Method and device for detecting defects during an additive manufacturing process

Abstract

The invention relates to a method for detecting anomalies during a powder bed-based additive manufacturing process, in particular for optically detecting defects of a powder bed during an SLM manufacturing process of a component, said method having the steps of: detecting the state of a powder bed layer by capturing an image of the powder bed; creating an anomaly-free image of the powder layer by inputting the image into an autoencoder and correcting the image by means of the autoencoder in order to obtain the anomaly-free image as an output image or inputting layer data into the autoencoder in order to simulate the state of the powder bed and creating the anomaly-free output image on the basis of the simulated state; and comparing the output image with the image of the powder bed in order to detect anomalies of the powder layer.

Description

[0001] Method and device for defect detection in additive manufacturing

[0002] The present invention relates to a method and a device for defect detection in additive manufacturing, in particular in a manufacturing system based on optical interaction, such as an SLM (Selective Laser Melting) manufacturing system. It proposes monitoring and detection of the powder bed in the workspace of the manufacturing system in order to efficiently and accurately detect defects in the powder layers.

[0003] Background of the invention

[0004] Additive manufacturing, and in particular the SLM manufacturing process, can be a very time-consuming manufacturing process, making it very costly to have a machine operator continuously monitoring a build job (or the manufacturing order for a component). In particular, manually monitoring a build job for each individual shift is difficult and very time-consuming. Satisfactory approaches for automating the monitoring of the manufacturing process are currently not sufficiently mature to gain market acceptance.

[0005] Approaches to monitor the manufacturing process by analyzing simple difference images of individual layers are currently insufficient in terms of accuracy and are too complex and difficult to automate. Furthermore, changes in individual layers can be incorrectly detected as defects, depending on the component geometry being printed. Therefore, a deviation between two layers can only be used as a defect model to a limited extent. Manual rework is usually necessary to detect defects during the manufacturing process. Errors in additive manufacturing typically result from different process parameters, different lighting conditions, and material changes.However, since certain changes to these parameters may also be desired, for example for the production of certain components, this must also be taken into account for the detection of errors in the manufacturing process.

[0006] Approaches to enable defect detection using supervised machine learning are currently still costly and labor-intensive. These models are usually trained based on images that already contain real defects, which poses the additional challenge of only a limited number of these images being available, and a large number of them would be required for a corresponding training process. Consequently, supervised machine learning approaches only lead to specific solutions, are usually difficult to generalize, and lack the required accuracy and reliability.

[0007] An object of the present invention is therefore to provide an optimized method and an optimized device to enable efficient and highly precise defect detection in additive manufacturing. In particular, it is an object to efficiently detect defects in the powder bed layer during layer-by-layer manufacturing processes, in particular SLM processes, thus improving the manufacturing process and, for example, also enabling an assessment of the quality of a finished component.

[0008] Detailed description of the invention

[0009] To achieve the above-mentioned objects, the features of the independent claims are proposed. The dependent claims relate to preferred embodiments of the present invention. A manufacturing system, for example an SLM system, for producing a workpiece by exposing powder material and / or a workpiece element can comprise at least one light source for irradiating the powder material and / or workpiece element provided in a process chamber. In addition, a light path generated by the light source and running in the process chamber of the manufacturing system can be present. By applying the powder material layer by layer and subsequently melting the material at predetermined locations by irradiation by the light source, a component to be manufactured can be produced step by step. After each exposure process, a new powder layer can be applied to the powder bed.To detect anomalies in this powder-bed-based additive manufacturing process, the following is proposed: Recording the condition of a powder bed layer by taking one or more images of the powder bed. For example, a camera module with one or more cameras can be used to create an optical image of the current powder bed layer as it exists in the workspace of the production machine. Such an image of the powder bed can, for example, show applied powder material as well as molten powder material due to the exposure process, as well as any defects. For example, individual images of the powder bed layer can be created at high frequency and combined to generate an evaluation image of each built layer.

[0010] The image (first image) is therefore preferably created directly after an exposure process of the powder bed. Alternatively or additionally, an image can also be created directly before the exposure process (after a new powder layer has been completely applied). In a particularly advantageous development, a large number of images are created during the exposure process, which are combined together to form an evaluation image (i.e., first image). In the following step, an anomaly-free image of the powder layer can be created based on the optical representation of the current powder bed. Thus, a step of creating an anomaly-free image can be carried out in the powder step. This step can preferably comprise the "input of the first image" to an autoencoder and the step of "correcting the input image" by the autoencoder so that an anomaly-free image can be generated as the output image.

[0011] The input image of the current powder bed layer, which is an optical representation of the current state of the powder bed, can be modified via filtering by the autoencoder to eliminate defects in the powder bed. The autoencoder can therefore be trained to filter out defects in the powder bed from a real optical image of a powder bed in order to ultimately obtain an anomaly-free image as the output image. Alternatively or additionally, layer data can be input to the autoencoder to create an anomaly-free image of the powder bed layer. The layer data can, in particular, include the information required for the production of the layer by the additive manufacturing system, such as the component geometry of the component to be manufactured in the respective layer.After inputting the layer data, the autoencoder can perform a "simulation" of the powder bed state to generate an image of the state after exposure. Optionally, the autoencoder can also receive layer data and / or an image of the previous, already applied layer. The autoencoder can be configured to create an anomaly-free output image based on the input data. Thus, the autoencoder creates a simulated state (or filtered state) based on the input data, which can be output as an anomaly-free output image (optical representation of a applied or already exposed build layer).

