Method and device for additive manufacturing of workpieces
By using previously trained statistical learning models in additive manufacturing to check the material layer image, identify layer defects and control the curing of particulate materials, the problem of material layer quality monitoring in additive manufacturing is solved, and efficient quality control and defects of the workpiece layer are achieved. correct.
Patent Information
- Application Number
- CN202180050499.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2021-01-29
- Filing Date
- 2021-07-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-07-23
AI Technical Summary
In the additive manufacturing process, it is difficult to conduct efficient quality monitoring of material layers, which makes it difficult to detect and correct layer defects early, which may lead to workpiece quality problems.
The previously trained statistical learning model was used to inspect the image of the material layer, identify possible layer defects by identifying defect vectors, and control selective curing of the particulate material based on this, including inspection of the material layer before and after selective curing to correct abnormalities .
It realizes efficient quality monitoring of additive manufacturing workpiece layers, early identification and correction of defects, avoid defect propagation, save materials and time, and ensure workpiece quality.
Smart Images

Figure CN115867403B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for additive manufacturing of a workpiece, the method comprising the following steps:
[0002] a) receiving a data set defining a workpiece in a plurality of successively arranged workpiece layers,
[0003] b) producing a material layer having a defined surface formed by particulate material,
[0004] c) recording at least one image of the material layer and examining the material layer based on the at least one image in order to determine respective properties of the material layer,
[0005] d) selectively solidifying the particulate material on the defined surface by means of a build tool using the data set and based on individual properties of the material layer, wherein the defined workpiece layer of the plurality of successively arranged workpiece layers is produced from the material layer, and
[0006] e) Repeating steps b) to d), wherein further defined workpiece layers of the plurality of successively arranged workpiece layers are generated.
[0007] The present invention also relates to a device for additive manufacturing of a workpiece, the device having a memory for receiving a data set defining the workpiece as a plurality of continuously arranged workpiece layers, having a production platform, having a layer forming tool, having a building tool, having a camera oriented toward the production platform, and having an evaluation and control unit, which is configured to generate a material layer having a defined surface from a granular material on the production platform with the aid of the layer forming tool, is further configured to record at least one image of the material layer with the aid of the camera, and is further configured to selectively solidify the granular material on the defined surface with the aid of the building tool, wherein a defined workpiece layer in the plurality of continuously arranged workpiece layers is generated from the material layer. Background Art
[0008] Such a method and such a device are known in principle from US 2015 / 0061170 A1.
[0009] Additive methods for the manufacture of workpieces are sometimes referred to as 3D printing. There are various additive manufacturing methods. In selective laser sintering (SLS) or selective laser melting (SLM), a so-called powder bed formed of a granular material is used. The granular material is usually a metallic material. However, there are also methods involving granular synthetic materials, in particular polymers. With the help of a laser beam or an electron beam, selected powder particles on the top of the powder bed are locally and selectively melted or at least begin to melt, and are therefore selectively solidified when cooled. Then, a new powder layer is distributed on the workpiece structure and the remaining unmelted powder, thereby manufacturing the workpiece in layers. The individual workpiece layers are usually produced from the bottom up on a production platform, which reduces the corresponding layer height with each workpiece layer.
[0010] Additive manufacturing of workpieces makes it possible to manufacture individual workpieces with a high degree of complexity and low material costs. However, at the same time, there are major challenges in terms of workpiece quality, since anomalies that may lead to workpiece defects may occur in any individual material layer. Anomalies may lead to defects such as holes in the layer structure, micropores, local layer detachment / delamination, tears, protrusions, shape differences and / or material stresses inside and / or on the surface. For this reason, there are many suggestions for detecting defects as early as possible during the manufacture of the layer sequence of the additively manufactured workpiece. The above-mentioned US2015 / 0061170A1 discloses, in particular, an optical measuring sensor with a camera, which can be configured to enable 3D coordinate measurement on the corresponding uppermost material layer. One difficulty in this case is reflections and shadows, which may be caused in particular by metal powder particles, but may also be caused by other objects in the working space of the device.
[0011] DE 10 2016 201 289 A1 discloses a method for additive manufacturing of workpieces, wherein first measurement data are acquired during the additive construction by thermographic material testing or by eddy current material testing. After the additive construction, second measurement data are acquired by means of computer tomography and compared with the first measurement data. The aim is to classify the results of the material test using an algorithm from the field of supervised machine learning (not described in more detail).
[0012] EP 3 459 715 A1 discloses a method for additive manufacturing of workpieces, wherein the aim is to predict defects in the current layer or in subsequent other layers with the aid of a classification function which has been trained using techniques from the field of machine learning (not described in greater detail).
[0013] WO 2015 / 020939 A1 discloses a method for additive manufacturing of workpieces based on filament machining, wherein the aim is to correlate a CAD input data set with parameters of a device used with the aid of an algorithm from the field of machine learning in order to predict the properties of a workpiece manufactured by the device and the time required for manufacturing. The actual quality control of the manufactured workpiece is carried out after the workpiece has been manufactured using a 3D scanner and with the aid of a calibration pattern on a production platform.
[0014] DE 10 2017 108 874 A1 and US 2020 / 158499 A1 of the same priority disclose an optical system for material testing by means of illumination from a plurality of different directions. In some variants, the system can be used to determine a height map of a material layer to be tested. Summary of the invention
[0015] In this context, an object of the present invention is to specify an alternative method and a corresponding device for the additive manufacturing of high-quality workpieces. In particular, an object is to monitor the quality of material layers efficiently and in close proximity to the process, so that any occurring or indicated layer defects can be corrected at an early stage.
[0016] According to a first aspect of the invention, the object is achieved by a method of the type mentioned at the outset, wherein at least one image of the material layer in step c) is examined using a previously trained statistical learning model, wherein the previously trained statistical learning model is used to determine at least one defect vector representing a plurality of individual defect probabilities, wherein each individual defect probability of the plurality of individual defect probabilities is an individual indicator as to whether a defined layer defect from a plurality of possible layer defects is present in the material layer, and wherein step d) is performed based on the at least one defect vector.
