DE102020121760A1 - Method and device for the additive manufacturing of a workpiece - Google Patents

Abstract

In a method for the additive manufacturing of a workpiece (14), a material layer (18) with a defined surface is formed from a particulate material (20). At least one image of the material layer (18) is recorded with a camera (40) in order to determine individual properties (110) of the material layer (18). Depending on the individual properties of the material layer (18), the particulate material (20) is selectively solidified using a structuring tool (28). The at least one image of the material layer (18) is inspected using a pre-trained statistical learning model (112). At least one error vector (100) is determined, which represents a large number of individual error probabilities (102a, 102b). Each individual error probability from the multiplicity of individual error probabilities (102a, 102b) is an individual indicator of whether a defined layer error from a multiplicity of possible layer errors is present in the material layer (18).

Description

1. a) Empfangen eines Datensatzes, der das Werkstück in einer Vielzahl von aufeinander angeordneten Werkstückschichten definiert,
2. b) Erzeugen einer Materialschicht mit einer definierten Oberfläche gebildet aus einem partikelförmigen Material,
3. c) Aufnehmen von zumindest einem Bild der Materialschicht und Inspizieren der Materialschicht anhand des zumindest einen Bildes, um individuelle Eigenschaften der Materialschicht zu bestimmen,
4. d) Selektives Verfestigen des partikelförmigen Materials an der definierten Oberfläche mithilfe eines Strukturierungswerkzeugs unter Verwendung des Datensatzes und in Abhängigkeit von den individuellen Eigenschaften der Materialschicht, wobei aus der Materialschicht eine definierte Werkstückschicht aus der Vielzahl von aufeinander angeordneten Werkstückschichten erzeugt wird, und
5. e) Wiederholen der Schritte b) bis d), wobei weitere definierte Werkstückschichten aus der Vielzahl von aufeinander angeordneten Werkstückschichten erzeugt werden.

The present invention relates to a method for the additive manufacturing of a workpiece, with the steps

1. a) receiving a data set that defines the workpiece in a plurality of workpiece layers arranged one on top of the other,
2. b) creating a material layer with a defined surface formed from a particulate material,
3. c) recording at least one image of the material layer and inspecting the material layer using the at least one image in order to determine individual properties of the material layer,
4. d) Selective solidification of the particulate material on the defined surface using a structuring tool using the data set and depending on the individual properties of the material layer, wherein a defined workpiece layer from the plurality of workpiece layers arranged one on top of the other is generated from the material layer, and
5. e) Repeating steps b) to d), further defined workpiece layers being produced from the plurality of workpiece layers arranged one on top of the other.

The invention also relates to a device for the additive manufacturing of a workpiece, with a memory for receiving a data set that defines the workpiece in a plurality of workpiece layers arranged one on top of the other, with a manufacturing platform, with a layer formation tool, with a structuring tool, with a camera that the production platform is directed, and with an evaluation and control unit that is set up to produce a material layer with a defined surface from a particulate material using the layer formation tool on the production platform, also taking at least one image of the material layer using the camera, and further to selectively solidify the particulate material on the defined surface using the structuring tool, a defined workpiece layer being produced from the plurality of workpiece layers arranged on top of one another from the material layer.

Such a method and such a device are basically over U.S. 2015/0061170 A1 famous.

Additive processes for the production of workpieces are sometimes referred to as 3D printing. There are various additive manufacturing processes. In selective laser sintering (SLS) or selective laser melting (SLM), a so-called powder bed made of a particulate material is used. Often the particulate material is a metallic material. However, there are also methods with particulate plastic materials, in particular polymers. Selected powder particles on the upper side of the powder bed are locally selectively melted or at least partially melted using a laser beam or electron beam and in this way selectively solidified during cooling. A new powder layer is then spread over the workpiece structure and the unmelted residual powder, and the workpiece is thus produced layer by layer. As a rule, the individual workpiece layers are produced from bottom to top on a production platform, which is lowered by the corresponding layer height after each workpiece layer.