[0012] In the subsequent step, anomalies in the powder layer can be detected based on the output image. For this purpose, advantageously, for example, a comparison of the detected state of the powder bed layer, and in particular the detected image of the powder bed (first image), can be performed with the output image output by the autoencoder, which is an anomaly-free image. Based on this comparison, it is possible to directly and effectively detect anomalies and, in particular, defects in the powder bed layer and / or the components. Based on the output, for example, the manufacturing process (e.g., during the execution of the build job) can be intervened in, in particular, to make readjustments or to detect component defects at an early stage in order to adapt the manufacturing process accordingly.Additionally or alternatively, it is also possible to create an analysis of the finished component based on the data with regard to possible manufacturing defects that may have occurred.

[0013] The current image of the powder bed, which is recorded in a first step to capture the state of the powder bed layer, can in particular correspond to a current state of at least part of the powder bed. However, it preferably corresponds to the state of the entire powder bed present in the work area. Anomalies are, in particular, defects in the powder layer in the current powder bed, wherein the image of the powder bed is preferably a real image recorded using an optical camera. Alternatively or additionally, the image can also be processed using filters, for example to simplify comparison with the output image of the autoencoder. This enables precise defect detection in additive manufacturing in an efficient manner.

[0014] The autoencoder can advantageously be trained based on a large number of defect-free images of the powder bed. These images are advantageously images before the exposure process. Alternatively or additionally, images after the exposure process (of the current layer or the previous layer) can also be used. Preferably, the output image and the image of the powder bed are the same layer of the layer structure of the component to be manufactured, so that representations of the same layer are compared. Training the autoencoder based on a large number of defect-free images of the powder bed can be easily enabled because defect-free images of a powder bed are available in large quantities and, particularly advantageously, manual detection and identification of specific defects for the learning process are not necessary in one embodiment, or are only necessary to a very limited extent.

[0015] The autoencoder can advantageously be an artificial neural network used to learn a compressed representation for an input dataset in order to extract essential features. The autoencoder can advantageously be trained using artificially created anomalies (artificially created defects) that are integrated into the defect-free images.

[0016] The method may include the step of determining a value indicative of the similarity between the input data generated to capture the state of the current powder bed layer and the anomaly-free image or reference data provided by the trained autoencoder. Thus, an output data set can be generated if the value indicative of a similarity is below a predetermined threshold. If the output data set is thus generated, a corresponding difference in the compared data exists, thus indicating an anomaly.

[0017] Method according to one of the preceding aspects, wherein the autoencoder is trained by unsupervised learning. Thus, a method for monitoring production can be provided in a particularly efficient manner. Training the autoencoder based on a large number of defect-free images of the powder bed can be easily enabled because defect-free images of a powder bed are available in large quantities and, particularly advantageously, manual detection and identification of specific defects are not necessary for the learning process in one embodiment.

[0018] Advantageously, the output of the method can be used to output a warning signal to inform the machine operator and / or a control signal to control the production machine. Defect-free images of the powder bed, which can preferably be assigned to a specific component layer, can be used as training data for the autoencoder. For this training data, the defect-free images can preferably be automatically loaded with artificially generated anomalies, so-called artificially generated defects (in particular artificially generated defect shapes and / or based on random numbers and / or through further processing of random numbers). The artificially generated defects can be automatically created based on a defect database. In particular, the artificially generated anomalies can be generated based on random numbers or through further processing of random numbers.

[0019] For example, a defect generator can be used that inserts artificially generated anomalies into defect-free images using defect shape data from a defect database. These artificially generated anomalies preferably correspond to defects that can occur in the powder bed and / or on the component, for example, before or after the laser is activated for exposure.

[0020] To train the autoencoder, a plurality of anomaly-free output images is advantageously provided as training data, into which one or more artificially created defects are inserted to obtain modified output images, and the autoencoder is trained to filter out the anomalies from the modified output images. At the end of the training phase, the autoencoder is thus trained such that anomalous input images are reconstructed into anomaly-free output images. The autoencoder is therefore preferably configured to detect defect-free features of a powder layer and, therefore, to reconstruct only what was trained as defect-free during reconstruction. This filters out defect features.Advantageously, to correct an image of the powder bed, which corresponds, for example, to an actual optical photograph of the powder bed, the autoencoder can compress the image using a filter and then generate the output image, which is free of anomalies, using a decoder.

[0021] In a further embodiment, the autoencoder is trained in such a way that, based on layer data, the state of the powder bed after applying another layer can be generated as a representation. The output of the autoencoder is thus an anomaly-free optical representation of the state of the powder bed after exposure.

[0022] In one embodiment, the autoencoder can be trained based on defect-free images of a powder bed. Images in this embodiment are real images (e.g., photographs or optically corrected photographs) of the powder bed before laser exposure. The powder bed is thus ready for laser exposure in this state. In particular, a new powder layer (in particular, completely) has been applied to the powder bed in this state by means of a coater.