[0017] According to a further aspect, the object is achieved by a device of the type mentioned at the outset, wherein the evaluation and control unit is further configured to inspect at least one image of the material layer using a previously trained statistical learning model, wherein the previously trained statistical learning model is used to determine at least one defect vector representing a plurality of individual defect probabilities, wherein each individual defect probability of the plurality of individual defect probabilities is an individual indicator as to whether a defined layer defect from a plurality of possible layer defects is present in the material layer, and wherein the evaluation and control unit controls a layer forming tool and a building tool based on the defect vector.
[0018] A method for training a statistical learning model used in the above method and / or the above device is also proposed herein, the method comprising the following steps:
[0019] - obtaining a plurality of training images showing a plurality of material layers, wherein the plurality of material layers contain a plurality of possible layer defects,
[0020] - marking the plurality of possible layer defects in the plurality of training images so as to obtain a plurality of annotated versions of the training images,
[0021] - define stopping criteria based on multiple annotated versions of these training images, in particular defect criteria and / or time criteria,
[0022] - repeatedly providing the plurality of training images as input data to a statistical learning model, and repeatedly obtaining output data from the statistical learning model in response to the plurality of training images provided accordingly, and
[0023] - Iteratively optimize the statistical learning model using the multiple training images and a stopping criterion.
[0024] The plurality of training images preferably comprises a plurality of groups, which have in each case at least three, in particular in each case four training images, wherein the training images of each group each show a defined material layer with at least one layer defect, and wherein the training images of each group show a defined material layer with mutually different irradiation directions. In terms of number and irradiation, the training images preferably correspond to those images which were examined using a previously trained statistical learning model according to the novel method for additive manufacturing of workpieces, and the following explanations therefore also apply analogously to the training images. In some preferred exemplary embodiments, the training images are recorded during the manufacture of the defined workpiece, and the stop criterion is defined based on the defined workpiece. In these preferred exemplary embodiments, the already trained learning model can advantageously be “retrained” in the current manufacturing process for manufacturing a single workpiece, so that the learning model is optimized with respect to the single workpiece. Then, using the learning model retrained in this way, other workpieces of the same type can be manufactured very advantageously in subsequent further manufacturing processes. In some advantageous exemplary embodiments, the learning model can be iteratively retrained based on a plurality of manufacturing processes in which workpieces of the same type are manufactured in each manufacturing process until one or more specified workpiece properties are achieved. The stop criterion can advantageously be defined based on the specified workpiece properties. In a preferred exemplary embodiment, a plurality of possible layer defects may be marked by a person entrusted with detecting layer defects.
[0025] The statistical learning model of the novel method and the novel device originates from the field of machine learning. Here, the statistical learning model implements a statistical evaluation of at least one image of the material layer based on previously trained parameters and delivers individual probability values, each probability value representing an individual probability of the presence of a defined layer defect in a plurality of possible layer defects. In a preferred exemplary embodiment, the plurality of possible layer defects include ridges on the surface of the material layer formed by the granular material, local agglomerations of the granular material on the defined surface, uneven particle sizes of the granular material, holes or depressions on the surface of the material layer formed by the granular material, and undesirable adhesion or melting of material particles. Such anomalies / inhomogeneities may lead to the workpiece defects mentioned further above. Since the material layer is checked for the presence of such anomalies with the aid of a previously trained statistical learning model, the occurrence of defects can be identified at an early stage. Therefore, the propagation of defects can be avoided or corrected, or, as an alternative, the printing process can be stopped to save material and time. Therefore, the selective solidification of the granular material in step d) is performed based on at least one defect vector. This may include not performing step d) at all due to identified or occurring defects, or performing step d) in a modified manner (e.g. with modified CAD data), or first reprocessing the material layer formed from the particulate material in step b) and then performing selective curing on the reprocessed material layer in step d).
[0026] The efficient use of previously trained statistical learning models allows the inspection of material layers in the vicinity of the process before and after the selective solidification of the granular material. The inspection of the material layer before the selective solidification makes it possible to correct any anomalies before the selective solidification, for example by smoothing the surface of the material layer again, distributing additional granular material and / or replacing the existing granular material. In addition, the inspection of the material layer after the selective solidification makes it possible to correct identified defects on a workpiece that is only partially manufactured by reworking an already manufactured workpiece layer (e.g., starting to melt the workpiece layer or causing the workpiece layer to melt) and / or by modifying a subsequent workpiece layer (e.g., making the workpiece layer thicker or thinner).
[0027] The statistical learning model delivers individual defect probabilities for various layer defects and allows for an empirically based surface inspection without having to know precisely in advance the exact appearance of each individual layer defect in at least one image of the material layer. The statistical learning model preferably provides individual defect probabilities for a plurality of different layer defects in each case so that for each layer defect under consideration an individual defect probability is obtained which represents the presence or absence of the respective layer defect. In contrast to the method mentioned at the beginning, the algorithm from the field of machine learning is therefore applied here in a targeted manner to the inspection of a powder bed comprising a material layer formed of granular material, rather than or only to the behavior and properties of a device for additive manufacturing and / or in connection with already manufactured workpiece layers. The novel method and the corresponding device contribute very efficiently to the early and proactive avoidance of defective workpieces and workpiece layers. The defect vector containing the individual defect probabilities for various layer defects very efficiently makes it possible to give qualitative and (at least in some exemplary embodiments) even quantitative statements relating to the quality properties of the manufactured workpiece (e.g. the size or distribution of the identified layer defects).
[0028] In some exemplary embodiments, selective solidification of the granular material can be stopped or postponed based on the defect vector until the material layer formed by the granular material has the desired uniformity achieved by appropriate reprocessing. In some exemplary embodiments, the manufacturing process can be stopped in advance when a layer defect is identified with a high probability that a defect-free workpiece cannot be expected. Since various layer defects may occur at the same time or at different times from each other, locally away from or directly close to each other or to different degrees according to a specific process sequence, it is particularly suitable to inspect the powder bed based on a machine learning method using a trained statistical learning model. When process parameters change, the statistical learning model can be efficiently "retrained". Therefore, the above-mentioned purpose is fully achieved.
[0029] In a preferred refinement of the invention, in step c) the material layer is irradiated from a plurality of mutually different directions and a plurality of images of the material layer are recorded, wherein each of the plurality of images shows the material layer with a different irradiation direction, and wherein the plurality of images are used to determine individual properties.
