estimate the probability of a given volume element exhibiting a specific defect. From this, a kind of probability map or matrix can be derived. It goes without saying that this can also include probability information for the individual defect classes.

[0040] The measurement data could be processed analogously or manually. This allows additional information to be incorporated into the training of the machine learning algorithm, information not apparent from the visual data alone. For example, individual defects spanning multiple volume elements can be marked as critical or non-critical. The training process then results in the algorithm not evaluating each individual volume element separately, but rather developing a prediction for larger areas of the process data. Tolerance limits for critical defects could also be trained in this way.

[0041] Furthermore, prior to correlation, the process data coordinate system and the measurement data coordinate system for each of the n objects can be registered using elastic, multi-modal registration. This transforms the measurement data and the initial process data, which exist in two independent coordinate systems due to deformations, into the same structure or a single coordinate system. Only then is a direct comparison between the detected defects and the corresponding process data possible. In a preferred embodiment, the measurement data and the initial process data are therefore registered with each other before the initial process data and the defect coordinates are correlated. Methods for this registration are known in the art, particularly in the field of medical technology.

[0042] In general, a model describing the image from the imaging procedure used to determine the initial defect data can be established for registration. For example, control points or support points can be defined, and the deformation or image between these points can be interpolated or extrapolated. The control points can be arranged in a regular or irregular grid. The irregular grid can have a higher resolution in relevant areas, such as near the surface, around defects, in areas with large deformation, or in areas where the image does not closely approximate the target geometry. This reduces the total number of control points and thus the required computation time. According to another example for describing the image, section-by-section normal rigid adjustments can be determined and, if necessary, interpolated between them.Furthermore, to describe the mapping, a possibly location-dependent mapping can be described globally and thus analytically for the entire three-dimensional space under consideration, e.g. using a Fourier series.

[0043] Furthermore, there are various ways to determine the necessary mapping in the case under consideration and thus to define the parameters of the aforementioned model. An error measure and an optimizer can be used for this purpose.

[0044] The error measure describes how well the target geometry and the measured data match. Reference points of identical structures are needed in both datasets. These can be, for example, the surface or surface structures such as corners and edges, but also internal defects that manifest as variations in gray values.

[0045] Other corresponding features can also be used, such as easily recognizable geometries, so-called landmarks, geometries or geometric areas defined by an evaluation rule, or manually defined geometries or landmarks. Distinctive geometries, such as surfaces or surface structures like corners or edges, can be selected as landmarks. It would also be conceivable to use clearly recognizable defects or desired cavities or edges inside the object to be manufactured.

[0046] In principle, this error measure can be defined based on geometries and thus interfaces between material and background, or directly on the image data or grayscale values ​​of the datasets. Both methods can also be combined.

[0047] The error measure can, for example, be a cross-correlation function between the gray values ​​of the process data and the defect data images or measurement data. Alternatively, a distance function based on the envelope of the component can be used as an error measure.

[0048] Transfer functions can be used to transform the initial defect data into a different coordinate system or structure. These transfer functions can be adjusted by the optimizer, which is based on an evaluation function. When calculating the evaluation function, an error measure can be determined for each computation step, which the optimizer then aims to minimize. Adjusting the function may require more computational power and / or time as the number of support points increases. For very simple defects and small deformations, a rigid transformation may suffice. Global scaling can provide an additional degree of freedom for the transformation.

[0049] The optimizer can, for example, be started with a reduced number of landmarks and, after an initial approximation, gradually increase the number of landmarks in an iterative process. For instance, defects could be identified that are present in both the measurement data and the initial process data. Only these defects would then be considered in a subsequent, elastic, multi-modal registration. This can accelerate or improve convergence overall. Regularization can prevent overfitting, which occasionally occurs in machine learning algorithms, thus ensuring convergence even if certain landmarks are only found in one of the relevant datasets.

[0050] In another example, the expected deformation of the object during additive manufacturing can be considered before the elastic, multimodal registration of the initial process data to the measurement data. These pre-deformed initial process data can be more easily and effectively adapted to the measurement data because they are already closer to the desired geometry. The expected deformation can be determined, for example, using a simulation of the additive manufacturing process.