[0023] The method for detecting anomalies in powder bed-based additive manufacturing, in particular for optically detecting defects in a powder bed in an SLM manufacturing process of a component, may comprise the steps:

[0024] Detecting the condition of a powder bed layer by taking at least one image of the powder bed before exposure;

[0025] Creating an anomaly-free image of the powder layer by:

[0026] Input of at least one image (image before exposure) to an autoencoder, variant “A”, and correcting the image by the autoencoder to obtain the anomaly-free image as the output image;

[0027] Comparison of the output image with the image of the powder bed to detect

[0028] Anomalies of the powder layer. In an additional or alternative embodiment, the autoencoder can be trained based on defect-free images of a powder bed, wherein images of this embodiment are real images (e.g., photographs or optically corrected photographs) of the powder bed after exposure by laser. In this state, the powder bed was thus at least partially or completely exposed by the laser. Advantageously, the powder layer was completely exposed in this state.

[0029] In addition, it is possible to use the layer data as additional data for training the autoencoder, which can further improve the accuracy of the autoencoder.

[0030] In an advantageous development, in addition to or as an alternative to the layer data, information (in particular photographs or optically corrected photographs) of the powder bed after exposure of the previous layer and information (in particular photographs or optically corrected photographs) of the unexposed powder bed of the current layer are used. The information before and after an exposure is thus additionally used in this embodiment. This allows the accuracy of the autoencoder to be further improved.

[0031] In a particularly preferred development, a scanning path (of the laser) of a current layer (“before” exposure) is combined with an image (in particular a photograph) of the unexposed powder bed of the current layer and additionally with an image of the already exposed powder bed of the previous layer. For example, the various images can be combined to form a combined image, whereby this combined image can be used to train the autoencoder in order to generate the state of the powder bed after exposure as an output image by the autoencoder. The autoencoder thus generates an output image of the current layer as it would appear after exposure. An autoencoder trained in this way can therefore be used to detect defects, for example by comparing an actual image of the powder bed (after exposure of the current layer) with the output image.

[0032] In a particularly advantageous embodiment, two autoencoders are provided, each trained differently: the first autoencoder with images "before exposure," and the second autoencoder with at least images "after exposure" (or additional data such as combined images). Depending on the application (before or after exposure), switching between the autoencoders is possible, resulting in very high flexibility and analysis accuracy.

[0033] Advantageously, the autoencoder can be trained in such a way that, based on layer data, a state of the powder bed after a manufacturing step can be represented and a generated image of the state of the powder bed is created, in particular an expected defect-free surface of the powder bed after a scanning process.

[0034] A method for detecting anomalies in powder bed-based additive manufacturing, in particular for optically detecting defects in a powder bed in an SLM manufacturing process of a component, may comprise the steps:

[0035] Detecting the condition of a powder bed layer by taking at least one image of the powder bed after exposure;

[0036] Creating an anomaly-free image of the powder layer by:

[0037] Input of slice data and / or the at least one image (“after exposure”) to the autoencoder, variant “B”, and creating the anomaly-free image as an output image based on the slice data and / or the at least one image;

[0038] Detecting powder layer anomalies based on the output image.

[0039] Another advantageous option is to train the autoencoder based on three different additional input data. For example, this can be done by combining the layer data (layer information) with information (especially photographs) of the powder bed after exposure of the previous layer and the unexposed powder bed of the current layer. During training, this input data can be fed to the autoencoder as input (e.g., as "prepared images"). The autoencoder attempts to generate an output image that is as close as possible to a real image of the exposed powder bed. The quality of the autoencoder is measured based on the "loss" (especially a loss function). During training, this "loss" is made as small as possible.Using the trained autoencoder, it is thus possible to generate an artificially generated image of the powder bed based on the input data. By comparing the artificially generated image with the real image of the powder bed, the position and size of anomalies, for example, can be precisely determined.

[0040] Further advantageously, the layer data can include at least geometric data of the components to be manufactured and / or production parameters of the production system. The layer data can also include, for example, layer thickness, type of powder material, and / or exposure parameters for the specific production of the respective layer.

[0041] The layer data may also include laser target position data, which may specify at least one predetermined position in the powder bed at which the laser processes, in particular melts, the powder material.

[0042] Advantageously, the method can also detect the state of a current powder bed layer during the manufacturing process. For this purpose, the powder layer in the workspace of the additive manufacturing system is illuminated from at least one direction and an image of the powder layer is captured. Thus, during the additional illumination of the powder layer, a preferably optical image of the powder layer in the powder bed in the workspace is created. Defect detection can thus be performed live during the manufacturing process, i.e., in particular, parallel to the manufacturing process of a component.

[0043] Advantageously, at least in the event that an anomaly is detected, a message can be issued and / or active intervention in the production process can take place.

[0044] In addition, a method for manufacturing a component by means of an additive manufacturing device, in particular an SLM system, can comprise the step of: comparing the output image with the image of the powder bed to detect anomalies of the powder layer after a scanning process, wherein, if no anomaly was detected, the scanning process is released for the next layer.