[0030] Refinement is particularly advantageous for inspecting powder beds formed by metallic granular materials. However, it can be used in the same way for inspecting powder beds formed by synthetic or mixed materials. Multiple images show material layers with mutually different light reflections and mutually different shadows. Therefore, the individual properties of the material layer can be acquired more reliably and in more detail. It is advantageous to record multiple images with a single camera arranged in a fixed position relative to the production platform and / or the material layer. This allows fast image recording and a simple assignment of different irradiation images to each other. Refinement makes it easier to semantically distinguish different layer defects in particular, because it allows a detailed inspection of the material surface. In step c), the material layer is preferably irradiated from at least three, in particular from four mutually different directions, and the multiple images in step c) are correspondingly composed of at least three, in particular four images, which show the material layer in each different irradiation direction. In a preferred exemplary embodiment, the image shows the material layer recorded by the camera with corresponding irradiation. Image distortion can be advantageously corrected and / or corrected. In addition, irradiation inhomogeneities, such as those caused by production tolerances of the light source used, can be corrected. However, it is preferred that the images in exemplary embodiments show the material layer itself, rather than, for example, a filtered view of the material layer or an otherwise modified view of pixel information, as this allows for particularly fast and efficient inspection during an ongoing manufacturing process.
[0031] In a further refinement, multiple images are provided together as input data to a previously trained statistical learning model.
[0032] In this refinement, each of the multiple images can form an inspection channel. The statistical learning model can process image information from various illumination images together, and thus associate various image information. For example, an edge in a material layer may be visible when illuminated from a first direction, while the edge is invisible when illuminated from another direction. On the other hand, the reflection occurring in the case of one illumination direction may resemble a layer defect that does not actually exist. Refinement advantageously helps to identify as many layer defects that actually exist as possible, and also helps to distinguish between layer defects that actually exist and layer defects that only appear to exist. Established statistical learning models from the field of machine learning (some of which will be described in further detail below) generally allow for joint processing of multiple input data sets, such as separate RGB color channels of a multi-color image. Refinement advantageously uses this ability for correlation analysis of multiple images with different illumination directions. Therefore, this refinement can be implemented relatively easily, and it allows very fast inspection of the material surface.
[0033] In a further refinement, at least one image is used in step c) to determine a height map of the material layer, wherein the height map is provided as input data set to a trained statistical learning model.
[0034] In a preferred exemplary embodiment, a height map is determined using multiple images showing the material layer in each case with different illumination directions. Preferably, the height map is determined as a 2.5D height map using the method described in DE 10 2017 108 874 A1 mentioned at the beginning and US2020 / 158499 A1 with the same priority, which are hereby incorporated by reference. As an alternative or in addition, in a further exemplary embodiment, the height map can be determined according to the triangulation principle (for example according to the strip light projection principle). The advantage of refinement is that the statistical learning model can obtain the height information of the current measurement in each case. This can reduce the scope and training time of the training data required in advance for the statistical learning model. In addition, highly critical layer defects, such as material missing or deep bulges caused in particular by damage to the layer forming tool, can be identified very quickly based on the height map.
[0035] In a further refinement, in step c), a plurality of defect vectors are determined, each of which represents a plurality of individual defect probabilities, wherein each defect vector of the plurality of defect vectors represents an individual defect probability associated with a selected pixel region in the at least one image. In this refinement, the selected pixel regions at least partially differ from each other, so that the defect vectors represent mutually different pixel regions of the at least one image.
[0036] It is particularly advantageous to represent, for each defect vector of the plurality of defect vectors, an individual defect probability associated with another individual pixel in the at least one image.
[0037] In these refinements, the defect vectors indicate the probability of defining a layer defect in a locally delimited area, in particular at the pixel level. The refinement makes it possible to determine the extent and / or shape of any layer defects that are present and / or their spatial contour. The defect vectors associated with individual pixels in at least one image in particular allow the size of the identified layer defects to be determined with high precision, so that, for example, the size of pores in the layer sequence of the workpiece can be estimated. The refinement contributes to a particularly efficient implementation of the novel method and the corresponding device, since, based on the individual requirements for the manufactured workpiece, layer defects that do not exceed a certain size, shape or form can be tolerated in a targeted manner. By means of this refinement, an advantageous classification of the individual layer defects and the workpiece defects that may arise therefrom also becomes easier.
[0038] In a further refinement, a plurality of defect vectors are used to determine morphological properties and / or dimensional properties of defined defects in the material layer.
[0039] Morphological properties contain information about the structure and / or type of defined defects. Dimensional properties contain information about the extent and / or shape. Determining morphological and / or dimensional properties makes it easier to classify any layer defects and to make targeted decisions about whether the manufactured workpiece can meet the defined specifications regarding robustness, durability, shape and / or size. Therefore, refinement is particularly beneficial for efficient process analysis.
[0040] In a further refinement, the previously trained statistical learning model comprises a convolutional neural network (CNN). The trained statistical learning model particularly preferably comprises a convolutional neural network with an encoder / decoder architecture, such as a so-called U-net.
[0041] The convolutional neural network uses in particular the mathematical operation "convolution" to analyze the input data set. A filter matrix is convolved with at least one image segment in a plurality of steps. The result of this convolution is a data set in which the presence or absence of the feature represented by the filter can be estimated. A plurality of such convolution operations are preferably performed sequentially in succession. A filter matrix is advantageously used for each relevant layer defect. By way of example, in each case one or more filter matrices can be provided for one or more of the following layer defects: ridges on the surface of a material layer formed by granular material, local agglomerations of granular material on a defined surface, uneven particle size of granular material, holes or depressions on the surface of a material layer formed by granular material, adhesion or melting of material particles. The plurality of convolution operations can comprise conventional convolution operations (convolutions) and / or modified convolution operations, such as strided convolutions, dilated convolutions or transposed convolutions, and produce a defect vector together with further operations performed by the convolutional neural network (e.g. by means of a rectified linear unit ReLU), such further operations in particular as normalization operations, pooling operations and delinearization. As demonstrated by tests, such a convolutional neural network makes it possible to determine defect vectors representing defects in the layer in an efficient manner.
[0042] A convolutional neural network with an encoder / decoder architecture (e.g. based on a U-net model) supplements the determination of the defect vector in a further subsequent convolution step with information from a first convolution step (known as upsampling) and thus segments the at least one image very accurately into different layer defect areas and a defect-free image background. Such convolutional networks are often used in the medical field. Tests have shown that such networks are very suitable for the specialized inspection of material layers formed by granular materials in order to detect layer defects in an accurate manner even at a very early stage.