[0051] In another example, the elastic, multimodal registration can be determined solely based on the surface data of the initial process data and the measurement data, and accordingly transferred to the inner areas.

[0052] This also allows for a mechanical simulation of the deformation of the internal areas, taking into account the known surface deformation. The defect coordinates determined in the measurement data can then be correlated with the initial process data using elastic, multi-modal registration, without requiring the detection of defects in the initial process data. This allows training data to be collected even without detecting defects in the process data. Therefore, the acquisition of training data is independent of defect detection in the initial process data.

[0053] In another example, it is conceivable that only data on individual defects are recorded. Consequently, from the overall view of all volume elements of an object, those areas considered to have a defect can be filtered out. This can simplify the collection of training data. It should be taken into account that corresponding individual features, which are to be interpreted as belonging to a defect, may also occur in only a single data set. As mentioned previously, a defect could, for example, be corrected by processing subsequent layers, so that it does not appear at all in the imaging study.

[0054] In another example, it is conceivable to perform the analysis of the second process data for defects in such a way that synthetic reference data (for example, volume data from a computed tomography measurement) are generated on the basis of the second process data and then analyzed for defects using conventional algorithms.

[0055] In the preceding description, certain features and properties that refer to process data without the term "first" or "second" apply to both first and second process data without explicitly stating so in each case.

[0056] The method according to the invention can be extended by additional functions that can provide information about the effects of detected defects from secondary process data. For example, an exact component geometry, taking the defects into account, could potentially be derived from an analysis of the secondary process data. This opens up the possibility of checking geometric dimensions and tolerances such as size, shape, and position, calculating strength, and simulating the mechanical, thermal, or electrical properties of the object. This would allow for an assessment of whether the object fulfills the required properties for its intended use.

[0057] The procedure can further include the following steps: capturing defect positions in the initial process data, identifying defect pairs from defect coordinates in the measurement data and captured defect positions in the initial process data that are associated with each other with a probability above a predefined probability threshold, and registering the process data coordinate system and the measurement data coordinate system based on the identified defect pairs. The defect positions can be coordinates of the process data coordinate system that define the positions of the defects in the process data.

[0058] This ensures that only defects or defect pairs are used that have been found with a certain degree of confidence in process and measurement data. This provides effective registration based on the defect pairs between the initial process data and the measurement data. This improves the registration and results in a better correlation between the defect coordinates and the initial process data, even for the areas between the defect pairs. Consequently, improved training data is generated, leading to better training of the machine learning algorithm.

[0059] Furthermore, the procedure may include the step: Repeating the steps Identify and Register, wherein each repetition of the Identify step is performed based on the immediately preceding registration, wherein the repetition of the Identify and Register steps is performed until a predefined termination condition is met, wherein after a termination the last performed registration is used for correlation.

[0060] This means that the improved registration from the previous iteration identifies more defect pairs in the current iteration, as the improved registration correlates the defects in the initial process data and the measurement data more effectively. This improved detection of defect pairs leads to the identification of more pairs, which can then be used to calculate a further improved registration. An improved registration results in a better correlation of the defect coordinates with the initial process data, allowing the machine learning algorithm to be trained with even more refined training data.

[0061] A termination condition can be that if the repeated sub-step "Identify" does not identify any additional pairs of defects in the initial process data and the measurement data that are associated with each other with a probability above a predefined probability threshold, compared to the identification in the previous iteration.

[0062] The procedure can further include the following steps: Repeating the steps: capturing defect positions in the initial process data, wherein the capturing of defect positions in the initial process data is carried out using the learning algorithm from the first repetition onwards; repeating the steps identifying and registering, wherein each repetition of the identifying step is based on the immediately preceding registration; wherein the repetition of the identifying and registering steps continues until a predefined termination condition is met, and after termination, the last registration is used for correlation; correlating at least one sub-area of ​​the measurement data coordinate system comprising the determined defect coordinates of the object representation with a corresponding sub-area of ​​the process data coordinate system for collecting training data.Training the machine learning algorithm to determine defect coordinates in spatially resolved process data acquired during an additive manufacturing process for the production of an object, using the training data, whereby the steps of acquiring, repeating, correlating, and training are performed until a further predefined termination condition is met.