[0045] Advantageously, a computer-implemented method for detecting anomalies in powder bed-based additive manufacturing can be proposed, comprising at least one of the steps of the preceding aspects.

[0046] A method for producing a component by means of additive manufacturing can carry out the steps for detecting anomalies according to at least one of the preceding aspects several times, in particular after each scanning process.

[0047] Furthermore, a method for training an autoencoder is proposed. The method for training an autoencoder to detect anomalies in powder-bed-based additive manufacturing based on training data, in particular training images, can comprise the following steps: providing a set of defect-free images of a powder bed during the production of a component (these images can also be referred to as training images); incorporating artificially generated anomalies as artificially generated defects into the defect-free training images to obtain modified training images; inputting the modified training images to the autoencoder; and adjusting the weighting of the autoencoder such that the input modified training images are filtered by the autoencoder to produce defect-free training images as output images.

[0048] In a further development, the autoencoder can be trained based on training time series.The method may comprise one or more of the steps of: generating a set of artificially generated defects for the one or more autoencoders based on testing the autoencoder with data from training images of at least a subset of the total training images; generating a probabilistic model of artificially generated defects for each autoencoder based on a set of artificially generated defects that occur across all training images processed by the autoencoder, and generating an aggregate model based on aggregating the probabilistic models for all autoencoders; calling at least a subset of the autoencoders; determining artificially generated defects; comparing the artificially generated defects for the autoencoders with the aggregate model to determine a distance therebetween; and identifying an anomaly in response to a determination that the distance exceeds a threshold.

[0049] Furthermore, a device for data processing in additive manufacturing can be proposed, in particular for optically detecting defects in a powder bed in an SLM manufacturing process. Means can be provided for carrying out the steps of one of the methods according to at least one of the preceding aspects.

[0050] Also proposed is a computer program element comprising computer program code for causing the computer, when loaded into and executed on a computer system, to perform the steps of a method as described above. Also proposed is a computer system including a processor and a memory storing computer program code for performing the steps according to one of the preceding aspects.

[0051] Short description of the characters

[0052] Fig. 1: shows a first representation of an application of the autoencoder A and the comparison between generated and real images of a powder bed;

[0053] Fig. 2: shows an exemplary structure of an image acquisition system of the present invention;

[0054] Fig. 3: shows the structure of the autoencoder A;

[0055] Fig. 4: shows schematically a training sequence of the autoencoder according to an embodiment;

[0056] Fig. 5: shows a schematic diagram of the anomaly detection process in powder bed-based additive manufacturing;

[0057] Fig. 6a, b, c: show exemplary layer data of different layers as well as real and generated images (or image data) and their difference images and in 6c a combination;

[0058] Fig. 7: shows a schematic structure of the training process of the

[0059] Autoencoder according to a first embodiment;

[0060] Fig. 8: shows the training sequence for training the autoencoder A according to a second embodiment;

[0061] Fig. 9: is another illustration of the training of the autoencoder according to an embodiment of the present invention;

[0062] Fig. 10: shows the schematic flow of a method for defect detection in additive manufacturing according to the present invention;

[0063] Fig. 11: shows the schematic process of an exposure;

[0064] Fig. 12: shows examples of training images. Detailed description of preferred embodiments

[0065] In the following, exemplary embodiments of the present invention are described in detail with reference to exemplary figures. The features of the exemplary embodiments can be combined in whole or in part, and the present invention is not limited to the described exemplary embodiments.

[0066] Fig. 1 shows exemplary images that can be used in a device and a method for detecting defects in powder bed-based additive manufacturing. The starting point is the image labeled "laser target position," which is a digital representation of a specific layer during additive manufacturing. On the one hand, extensive dark regions can be seen, which are regions in which remelting of the material powder does not take place for this layer. In addition, gray or light areas and geometric shapes are visible, which represent the layer of the component to be manufactured. For example, the laser for the scanning process or for the exposure of the respective layer can be controlled in such a way that it exposes the light regions in order to remelt the powder. The representation S1 in Fig.1 is thus representative of the layer data (or part thereof) of a layer to be produced. In addition to the geometric dimensions and positioning of the individual regions of the layer to be melted, the exact positioning of the individual sections relative to an origin can also be specified, for example, using a scale in the X and Y directions.

[0067] In additive manufacturing, for example, three-dimensional objects (workpieces or components) are created by solidifying successive layers of powdered build material. The properties of the created objects can depend on the type of powder and the type of solidification mechanism. Parameters that influence this can optionally also be directly contained in the layer data, which, according to the proposed embodiment, can be used for the method for detecting anomalies in powder-bed-based additive manufacturing. The layer data can also include vector data or mask data, which are adapted, for example, to the respective manufacturing process of the layer.