[0043] In a further refinement, at least one image of the material layer is normalized by means of a reference image, wherein the reference image shows a homogeneous diffuse reflecting surface or a defect-free material layer.
[0044] This refinement improves the accuracy when detecting any layer defects, since the previously trained statistical learning model obtains one or more input images that are independent of the individual irradiation. Unfavorable influences caused by production tolerances during irradiation, aging phenomena, different relative positions of individual irradiation elements in the working space of the device with respect to the camera, etc. can thereby be reduced. In some exemplary embodiments, the uniform diffuse reflecting surface of the reference image can be a white paper. The reflection properties of the diffuse reflecting surface preferably correspond to a Lambertian emitter. In other preferred exemplary embodiments, each of the multiple images is normalized with a reference image, which shows a defect-free material layer in the irradiation direction in which the corresponding image of the multiple images was recorded. The image of the defect-free material layer can advantageously be low-pass filtered. Such normalization results in particularly good inspection results.
[0045] In a further refinement, at least one defect vector is stored as a historical defect vector together with a time stamp identifying the material layer, wherein further defined workpiece layers are generated based on the historical defect vector.
[0046] In some exemplary embodiments, it is therefore possible to check in a repeated step c) (that is to say in the inspection of subsequent material layers) whether a layer defect detected in the current material layer persists on multiple layers. Refinement helps to reduce "false positives" and helps to stop and / or modify the selective solidification of the granular material of the current material layer only when it is actually necessary to achieve the desired workpiece quality. For example, individual layer defects that are limited to one material layer, or even obvious layer defects that do not actually exist at all and appear to exist, for example due to light reflections, can be efficiently eliminated. In some exemplary embodiments, after the inspection with the aid of a statistical learning model, the corresponding current defect vector can be compared in a deterministic manner with one or more historical defect vectors, that is to say as part of post-processing. For example, potential layer defects can be ignored when the historical defect vectors do not indicate the same layer defect in a previous material layer.
[0047] In a further refinement, at least one image of the material layer is stored as a historical image together with a time stamp identifying the material layer, wherein the further defined workpiece layer is generated based on the historical image.
[0048] In this refinement, which can be used as an alternative or supplement to the above-described refinement, a history of the identified layer defects is provided with the aid of the recorded images. The refinement makes it possible to provide one or more historical images together with the corresponding current image as a common multi-channel input data set to a trained learning model. The statistical learning model can then advantageously perform an inspection of the material surface taking into account the history and identify temporal correlations. In a preferred exemplary embodiment, at least one current image and one or more historical images can each form a channel of a convolutional neural network. In a further exemplary embodiment, the statistical learning model can have a short-term memory, for example implemented by an LSTM (Long Short-Term Memory) network. In some exemplary embodiments, the statistical learning model can accordingly be a convolutional neural LSTM network.
[0049] The refinement can be particularly advantageously combined with the determination of a height map of a material layer, wherein a convolutional neural network is applied which performs a convolution operation in three dimensions. For example, the first two dimensions can be spatial pixel information along the X-axis and the Y-axis of the height map, while the third dimension of the convolution operation can be time, wherein a current height map and one or more historical height maps are used. The input data set can be, for example, a tensor whose dimensions correspond to the width and height of the height map and the number of historical and current height maps. The refinement allows a very advantageous implementation of novel methods and corresponding devices and monitoring of manufacturing processes and process sequences for a plurality of different workpieces.
[0050] In a further refinement, the individual properties of the material layers in the repeated step c) are each stored together with a time stamp identifying the respective material layer, wherein the workpiece is authorized for use based on the stored individual properties.
[0051] In this refinement, after the manufacturing process is finished, the material layer inspection accompanying this process is advantageously used to assess the suitability of the manufactured workpiece for its intended use. The stored individual properties can also be advantageously used to document the quality assurance process. Refinement contributes in a very efficient way to achieving high product quality in additive manufacturing processes.
[0052] It goes without saying that the features mentioned above and the features yet to be discussed below can be used not only in the respectively specified combination but also in other combinations or alone, without departing from the scope of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Exemplary embodiments of the present invention are shown in the drawings and are explained in more detail in the following description. In the drawings:
[0054] Figure 1 A schematic diagram showing an exemplary embodiment of the novel device is shown,
[0055] Figure 2A flow chart is shown for illustrating an exemplary embodiment of the novel method.
[0056] Figure 3 The diagram is used to illustrate the Figure 2 A flow chart of an exemplary embodiment of a method for material surface inspection,
[0057] Figure 4 shows a simplified diagram for illustrating the operation of a convolutional neural network that can be advantageously used in exemplary embodiments of the novel method and the novel apparatus,
[0058] Figure 5 A simplified diagram of an exemplary embodiment having an encoder / decoder architecture is shown, and
[0059] Figure 6 A simplified diagram of a further exemplary embodiment with an encoder / decoder architecture is shown. DETAILED DESCRIPTION
[0060] exist Figure 1 In the present invention, an exemplary embodiment of the novel device is designated as a whole by reference numeral 10. The device 10 has a production platform 12, on which a workpiece 14 is additively manufactured according to an exemplary embodiment of the novel method. The workpiece 14 is constructed layer by layer from the bottom to the top of a material stack in temporally consecutive steps. Reference numeral 16 indicates the instantaneous uppermost workpiece contour or the uppermost workpiece layer. The uppermost material layer that produces the workpiece layer 16 is designated by reference numeral 18.
[0061] After the respective workpiece layer 16 has been completed, a new material layer 18 of granular material 20 (e.g. metallic material and / or synthetic material) is distributed on the layer stack by means of a scraper 22. To this end, the production platform 12 is usually lowered in the direction of the arrow 24 by the height of the next material layer, and the granular material 20 is removed from a reservoir 26 and distributed on the existing layer stack by means of the scraper 22.
[0062] Reference numeral 28 shows a simplified form of the build tool. In some exemplary embodiments, the build tool 28 generates a laser beam 30 and moves it relative to the production platform 12 and the material layer to be built 18. The laser beam 30 selectively melts and / or begins to melt the material particles so that they solidify when cooled.