[0063] Improved registration, with better training data available, allows the machine learning algorithm to be trained more effectively at detecting defects in the process data. This can be used to identify more defect pairs in the initial process data and the measurement data. Therefore, further improvements to the registration can be leveraged to further enhance the training of the machine learning algorithm.

[0064] The further predefined termination condition can occur if the learning algorithm does not detect any new defect positions in the initial process data compared to the previous run.

[0065] The invention further relates to a computer program product with instructions executable on a computer, which, when executed on a computer, cause the computer to carry out the method according to the preceding description.

[0066] The advantages and further developments of the computer program product correspond to the advantages and further developments of the procedure described above. Therefore, reference is made to the description above in this regard.

[0067] Further features, details and advantages of the invention will become apparent from the wording of the claims and from the following description of exemplary embodiments with reference to the drawings. The drawings show: Fig. 1 a flowchart of a method according to the invention; and Fig. 2 a schematic representation of the correlation of defect coordinates with the process data coordinate system.

[0068] Fig. 1 Figure 100 shows a flowchart of a computer-implemented method for determining defects in an object manufactured using an additive manufacturing process. The method is essentially divided into two successive steps. A first, preparatory step A involves determining spatially resolved initial process data from n objects, which are acquired during an additive manufacturing process to produce the n objects. This initial process data step can therefore also be part of an additive manufacturing process for an object. The initial process data defines a process data coordinate system for each of the n objects.

[0069] For each volume element or spatial coordinate processed during manufacturing, one or two (or more) physical measurements can be recorded for the spatially resolved process data.

[0070] The spatially resolved first and second process data can contain information about the sequence in which volume elements or spatial coordinates are processed during manufacturing.

[0071] Subsequently, after their manufacture, the n objects are recorded in a further step using imaging non-destructive methods. Measurement data of the n objects are determined. The measurement data defines a measurement data coordinate system for each of the n objects. The determination of the measurement data will be carried out using CT measurements.

[0072] An investigation of at least a subset of the measurement data follows. Defect coordinates are determined 106, which are assigned to defects in the object representation. The resulting measurement data from a CT scan are segmented into volume elements affected by a defect 122, whereby the defects are also segmented into several defect classes.

[0073] In a further optional step, defect positions are recorded in the initial process data 126. The defect positions can specify coordinates in the process data coordinate system that are assigned to defects of the object in the process data.

[0074] The recorded defect positions from the initial process data and the determined defect coordinates from the measurement data can optionally be compared. In a further optional step, defect pairs are identified consisting of defect coordinates in the measurement data and recorded defect positions in the initial process data that are assigned to each other with a probability above a predefined probability threshold. This means that the defect pairs have defect positions in the initial process data that are highly likely to be found at a defect coordinate in the measurement data, and vice versa.

[0075] Based on the identified defect pairs, the process data coordinate system and the measurement data coordinate system can be registered. This means that a mapping is determined that maps the coordinates of the process data coordinate system and the measurement data coordinate system onto each other. This registration can be an elastic, multimodal registration.

[0076] Fig. 2 To explain registration 124, a visualization 10 of process data of an object and a visualization 12 of measurement data of an object that was manufactured using an additive manufacturing process are shown.

[0077] The object has a defect, which is identified by reference number 14 in visualization 10 and by reference number 16 in visualization 12. Visualizations 10 and 12 use different coordinate systems: visualization 10 uses a process data coordinate system, and visualization 12 uses a measurement data coordinate system.

[0078] Arrows 18 represent a registration using a mapping that maps various coordinates of the process data coordinate system to the corresponding coordinates of the measurement data coordinate system. An inverse mapping can also be provided, which maps the coordinates of the measurement data coordinate system to the process data coordinate system.