[0068] Based on the layer data, an artificially generated image S2 can be created via the autoencoder A, as shown in Fig. 1. This image is representative of the result of a simulation of an exposure process or scanning process of the additive manufacturing system. The simulation thus creates an artificially generated representation that depicts the optical conditions in the powder bed after a scanning process has been carried out. The autoencoder is trained in such a way that, based on the layer data S1, an artificial representation is generated that represents the optical state in the powder bed during error-free production. As shown in Fig. 1, various melted areas and component sections can be seen, surrounded by unmelted powder. Finally, Fig. 1 shows a real optical representation of a state of the powder bed during actual additive manufacturing using the real image R.This real image can contain not only the powder bed and the produced component layers, but also defects that were created by the manufacturing process. As can be seen from this illustration, it can be very difficult in practice to identify defects, and in particular component defects, from such a visual representation of the condition of the powder bed. On the one hand, the color of the unmelted powder sometimes differs only slightly from component sections that are melted, as these are both usually represented in a certain shade of gray. Distinguishing between these different shades of gray is a particular challenge. In addition, there can be a wide variety of types of defects. These so-called printing defects can be line defects or point defects, for example, which can result from incorrect application of the powder material or faulty melting.This makes it difficult in many cases to distinguish between a desired component and a defect. A further challenge is the small difference between individual pixel values, which sometimes makes deviations or defects difficult to detect visually. Human intervention is often necessary to monitor the build process, which can be very time-consuming and inefficient.

[0069] Particularly in an SLM process, error monitoring can be very time-consuming and inefficient, and continuous monitoring by a machine operator can also be difficult to implement. Due to the boundary conditions described, automation is also not easily possible. The optical detection of the condition of the powder bed is shown as an example in Fig. 2 using the camera system KS, which detects the powder bed (preferably optically) in at least one direction. The camera system can, for example, have additional lighting devices to generate appropriate brightness in the area of ​​the powder bed for taking photographs. The camera system can advantageously comprise two cameras, each with a corresponding lighting device, in order to produce the most precise images of the powder bed. A camera can, for example, be a camera detector or camera sensor, or a camera can contain the detector.The detector can be designed as a spatially resolving detector, so that the detector has a plurality of pixels, in particular pixel arrays. For example, the detector can contain several thousand pixels, in particular an array with a number of pixels in the range of 100,000 pixels to 20,000,000 pixels. By using such a spatially resolving detector, the area to be detected in the powder bed can be imaged quickly, with high resolution and high quality. The camera can be controlled via a control unit to dynamically change the image area. Fig. 3 shows the exemplary structure of an autoencoder A, which initially comprises an encoder and a decoder as well as a "bottleneck". The input of a data set, for example in Fig.3 an image with a black background and a white number 4, can be filtered via the encoder so that a compressed representation of this output image is present at the bottleneck, whereby in turn an output image is generated via the decoder which comes very close to the optical representation of the input image. The autoencoder is thus a neural network type algorithm which is trained in this case in such a way that, for example, the input is in turn reproduced as output. In the encoder or encoder section, the essential information or characteristic features of the input information are filtered out until a compressed representation is achieved. This compressed representation is in turn developed into an output image via the decoder section or decoder.By appropriately training the autoencoder, defects can be filtered out (by the encoder part) so that only defect-free images or information are stored in the compressed representation. Applying the decoder part to the compressed representation, in turn, ensures that the resulting image is defect-free. By comparing an input image and the resulting image, it is possible to locate specific defects.

[0070] Fig. 4 shows, by way of example, a first application of the present invention with data in the state before exposure ("before exposure"). Starting from an anomaly-free (or defect-free) data set, which, for example, represents anomaly-free representations of a layer of a powder bed ("before exposure"), a modification can be carried out in order to insert, in particular, artificial errors or defects, so-called artificially created defects, into these images. The autoencoder according to variant "A" is subsequently used to modify the input modified data set such that "repaired images" or a repaired data set results as the output of the autoencoder. The autoencoder thus filters out the (artificial) errors from the input data set. Since the original anomaly-free data set ("before exposure") is also known, efficient and precise training of the autoencoder can be achieved using the proposed procedure.Training is therefore essentially based on a dataset (e.g., with pre-exposure data) that is free of actual defects. The required image data is usually available in large quantities. By inserting artificial defects, precise training of the autoencoder A can be achieved efficiently and on a large scale. The application of an autoencoder trained in this way (in this variant; also referred to as autoencoder variant "A") is based on the principle that a repaired powder bed image ("before exposure") is generated, for example, by filtering out (real) defects. Anomalies can be detected by comparing the real image (input image; "before exposure") of the powder bed with the repaired image ("before exposure").

[0071] Fig. 4 shows a first possible method for defect detection in additive manufacturing, in particular an SLM manufacturing process for a component. The autoencoder A is trained to filter out defects from input data or an input data set in order to output repaired representations, i.e., essentially defect-free representations. To train the autoencoder, a large number of defect-free real representations ("before exposure") are collected, which are then in turn imbued with artificial defects. The resulting modified data set is fed into the autoencoder as an input data set or modified input data set during the training phase. The autoencoder A is then trained to filter out the defects.In the encoder and decoder sections, autoencoder A generates a representation of the original defect-free data set or the original defect-free input image ("before exposure"). These so-called repaired images thus essentially correspond to the images of the original anomaly-free data set. The autoencoder trained in this way is therefore preferably applied to data sets that represent the powder bed "before" exposure. Such an autoencoder can also be referred to as an autoencoder of variant "A." The autoencoder can thus be trained based on defect-free images of a powder bed. In this embodiment, images are real images (e.g., photographs or corrected photographs) of the powder bed before ("before exposure") laser exposure. The powder bed is thus ready for laser exposure in this state.In particular, in this state of the powder bed, a (new) powder layer has already been applied by means of a coater, but this has not yet been exposed to light.