[0063] In other exemplary embodiments, build tool 28 may generate an electron beam to build workpiece layers on production platform 12. In some exemplary embodiments, apparatus 10 may include more than one build tool 28, that is, use two or more laser beams and / or electron beams to generate workpiece layers, for example.
[0064] The building tool 28 (hereinafter sometimes referred to as a scribing laser) is connected to an evaluation and control unit (hereinafter referred to as a controller 32), which controls the movement of the laser beam 30 along the material surface. The controller 32 has an interface 34, through which a data set 36 can be read in, which defines the workpiece 14 to be manufactured in a plurality of successively arranged layers. The controller 32 controls the movement of the laser beam 30 relative to the material stack based on the data set 36, wherein the laser beam 30 describes a trajectory generated by the data set 36 in each workpiece layer 16 to be manufactured. In some exemplary embodiments, the controller 32 is implemented by means of one or more commercially available personal computers, on which an operating system (for example Microsoft Windows, MacOS or Linux) and one or more control programs for implementing exemplary embodiments of the novel method are run. In some exemplary embodiments, the controller 32 can be implemented in the form of a soft SPS on a commercially available PC. As an alternative or in addition, the controller 32 can be implemented by means of dedicated control hardware including one or more ASICs, FPGAs, microcontrollers, microprocessors or equivalent logic circuits.
[0065] The equipment 10 also has a measuring device 38, 40, which is configured to check the surface of the layer stack. The measuring device includes an irradiation device 38 and a camera 40, which are each connected to the controller 32 (or connected to a separate controller for the measuring device, not shown here). In this preferred exemplary embodiment, the camera 40 is configured to record multiple images of the surface of the material stack, wherein the surface is irradiated from different directions. For example, the irradiation device 38 includes a plurality of irradiation modules 38a-38f, which are arranged at different positions relative to the production platform 12. Alternatively or additionally, the irradiation device 38 can be moved relative to the production platform 12 to irradiate the material surface from different directions. For example, the production platform 12 can be arranged on a turntable.
[0066] exist Figure 1In the figure, three irradiation modules 38a, 38b, 38c are arranged adjacent to each other. The irradiation modules 38a, 38b, 38c are therefore capable of generating three irradiation directions 42a, 42b (not shown here), 42c running substantially in parallel. Three further irradiation modules 38d, 38e, 38f are arranged here on the second side of the production platform 12, transverse to the irradiation modules 38a, 38b, 38c and parallel to each other. The irradiation modules 38d, 38e, 38f are capable of generating three further irradiation directions 42d (not shown here), 42e, 42f. In some exemplary embodiments, the device 10 has six further irradiation modules (not shown here), three of which may be arranged opposite to the irradiation modules 38a-38c, and the other three may be arranged opposite to the irradiation modules 38d-38f. In an exemplary embodiment, the device 10 is capable of irradiating the material surface from at least three different main directions. As Figure 1 As shown, in some exemplary embodiments, the main directions can each contain three illumination sub-directions that are offset parallel to one another. This allows a very advantageous determination of a 2.5D height map of the material surface according to the method described in DE 10 2017108 874 A1 mentioned at the beginning or US 2020 / 158499 A1 with the same priority. Both documents are hereby incorporated herein by reference.
[0067] As Figure 1 As an alternative or in addition to the arrangement shown, the device 10 can have a ring light with a plurality of light sources that can be driven individually and / or in segments, wherein the ring light (not shown) is, for example, arranged above the production platform 12 and preferably around the material stack so as to allow illumination of the material surface 18 from different directions.
[0068] can be controlled by one or more control programs according to Figure 1 Exemplary embodiments of the novel method implemented on the device are additionally referred to below Figures 2 to 6 Explanation. In step 50, a data set 36 defining the workpiece 14 in a plurality of successively arranged workpiece layers 16 is read into the controller 32. As an alternative or in addition thereto, the controller 32 may first receive a data set defining the workpiece to be manufactured "as a whole", such as a CAD data set, via the interface 34, and determine a plurality of successively arranged workpiece layers 16 based thereon. Also in this case, the controller 32 ultimately receives a data set defining the workpiece 14 in a plurality of successively arranged workpiece layers 16. In step 52, a material layer 18 is produced on the layer stack. For example, the controller 32 may remove the granular material 20 from the reservoir 26 by means of a scraper 22 and distribute it on the layer stack. The distribution of the granular material 20 should be as consistent and uniform as possible.
[0069] In step 54, the surface of the material layer 18 is then checked with the aid of the measuring devices 38, 40 in order to identify any anomalies, in particular such as ridges, holes, depressions, undulations, material agglomerates, thickness variations and / or grain inhomogeneities (e.g. agglomerations) in the material layer 18. If the surface of the new material layer 18 meets all desired criteria, the method branches in step 56 to step 58, in which a new workpiece layer 16 is produced in the uppermost material layer 18 with the aid of the scribing laser 28. The scribing laser 28 selectively melts the material grains along defined paths, thereby connecting the melted grains or grains that have already begun to melt to one another.
[0070] If the surface of the new material layer 18 does not meet or does not sufficiently meet the desired criteria, the method may return to step 52 in loop 60 to re-process or completely regenerate the surface of the material layer 18. In step 62, steps 52-58 are repeated until the workpiece 14 is completed according to the data set 36.
[0071] In some exemplary embodiments, the newly manufactured workpiece layer 16 can be inspected in a targeted manner by means of the measuring devices 38, 40, which is indicated by reference numeral 64. Based on this, the subsequent workpiece layer can then be modified in order to correct, for example, shape or size differences. In step 66, based on the inspection history from repeated steps 52 and / or 64, the manufactured workpiece can be authorized for the intended use.
[0072] Figure 3 Shown in Figure 2 In an exemplary embodiment of the material layer 18, the surface of the material layer 18 is preferably irradiated with short-wave light. In step 68, the material surface is irradiated from a first direction. The first direction can be, for example, according to Figure 1 In step 70, a first image I of the illuminated material surface is recorded by means of the camera 40. In step 72, the material surface is then illuminated from another direction, for example from a direction according to Figure 1 In step 74, a further image J of the material surface is recorded by means of the camera 40. k , while illuminating the material surface from another direction 42c. In step 76, it is determined whether the material surface should be recorded from another direction (e.g., from the material surface according to the embodiment of the present invention). Figure 1 In some exemplary embodiments, the first illumination 68 and the further illumination 72 may be generated simultaneously with mutually different wavelengths / light colors, and the different images may be separated from each other based on the different wavelengths / light colors.