[0079] Defect 14 and defect 16, which represent the same defect in the object in the process data and the measurement data, can be identified as defect pair 20 128.

[0080] The steps Identify 128 and Register 124 can be performed according to Figure 1 In an optional embodiment, the process is repeated 130 times. The repetition 130 results in an iterative improvement of the registration 124, since the process data coordinate system and the measurement data coordinate system are aligned using the registration 124. In this process, defects in the initial process data and the measurement data, which have not yet been identified as defect pairs, can be linked by mapping the registration, so that they are recognized as new defect pairs during the subsequent repeated identification 128. This increases the number of identified defect pairs in each repetition 130. The increased number of defect pairs leads to an improvement of the registration 124.

[0081] Repetition 130 can be performed until a predefined termination condition is met. This predefined termination condition is met if the repeated substep "Identify 128" does not identify any additional defect pairs in the initial process data and measurement data compared to the "Identify 128" step in the previous iteration. In other words, if repeating the steps does not result in any further improvement in the registration.

[0082] The defect coordinates determined from the measurement data are correlated in a further step with coordinates of the process data coordinate system.108 This means that the defects from the measurement data are compared with the defects from the initial process data and related to each other. Training data is thus collected using the correlated initial process data. Furthermore, defect-free areas of the measurement data can also be correlated with the corresponding defect-free areas of the initial process data. The correlation of the defect-free areas can also contribute to the training data.

[0083] If 122 defect classes were considered during segmentation, these defect classes can be linked to the corresponding correlated coordinates of the process data coordinate system during correlation. The defect classes then also become part of the training data and can later be used when evaluating the second set of process data with the trained machine learning algorithm.

[0084] In a further step, the machine learning algorithm is trained. Using the training data, the machine learning algorithm is trained to recognize the defects in the initial process data and to determine the coordinates of the defects in the initial process data.

[0085] An additional optional step can be performed, in which the steps Capture 126, Repeat 130, Correlate 108, and Train 110 are repeated 132. First, the defect positions in the initial process data are captured 126, with the capture 126 of defect positions in the initial process data being performed using the trained machine learning algorithm. The trained machine learning algorithm detects more defect positions in the initial process data with a high probability and with a tendency toward higher accuracy than were detected during the initial capture 126 or any previous capture 126, because the registration 124 has been improved by the repetitions 130. This very likely results in an increase in the number of identified defect pairs.This further improves the registration 124 through a further repetition 130, allowing for the collection of even better training data, which the machine learning algorithm uses for improved training. With this improved training, the machine learning algorithm can then, in a further repetition 132, detect even more defect positions in the initial process data 126.

[0086] Repetition 132 can be performed until another predefined termination condition is met. This further predefined termination condition might be met, for example, if the machine learning algorithm does not detect any new defect locations in the initial process data compared to the previous iteration. It can then be assumed that either no or very little optimization potential can be realized through further repetitions 132.

[0087] This completes preparatory process A. Following preparatory process A, process B can be performed, in which potential defects can be detected without an imaging analysis method using the trained machine learning algorithm. For this purpose, spatially resolved secondary process data are acquired 112, which are recorded during an additive manufacturing process for the production of an object. Subsequently or in parallel, the secondary process data are analyzed for possible defects by examining the secondary process data section by section using the trained machine learning algorithm 114. It is understood that process B can be repeated for any further objects produced using the additive manufacturing process.

[0088] The analysis 114 can, for example, involve examining 118 the second process data for deformations in the form of deviations in surface shape, in the form of a target geometry of the object, or at least of a volume element, as defects. Furthermore, the examination 120 of the second process data can also be performed to detect open or closed cavities containing raw material or fluid as defects.