[0072] When using the autoencoder according to variant "A", detection of anomalies in the powder layer ("before exposure") can be achieved. For this purpose, for example, a comparison of the detected state of the powder bed layer and in particular the captured image of the powder bed with the repaired image output by the autoencoder can be advantageously performed. Based on this comparison, it is possible to directly and effectively detect anomalies and in particular defects in the powder bed layer and / or the components before exposure.

[0073] The application of the thus trained autoencoder A according to variant “A” can be captured, for example, in Fig. 5. Based on the (real) input image Bl as input, the autoencoder A modifies this input image Bl to the repaired output image B2. This repaired output image B2 is an essentially defect-free image of the input image Bl and thus a modified input image Bl. The repaired representation B2 is in turn used for comparison with the input image Bl, as shown in Fig. 5. By comparing the input image Bl with the repaired image B2, it is possible to record the difference between these representations and, for example, to output a difference image DB which corresponds to a representation of the existing defects or errors in the powder bed. The process shown can, for example, be carried out during a build job after each scanning process, i.e., for each layer of the component.Defect detection can therefore be integrated into a manufacturing process for producing a component using additive manufacturing methods, such that any anomalies, and in particular defects in the powder bed, can be detected as quickly as possible. This advantageously eliminates the need for manual defect detection by a machine operator. Although the focus here is on input images and output images, the present application also generally encompasses input and output data, which can go beyond a mere visual representation of the powder bed.

[0074] Fig. 6a illustrates another embodiment. In this embodiment, anomaly-free images after exposure and / or the layer data of a component layer are used as input data during training. Autoencoder A (here variant "B") is therefore trained to represent the state of the powder bed after the scanning process ("after exposure"). This image generation is based, for example, on the input layer data. For different layers, for example, the layer data of a 3D model are input into autoencoder A, whereupon it creates an artificial representation of the powder bed after a scanning process. These generated representations are shown as "generated" in Fig. 6, for example. These generated representations can in turn be compared with the real images to create a difference image.A further advantage of this solution is that the status image can be generated, for example, in parallel to the actual scanning process in the production plant, so that the actual image can be created directly after the production process for a layer and compared with the generated image. In addition, a particularly high level of accuracy in error detection has been observed in practice. To train the autoencoder A for this design, a large number of layer data can be entered, which, after processing by the autoencoder A according to variant "B", can be compared with the actual, real, defect-free images of the powder bed ("after exposure"). Since the defect-free images of the powder bed are available in large quantities, a large data set can be used to train the autoencoder A. The precision of the trained autoencoder A can therefore be classified as correspondingly accurate.This variant of the autoencoder can be used in addition to or as an alternative to variant "A." In variant "B," information (especially photographs) of the powder bed after exposure of the previous layer and information (especially photographs) of the unexposed powder bed of the current layer are optionally used in addition to the layer data. The information before and after an exposure is thus additionally used in this embodiment. This allows the accuracy of the autoencoder to be further improved.

[0075] In a particularly preferred development, a scanning path (of the laser) of a current layer (before exposure) is combined with an image (in particular a photograph) of the unexposed powder bed of the current layer and additionally with an image of the already exposed powder bed of the previous layer. For example, the various images can be combined to form a combined image, whereby this combined image can be used to train the autoencoder in order to generate the state of the powder bed after exposure as an output image by the autoencoder. The autoencoder thus generates an output image of the current layer as it would appear after exposure. An autoencoder trained in this way can therefore be used to detect defects, for example by comparing an actual image of the powder bed (after exposure of the current layer) with the output image.

[0076] Figure 6b shows a representation of example layer data. However, this representation is only an example and the layer data may well include further or even different data or representations. This may, for example, include the use of laser data (as layer data) in the form of an image. This data, as in Figure 6b, is a representation of the cross-section of the component at a specific layer height. This data is crucial for training the autoencoder model to generate images of the powder bed after exposure. The inclusion of the layer data is advantageous for several reasons. For example, accuracy can be improved. The layer data provides a detailed view of the powder bed, which enables the autoencoder to generate accurate and high-resolution images. Another advantage is completeness.The layer data ensures that all relevant information is taken into account for model training. This is particularly advantageous because the powder bed has a complex structure after exposure. Another advantage is performance. Using layer data can improve the performance of the autoencoder model by training it more accurately to generate images of the powder bed after exposure.

[0077] As described, in a particularly advantageous further development (see Figure 6c), a combined image “B1” is created and used to train the autoencoder (modified variant “B”). For this purpose, the layer information (A3) can be combined with the information of the powder bed after exposure of the previous layer (A2) and the unexposed powder bed of the current layer (A1). The trained autoencoder thus generates an output image of the current layer as it would appear after exposure. An autoencoder trained in this way can therefore be used to detect defects, for example by comparing an actual image of the powder bed (after exposure of the current layer) with the output image.