[0073] In step 78, in some preferred exemplary embodiments, the recorded images are normalized using a correspondingly selected reference image in order to obtain a consistent irradiation level independently of the arrangement of the irradiation module in the working space of the device 10 and any production tolerances. In some exemplary embodiments, the reference image can show a white sheet of paper recorded by the camera 40. In other exemplary embodiments, the recorded images are each normalized with an associated reference image showing a defect-free material layer including granular material and an irradiation that is in each case identical to the irradiation of the recorded image to which the normalization was performed. In some exemplary embodiments, the recorded and possibly normalized images can be used to determine a 2.5D height map of the material surface 18, as has been further mentioned above with reference to DE 10 2017108 874 A1 or US 2020 / 158499 A1. In step 80, the height map and / or the normalized image is provided here with a corresponding time stamp representing the material layer 18 currently being recorded. In some exemplary embodiments, the time stamp can be a serial number that is incremented with each new material layer 18, independent of the current time.
[0074] In steps 82, 84, 86, 88, the height map and / or normalized image from step 78 is provided to a previously trained statistical learning model, which is preferably implemented as a convolutional neural network. The basic operation of such a network is Figures 4 to 6 Explained in more detail in.
[0075] according to Figure 4 , where one or more images 90a, 90b of a material surface 18 are provided to a statistical learning model as an input data set. In some exemplary embodiments, image 90a contains a current height map of an uppermost material layer 18, while image 90b contains a height map of a previous material layer located further below. Multiple such historical height maps of previous material layers of the layer sequence may be compared to the current height map. Figure 1 The stack of input images 90a, 90b may contain height maps for two, three, four or five previous material layers.
[0076] In other exemplary embodiments, the stack of input images 90a, 90b may in each case contain a current (preferably normalized) image of the uppermost material layer 18 with a respectively different illumination direction. Thus, in these exemplary embodiments, the determination of a height map may be dispensed with. In further exemplary embodiments, the stack of input images 90a, 90b may contain a current image and a historical image with a respectively different illumination direction. In further exemplary embodiments, the stack of input images 90a, 90b may contain a current image and / or a historical image with a respectively different illumination direction and one or more (current and / or historical) height maps.
[0077] Then, the input images 90a, 90b are each convolved with one or more filter masks 92a, 92b. The convolution operation generates an image stack 94, which contains the corresponding convolution results of each input image 90a, 90b and each filter mask 92a, 92b. In some preferred exemplary embodiments, the stack 94 also includes a step referred to as a ReLU step by those skilled in the art. Using the ReLU step (according to Figure 3 Step 84) delinearizes the stack 94 in a targeted manner. Figure 3 , which may be followed by a so-called pooling step, which may in particular be implemented as so-called maximum pooling. The pooling step 86 produces a stack 96 in which redundant image information is reduced. Figure 3 As indicated by reference numeral 88 in FIG. 8 , steps 82, 84, 86 may be repeated multiple times, wherein the stack from the previous steps is used in each case as an input data set for further convolution and pooling steps. In some exemplary embodiments, between three and ten convolution steps 82, ReLU steps 84 and pooling steps 86 may be performed in cascade in succession. It is also possible to perform a series of multiple convolution steps and / or delinearization steps before performing the corresponding pooling step.
[0078] refer to Figure 3 In a further step 98, one or more defect vectors 100 may be determined, in particular using a so-called Softmax function, which transfers the probability distribution based on the previous stack 96'. Thus, each defect vector 100 contains a plurality of individual defect probabilities 102a, 102b, wherein each individual defect probability 102a, 102b is an individual indicator as to whether a defined layer defect from a plurality of the above-mentioned layer defects is present in the inspected material layer 18.
[0079] Possible layer defects are associated with filter matrices 92a, 92b, which in turn can be predetermined based on the provided training data. In some preferred exemplary embodiments, the training data comprises a height map and / or a normalized image of a material surface containing one or more of the above-mentioned layer defects, and at least one height map and / or image of a defect-free material surface. In order to generate such training data, the manufacturing process can be interrupted in a targeted manner, for example by overfilling powder material, introducing ridges or undulations on the material surface, mechanical vibrations, targeted incorrect driving during the printing process, such as too fast movement of the scribing laser and / or too high scribing intensity, etc. The images or height maps recorded in this way can also be reflected or manipulated using image processing methods in order to obtain a wide range of training data. The training data are preferably provided to the convolutional neural network 112 in advance in a plurality of training cycles. The resulting defect vectors are analyzed for each provided training data set. Various defect metrics can be envisioned for analysis, which can be optimized during the learning step, such as the average error over all pixels of the defect map (usually cross entropy) or a perceptual defect metric (e.g. adversarial loss). As part of a so-called back-propagation algorithm, the parameters of the filter masks 92a, 92b are modified using a suitable optimization method until the selected objective function is sufficiently optimized, e.g. until the resulting defect vector correctly represents the layer defects in the training data. Such an optimization method preferably comprises a gradient descent variant (e.g. stochastic gradient descent with or without momentum, Adam, RMSProp, etc.). The back-propagation algorithm is preferably used to efficiently compute the gradients. Alternatively, the stopping point can also be earlier if, for example, a previously defined training time budget is consumed, or if the model error evaluated on a separate validation dataset no longer decreases during training. The statistical learning model is then fully trained.
[0080] In addition to the above explanations, reference is made here by way of example to some publications which describe advantageous embodiments of convolutional neural networks and are hereby incorporated herein by reference, namely:
[0081] -Ronneberger, O.; Fischer, P.; Brox, T., "U-Net: Convolutional Networks for Biomedical Image Segmentation," Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015.