Claims

1. Computer-implemented method for the determination of defects of an object produced by means of an additive manufacturing process, wherein volume elements or spatial coordinates of the object to be produced are processed during the additive manufacturing process, wherein the method (100) has the following steps: - determining (102) locally resolved first process data of n objects which are recorded during an additive manufacturing process for producing the n objects, the first process data defining a process data coordinate system for each of the n objects, - determining (104) measurement data of the n objects by non-destructive imaging methods after the production of the n objects, the measurement data defining an object representation in a measurement data coordinate system for each of the n objects, wherein the determination (104) of measurement data of the n objects is carried out by a computed tomography (CT) measurement, - determining (106) the coordinates of at least one subregion of the measurement data coordinate system which are defect coordinates, which are assigned to a defect in the object representation; wherein the examination (106) of the measurement data has the substep: segmentation (122) of defects in the measurement data determined in order to identify volume elements which have a defect, - correlating (108) the at least one subregion of the measurement data coordinate system comprising the defect coordinates determined with a corresponding subregion of the process data coordinate system in order to collect training data, - training (110) an adaptive algorithm for the determination of defect coordinates in locally resolved process data, which have been recorded during an additive manufacturing process for producing an object, by means of the training data, - determining (112) locally resolved second process data which are recorded during an additive manufacturing process for producing an object, and - analysing (114) the second process data in respect of defects by means of the adaptive algorithm.

2. Method according to Claim 1, characterized in that the step of analysing the second process data in respect of defects by means of the adaptive algorithm is carried out during the determination (112) of the second process data.

3. Method according to one of the preceding claims, characterized in that, during the production of the object, at least two different physical measurement quantities for the locally resolved first and second process data are respectively recorded (116) for a volume element being processed during the production or for a spatial coordinate.

4. Method according to one of the preceding claims, characterized in that the locally resolved first and second process data have information relating to the order in which volume elements or spatial coordinates are processed during the production.

5. Method according to one of the preceding claims, characterized in that the second process data are examined (118) in respect of deformations in the form of deviations of the surface configuration of at least one volume element from a setpoint geometry of the object as a defect.

6. Method according to one of the preceding claims, characterized in that the second process data are examined (120) in respect of open or closed cavities with inclusion of raw material or fluid as a defect.

7. Method according to one of the preceding claims, characterized in that the segmentation (122) of defects in the measurement data determined is carried out into at least two defect classes, the at least two defect classes being linked with the corresponding correlated coordinates of the process data coordinate system in the correlation step (108).

8. Method according to Claim 7, characterized in that the at least two defect classes have the defect classes of pores and / or inclusions of extraneous material.

9. Method according to one of the preceding claims, characterized in that, before the correlation (108), the process data coordinate system and the measurement data coordinate system are registered (124) for each of the n objects by means of an elastic multimodal registration.

10. Method according to Claim 8 or 9, characterized in that the method furthermore has the following steps: - recording (126) defect positions in the first process data, - identifying (128) defect pairs of defect coordinates in the measurement data and recorded defect positions in the first process data, which are assigned to one another with a probability above a predefined probability threshold value, - registering (124) the process data coordinate system and the measurement data coordinate system on the basis of the defect pairs identified.

11. Method according to Claim 10, characterized by repetition (130) of the steps of identification (128) and registration (124), each repetition (130) of the identification step (128) being carried out on the basis of the registration carried out immediately before, the repetition (130) of the identification and registration steps being carried out until a predefined termination condition is satisfied, the registration (124) carried out last being used for the correlation (108) after a termination.

12. Method according to Claim 11, characterized by repetition (132) of the steps: - recording (126) defect positions in the first process data, the recording of defect positions in the first process data being carried out at the latest starting with the first repetition by means of the adaptive algorithm, - repeating (130) the steps according to Claim 11, - correlating (108) the at least one subregion of the measurement data coordinate system comprising the determined defect coordinates of the object representation with a corresponding subregion of the process data coordinate system in order to collect training data, - training (110) the adaptive algorithm in order to determine defect coordinates in locally resolved process data, which are recorded during an additive manufacturing process for producing an object, by means of the training data, the repetition of the steps of recording (126), repetition (130), correlation (108) and training (110) being carried out until a further predefined termination condition is satisfied.

13. Computer program product having instructions that can be executed on a computer and when executed on a computer cause the computer to carry out the method according to one of the preceding claims.