[0078] Fig. 7 shows an example of a training process for autoencoder A according to variant "B." After starting the process, data is collected. Data collection includes, in particular, a large number of powder bed images ("after exposure") that are error-free, i.e., depict a defect-free powder bed. These powder bed images are preferably geometrically corrected, for example, by perspective correction, to enable optimal processing of the image data. Data collection, in turn, also includes the collection of layer data or slice data. Optionally, this slice data can be converted into layer images or slice images. After data collection, the actual model training of autoencoder A follows data processing. For this purpose, the collected data is used as previously described.After model training is complete, the trained autoencoder A is evaluated and tested as part of a validation process. If necessary, this is followed by an iterative training phase, for example, to further improve the precision of the autoencoder. Once the image correction performed by autoencoder A reaches a specified accuracy level, model training can be considered complete, and a final model is provided, which can be used in the additive manufacturing system.

[0079] Fig. 8 shows a further sequence of a model training process, whereby this training process only uses the actual powder bed images (“before exposure”; variant “A”), which are free of defects.

[0080] The training process, as described, is shown schematically again in Fig. 9 with corresponding representations of the images used. After the start of data collection, the actually recorded optical representations within the real images CI are subjected to a geometric correction in order to arrive at the processed images C2. These processed images are fed into the training process via data processing to train the autoencoder. The autoencoder is trained. The input representation Bl, which corresponds, for example, to one or more images C2, is processed. The autoencoder is trained in order to arrive at the output image B2. A difference image DB can be created by forming a difference or comparing it.

[0081] In a further development, an autoencoder can be used particularly advantageously, which includes variant “A” and also variant “B” and can, for example, be switched depending on the application. Fig. 10 shows an example of the method for detecting anomalies during powder bed-based additive manufacturing. In particular, this involves the optical detection of defects in the powder bed during an SLM manufacturing process for a component. In an initial step, the state of the powder bed can first be recorded at time t0. For example, this can be done via the camera system. In one embodiment, during the additive manufacturing of a component, after a scanning process for a specific layer has been completed, the powder bed is (preferably) illuminated and an image is created or recorded with the camera. For example, this time is referred to as t0.In an advantageous development, the image of the layer is also available beforehand (for example, before the powder is applied for the current layer), i.e. the image taken at time t i. After recording the actual state of the powder bed at time t and thus recording the real optical state, the data set generated can be used as input for the autoencoder A. For example, an autoencoder can be used that has different operating modes, for example a mode A and a mode B. Alternatively, two different autoencoders can be used, between which switching is possible. If the autoencoder is operated in mode A, a repaired (or corrected) image of time t can be created from the input data set, in which the defects have been filtered out. Thus, a repaired image at time t is obtained.This repaired image can in turn be used for a comparison process in which the real image at time t, which results from the capture in the powder bed, is compared with the repaired image at time t, for example to create a difference image in order to be able to detect defects in the final step and, for example, intervene in the manufacturing process. If the autoencoder A is operated in mode B, for example, the slice data, i.e. layer data, of the respective manufactured layer at time t can also be used as input. Based on the input data, a generated image can in turn be created at time t, which in turn corresponds to a representation of how the powder bed would appear in defect-free production at time t. This generated image can in turn be used as input for defect detection by comparison.In this case, a comparison of the generated image at time to with the actually acquired image to can be made, for example, to create a difference image and detect the defects.

[0082] Figure 11 shows the process of exposing a layer in additive manufacturing. In an initial state, a new layer of powder is applied to the powder bed using a coater. After the new layer has been applied, the powder bed is ready for exposure; this corresponds to the illustration on the left side of Figure 11. The state of the powder bed before exposure is also referred to as "before exposure." The next step is complete exposure, as can be seen in the adjacent image, when the laser acts on the powder bed. After the exposure is (fully) completed, the "after exposure" state is reached. The next step involves applying another layer of powder using a coater; see the illustration on the far right in Figure 11.

[0083] Artificially created defects are used to train the autoencoder. Examples of representations of a powder bed with artificially created defects are shown in Figure 12. The autoencoder can be trained using these images. Defects can be, for example, horizontal lines (in the coating direction). Such an anomaly extends, for example, only a few pixels in the Y direction (orthogonal to the coating direction) but many times over in the X direction (in the coating direction). Another defect can be, for example, a vertical line (orthogonal to the coating direction). Such an anomaly extends, for example, only a few pixels in the X direction (in the coating direction), but many times over in the Y direction (orthogonal to the coating direction). Another defect can be, for example, powder accumulation (inhomogeneities in front of and behind components; at the build plate edge).These anomalies can be represented as continuous lines, which tend to be parallel to the Y-axis and can be distinguished from lines orthogonal to the coating direction if they lie outside the component contours. These predominantly occur behind the component on the coater's path and are often accompanied by protruding support structures. A variety of other defects and defect types can be included in a defect catalog. For example, defects can result from protruding components. In addition, craters, i.e. depressions in the powder bed, can occur on and next to components. Other defects can result from a lack of powder, i.e. regions where the components are exposed. Other defects can occur due to interfering particles, such as splashes or agglomerates. The artificially created defects can be created based on these defects.