[0082] - Milletari F, Navab N, Ahmadi SA. V-net: “Fully convolutional neural networks for volumetric medical image segmentation”; 2016 Fourth International Conference on 3D Vision (3DV), October 25, 2016 (pp. 565-571), IEEE;
[0083] - Diba A, Fayyaz M, Sharma V, Karami AH, Arzani MM, Yousefzadeh R, Van Gool L., “Temporal 3d convnets: New architecture and transfer learning for video classification,” arXiv preprint arXiv:1711.08200, November 22, 2017;
[0084] Nabavi, Seyed and Rochan, Mrigank and Yang and Wang, (2018). “Future Semantic Segmentation with Convolutional LSTM”
[0085] - Xingjian SH, Chen Z, Wang H, Yeung DY, Wong WK, Woo WC. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems 2015 (pp. 802-810).
[0086] -Saeid Asgari Taghanaki, Kumar Abhishek, Joseph Paul Cohen, Julien Cohen-Adad, and Ghassan Hamarneh, “Deep Semantic Segmentation of Natural and Medical Images: A Review”; https: / / arxiv.org / pdf / 1910.07655.pdf.
[0087] Reference again Figure 3 In some exemplary embodiments, a defect vector is determined for each pixel of the height map and / or each pixel of the normalized image in step 98 in order to give a probability of the presence of a layer defect at the pixel level. In other exemplary embodiments, individual defect vectors may be determined for the material layer 18 to be inspected, or sub-regions may be defined within the material layer 18, and for each of these sub-regions, an individual defect vector is determined with the aid of a convolutional neural network.
[0088] In step 104, in some exemplary embodiments, the defect vector has a corresponding time stamp identifying the material layer currently being inspected. The defect vector with the time stamp may be stored in memory 106 as a historical defect vector (see Figure 1 ). In step 108, the corresponding current defect vector may be compared with the historical defect vector in the memory 106 to identify layer defects that persist across multiple material layers. Conversely, layer defects that are detected only in the current material layer and do not appear in subsequent material layers, for example, may also be discarded as “outliers”. In some exemplary embodiments, the defect vectors may be compared in step 108 as part of the determined comparison.
[0089] If already referenced Figure 4As indicated, in further exemplary embodiments, historical layer defect information may be provided to the statistical learning model as a set of height maps or input images 90a, 90b such that the resulting defect vector 100 already represents individual defect probabilities taking into account the persistence of any layer defect. In step 110, individual properties of the material layer 18 are determined based on the defect probabilities 102a, 102b of the defect vectors, in particular the presence or absence of a layer defect is defined. In some exemplary embodiments, in step 110, the extent and / or shape of the identified layer defect is determined by comparing the defect vectors / defect probabilities of different pixels or pixel regions. Advantageously, the inspected material surface is segmented into defect-free sub-regions and defective sub-regions and classified in step 110, wherein a distinction is made between the different defective sub-regions depending on the type of layer defect identified accordingly. Based on these properties, in Figure 2 In step 56, a decision is made whether and how to continue the method for manufacturing the workpiece.
[0090] Refer to the simplified diagram, Figure 5 Further exemplary embodiments of the novel method and exemplary embodiments for training a statistical learning model are shown. The same reference numerals denote the same elements as before. As already further indicated above, the statistical learning model 112' can be trained based on training data, which in particular comprises a group of normalized images of a material surface, each group having one or more of the layer defects to be detected. In the exemplary embodiment shown, the training data comprises a plurality of training images 120a, 120b and annotated versions 122a, 122b of the training images 120a, 120b. The layer defects to be detected are marked in the annotated versions 122a, 122b. Therefore, if the training images 120a, 120b are provided again as input data, the annotated versions 122a, 122b represent the target results that the statistical learning model 112' wants to deliver after the training is completed. In an iterative optimization process, the statistical learning model 112' (see Figure 4) are modified, for example, using stochastic gradient descent with or without momentum, Adam, RMSProp, etc., until the statistical learning model 112' delivers annotated versions 122a, 122b as part of the defined defect criteria and / or stopping criteria. The annotated versions 122a, 122b can advantageously be generated by an experienced expert with the aid of visual inspection of the training images 120a, 120b and individual marking of the identified layer defects. After the statistical learning model 112' has been sufficiently trained, in a preferred exemplary embodiment of the method for additive manufacturing of a workpiece, the statistical learning model is provided with the corresponding current image 92 and preferably also with historical images of previous material layers and / or current height maps and / or historical height maps. In a preferred exemplary embodiment, for checking each new material layer 16, 18, at least three, in particular four, current and preferably normalized images of the new material layer with mutually different illumination directions are used in each case ("normalized raw images", possibly rectified, but without height maps), while the annotated versions 122a, 122b can be produced by additionally using the height maps. In a further exemplary embodiment, the training images 120a, 120b can be images which were recorded during the layered manufacturing of a given workpiece according to the novel method and which therefore show a plurality of material layers produced during the manufacturing process. These training images can be used to retrain the statistical learning model and optimize it for the manufactured workpiece.
[0091] In the exemplary embodiment presented here, a statistical learning model 112 ' with an encoder / decoder architecture is used. This architecture can be advantageously used not only to determine a single defect vector 100. Instead, by means of a deconvolution step, a defect map 124 can be created, which represents the identified layer defects in a spatially resolved manner. For example, the defect map 124 can contain a single defect probability or a defect vector containing multiple individual defect probabilities for each pixel, wherein each defect probability represents whether the corresponding pixel shows a layer defect to be detected.
[0092] Figure 6A further statistical learning model 112" with an encoder / decoder architecture is shown. Here, four corresponding current images 92a, 92b, 92c, 92d of the material layers 16, 18 to be inspected, each image having a different illumination direction, are provided to the learning model 112" as input data. The images 92a, 92b, 92c, 92d may be normalized and corrected. Optionally, historical images and / or current and historical height maps may be additionally provided to the statistical learning model 112" as input data. The statistical learning model 112" determines a plurality of gradually filtered image stacks 94, 94' in a number of consecutive convolution steps, delinearization steps and pooling steps. The image stacks are then deconvolved again in an upsampling path. In the deconvolution step, the respective stacks are combined here with the respective intermediate results from the previous convolution steps, e.g. concatenated, as indicated by the arrow 126. Such a statistical learning model 112" is advantageously based on the above-described U-net model. The statistical learning model 112" thus delivers a defect map 124 containing spatially resolved defect probabilities that locally represent the presence or absence of layer defects. Here, the defect map 124 contains a plurality of defect vectors containing preferably a plurality of individual defect probabilities for each pixel of the images 92a, 92b, 92c, 92d. Such a defect map 124 can advantageously be used to determine the spatial dimension of the detected layer defects.