[0084] The defect detection process in additive manufacturing can, for example, be integrated into a dedicated monitoring system, which can be installed in the additive manufacturing system or independent of it and simply connected via a network. The autoencoder can advantageously be controlled in such a way that, for the highest precision, defect detection occurs by comparison in both Mode A and Mode B, so that two difference images are created for the time t0 and only those errors or defects are recorded that were detected in both difference images. In this way, a high level of accuracy in defect detection can be achieved in a particularly efficient manner. Component quality can be achieved through the continuous defect detection process, which is integrated into the manufacturing process of a component in additive manufacturing.

[0085] In summary, a method for detecting defects during the additive manufacturing of a component is proposed. The method comprises several steps, including recording the state of a powder bed layer by capturing an (optical) image of the powder bed. Subsequently, an anomaly-free image of the powder layer is created, either by inputting the image to an autoencoder and correcting the image to obtain the anomaly-free image as the output image, and / or by inputting layer data to the autoencoder to simulate the state of the powder bed and creating the anomaly-free output image based on the simulated state. Finally, the output image is compared with the image of the powder bed to detect anomalies in the powder layer. The advantages of this method lie in the precise and efficient detection of defects, leading to improved quality and reliability of additive manufacturing.

Claims

Patent claims 1. A method for detecting anomalies in powder bed-based additive manufacturing, in particular for optically detecting defects in a powder bed in an SLM manufacturing process of a component; the method comprising the steps: Detecting the condition of a powder bed layer by taking at least one image of the powder bed; Creating an anomaly-free image of the powder layer by: Inputting the at least one image to an autoencoder and correcting the image by the autoencoder to obtain the anomaly-free image as an output image; or inputting slice data and / or the at least one image to the autoencoder and creating the anomaly-free image as an output image based on the slice data and / or the at least one image; Detecting powder layer anomalies based on the output image.

2. The method according to claim 1, wherein the image of the powder bed captures a current state of at least a part of the powder bed, preferably of the entire powder bed present in the work area, the anomalies are defects of the powder layer and the at least one image of the powder bed is preferably a real image.

3. The method according to claim 1 or 2, wherein the autoencoder was trained based on a plurality of defect-free images of the powder bed and wherein preferably the output image and the image of the powder bed represent the same layer of the layer structure of the component to be manufactured.

4. The method according to claim 3, wherein the defect-free images of the powder bed are subjected to artificially generated anomalies for training the autoencoder, and the defect-free images preferably show the powder bed before the exposure process.

5. Method according to at least one of the preceding claims, wherein the artificially generated anomalies are created automatically based on an error database.

6. Method according to at least one of the preceding claims, wherein for training the autoencoder a plurality of anomaly-free output images are provided, into which one or more artificially generated anomalies are inserted in order to obtain modified output images and wherein the autoencoder is trained to filter the anomalies from the modified output images.

7. Method according to at least one of the preceding claims, wherein, in order to correct the image of the powder bed, the autoencoder compresses the image via a filter and then generates the output image via a decoder 8. Method according to at least one of the preceding claims, wherein the autoencoder has been trained at least based on the layer data in order to generate the state of the powder bed after application of a further layer as an output image.

9. The method according to claim 8, wherein the autoencoder simulates a manufacturing step based on the layer data and creates a generated image of the state of the powder bed and in particular of an expected defect-free surface of the powder bed after a scanning process.

10. Method according to at least one of the preceding claims, wherein the layer data comprise geometric data of the components to be manufactured and / or manufacturing parameters of the manufacturing plant.

11. Method according to at least one of the preceding claims, wherein the layer data comprise laser target position data which indicate at least one predetermined position in the powder bed at which the laser processes, in particular melts, the powder material.

12. Method according to at least one of the preceding claims, wherein the state of a current powder bed layer is detected during a manufacturing process and, for this purpose, the powder layer in the working space of the additive manufacturing system is illuminated from at least one direction and the image of the powder layer is recorded.

13. Method according to at least one of the preceding claims, wherein if at least one anomaly is detected, a message is output and / or an active intervention in the manufacturing process of the component takes place.

14. Method according to at least one of the preceding claims, wherein the comparison of the output image with the image of the powder bed to detect anomalies of the powder layer is carried out after a scanning process and if no anomaly was detected, the scanning process is released for the next layer.

15. A computer-implemented method for detecting anomalies in powder bed-based additive manufacturing, comprising the steps of at least one of the preceding claims.

16. A method for producing a component by means of additive manufacturing, wherein, during the production of a component by an additive manufacturing system, the steps for detecting anomalies according to at least one of the preceding claims are carried out several times, in particular after each scanning process.

17. A method for training an autoencoder for detecting anomalies in powder bed-based additive manufacturing on the basis of training data, in particular training images; the method comprising the steps of: providing a set of defect-free training images of a powder bed; Incorporation of artificially generated anomalies as form errors into the defect-free training images in order to obtain modified training images; Input of the modified training images into the autoencoder and adjustment of the weighting of the autoencoder such that the input modified training images are filtered by the autoencoder to produce defect-free training images as output images.

18. Device for data processing in the additive manufacturing of objects, in particular for the optical detection of defects in a powder bed in an SLM manufacturing process, comprising means for carrying out the steps of the method according to at least one of the preceding claims.

Images