Claims
1. A method for additive manufacturing of a workpiece (14), the method comprising the following steps: a) receiving (50) a data set (36) defining the workpiece (14) as a plurality of successively arranged workpiece layers, b) producing (52) a material layer (18) having a defined surface formed by particulate material (20), c) recording (70, 74) at least one image of the material layer (18) and inspecting the material layer (18) based on the at least one image in order to determine respective properties of the material layer (18), d) selectively solidifying (58) the particulate material (20) on the defined surface by means of a building tool (28) using the data set (36) and based on individual properties of the material layer (18), wherein a defined workpiece layer (16) of the plurality of successively arranged workpiece layers is produced from the material layer (18), and e) repeating steps b) to d), wherein further defined workpiece layers of the plurality of successively arranged workpiece layers are generated, Characterized in that at least one image of the material layer (18) in step c) is examined using a previously trained statistical learning model (112), wherein the previously trained statistical learning model (112) is used to determine at least one defect vector (100) representing a plurality of individual defect probabilities (102a, 102b), wherein each individual defect probability of the plurality of individual defect probabilities (102a, 102b) is an individual indicator as to whether a defined layer defect among a plurality of possible layer defects is present in the material layer (18), and wherein individual properties of the material layer (18), namely morphological properties and / or dimensional properties of the defined defects in the material layer, are determined based on the defect probabilities (102a, 102b) of the defect vectors, and wherein step d) is performed based on the at least one defect vector (100).
2. The method according to claim 1, characterized in that In step c), the material layer (18) is irradiated from a plurality of mutually different directions (42a, 42c, 42e, 42f) and a plurality of images of the material layer (18) are recorded, wherein each of the plurality of images shows the material layer in a different irradiation direction (42a, 42c, 42e, 42f), and wherein the plurality of images are used to determine individual properties of the material layer (18).
3. The method according to claim 2, characterized in that The plurality of images (90a, 90b) are provided together as input data to the previously trained statistical learning model (112).
4. The method according to claim 1, characterized in that The at least one image is used in step c) to determine a height map of the material layer (18), wherein the height map is provided as an input data set to the previously trained statistical learning model (112).
5. The method according to any one of claims 1 to 4, characterized in that In step c), a plurality of defect vectors are determined, each of which represents a plurality of individual defect probabilities, wherein each defect vector of the plurality of defect vectors represents an individual defect probability associated with a selected pixel region in the at least one image.
6. The method according to claim 5, characterized in that Each defect vector of the plurality of defect vectors represents an individual defect probability associated with another individual pixel in the at least one image.
7. The method according to claim 5, characterized in that Morphological properties and / or dimensional properties of defined defects in the layer of material are determined using the plurality of defect vectors.
8. The method according to claim 1, characterized in that The previously trained statistical learning model includes a convolutional neural network (112).
9. The method according to claim 8, characterized in that The convolutional neural network has an encoder and decoder architecture.
10. The method according to claim 1, characterized in that At least one image of the material layer (18) is normalized with the aid of a reference image, wherein the reference image shows a uniform, diffusely reflecting surface or a defect-free material layer.
11. The method according to claim 1, characterized in that The at least one defect vector (100) is stored as a historical defect vector (104) along with a time stamp identifying the material layer (18), wherein the additional defined workpiece layers are generated based on the historical defect vector.
12. The method according to claim 11, characterized in that At least one image of the material layer (18) is stored as a historical image (80) together with a time stamp identifying the material layer, wherein the further defined workpiece layers are generated based on the historical image.
13. The method according to claim 12, characterized in that The individual properties of the material layers in the repeated step c) are each stored together with a time stamp identifying the respective material layer, wherein the use of the workpieces is authorized (66) based on the stored individual properties.
14. A device for additive manufacturing of a workpiece (14), the device having - a memory for receiving a data set (36) defining the workpiece (14) as a plurality of successively arranged workpiece layers, - production platform (12), - layer forming means (22), - Build Tools (28), - a camera (40) oriented towards the production platform, and - an evaluation and control unit (32) configured to generate a material layer (18) having a defined surface from a granular material (20) on the production platform (12) by means of the layer forming tool (22), further configured to record at least one image of the material layer (18) by means of the camera (40), and further configured to selectively solidify the granular material (20) on the defined surface by means of the building tool (28), wherein A defined workpiece layer (16) of the plurality of successively arranged workpiece layers is produced from the material layer (18), Characterized in that the evaluation and control unit (32) is also configured to use a previously trained statistical learning model (112) to inspect at least one image of the material layer (18), wherein the previously trained statistical learning model (112) is used to determine at least one defect vector (100) representing multiple individual defect probabilities (102a, 102b), wherein each individual defect probability of the multiple individual defect probabilities (102a, 102b) is a separate indicator as to whether a defined layer defect among multiple possible layer defects exists in the material layer (18), and wherein the evaluation and control unit (32) controls the layer forming tool (22) and the building tool (28) based on the defect vector (100).
15. A method for training a statistical learning model (112) for use in the method as claimed in claim 1 and / or in the device as claimed in claim 14, the method comprising the following steps: - obtaining a plurality of training images (120a, 120b) showing a plurality of material layers (18), wherein the plurality of material layers (18) contain a plurality of possible layer defects, - marking the plurality of possible layer defects in the plurality of training images (120a, 120b) so as to obtain a plurality of annotated versions (122a, 122b) of the training images (120a, 120b), - defining a stopping criterion based on a plurality of annotated versions (122a, 122b) of these training images (120a, 120b), - repeatedly providing the plurality of training images (120a, 120b) as input data to the statistical learning model (112), and repeatedly obtaining output data from the statistical learning model (112) in response to the plurality of training images (120a, 120b) provided accordingly, and - iteratively optimizing the statistical learning model (112) using the plurality of training images (120a, 120b) and the stopping criterion, - the plurality of training images (120a, 120b) are recorded during the manufacture of a definition workpiece (14) on the device as claimed in claim 14, and wherein the stop criterion is defined based on the definition workpiece (14).
16. The method of claim 15, wherein: Defect criteria and / or time criteria are defined based on multiple annotated versions (122a, 122b) of the training images (120a, 120b).