The additive manufacturing of workpieces makes it possible to produce individual workpieces with a high degree of complexity and low material costs. At the same time, however, there are major challenges in terms of workpiece quality, since anomalies can occur in each individual layer of material, which can lead to defects in the workpiece. Anomalies can result in defects such as pores in the layer structure, micropores, local delamination/delamination, cracks inside and/or on the surface, dents, shape deviations and/or material stresses. For this reason, there are numerous proposals for detecting defects in an additively manufactured workpiece as early as possible during the manufacture of the layer sequence. The above US2015/0061170A1 discloses, among other things, an optical measurement sensor with a camera that can be set up to enable a 3D coordinate measurement on the topmost material layer in each case. A difficulty here are reflections and shadows that can be caused in particular by metallic powder particles, but by other objects in the working space of the device.

DE 10 2016 201 289 A1 discloses a method for the additive manufacturing of a workpiece, with first measurement data being recorded during the additive construction with a thermographic material test or with an eddy current material test. After the additive build-up, second measurement data is recorded using computed tomography and compared with the first measurement data. Material testing results are to be classified using an unspecified algorithm from the field of supervised machine learning.

EP 3 459 715 A1 discloses a method for the additive manufacturing of a workpiece, in which defects in a current layer or in subsequent further layers are to be predicted using a classification function that was trained with a technique from the field of machine learning that is not described in detail.

WO 2015/020939 A1 discloses a method based on the processing of filaments for the additive manufacturing of a workpiece, using algorithms from the field of machine learning to correlate a CAD input data set with parameters of the device used in order to determine the properties of the workpiece manufactured with the device and those for manufacturing predict required time. The actual quality control of the manufactured workpiece takes place after it has been manufactured using a 3D scanner and calibration patterns on the manufacturing platform.

DE 10 2017 108 874 A1 and the same priority U.S. 2020/158499 A1 disclose an optical system to enable material testing using illumination from multiple different directions. In some variants, the system can be used to determine a height map of a layer of material to be inspected.

Against this background, it is an object of the present invention to specify an alternative method and a corresponding device for the additive manufacturing of workpieces in high quality. One task in particular is to efficiently monitor the quality of the material layers close to the process in order to be able to correct any layer defects that occur or are imminent at an early stage.

According to a first aspect of the invention, this object is achieved by a method of the type mentioned at the outset, wherein the at least one image of the material layer is inspected in step c) using a pre-trained statistical learning model, wherein at least one error vector is determined using the pre-trained statistical learning model which represents a large number of individual error probabilities, each individual error probability from the large number of individual error probabilities being an individual indicator of whether a defined layer error from a number of possible layer errors is present in the material layer, and step d) being dependent from which at least one error vector is executed.

According to a further aspect, this object is achieved by a device of the type mentioned at the outset, the evaluation and control unit also being set up to inspect the at least one image of the material layer using a previously trained statistical learning model at least one error vector is determined, which represents a multiplicity of individual error probabilities, each individual error probability from the multiplicity of individual error probabilities being an individual indicator of whether a defined layer error from a multiplicity of possible layer errors is present in the material layer, and wherein the evaluations - and control unit controls the layer formation tool and the structuring tool depending on the error vector.

The statistical learning model of the new method and the new device comes from the field of machine learning. It implements a statistical evaluation of the at least one image of the material layer based on previously trained parameters and supplies individual probability values that each represent an individual probability for the presence of a defined layer defect from a large number of possible layer defects. In the preferred embodiments, the multitude of possible layer defects includes grooves in the surface of the material layer made of the particulate material, local accumulations of particulate material on the defined surface, non-uniform grain sizes of the particulate material, holes or depressions in the surface of the material layer made of the particulate material as well undesired adhesion or fusion of material particles. Such anomalies/inhomogeneities can lead to the workpiece defects mentioned above. By inspecting the material layer for the presence of such anomalies using the previously trained statistical learning model, the development of the defects can be detected at an early stage and, if necessary, avoided.

The use of the pre-trained statistical learning model efficiently enables a process-near inspection of the material layer both before and after the selective solidification of the particulate material. Inspection of the layer of material prior to selective solidification makes it possible to correct any anomalies prior to selective solidification, for example by smoothing the surface of the layer of material again, spreading additional particulate material, and/or replacing existing particulate material.

The statistical learning model provides individual error probabilities for the different which layer defects and allows a surface inspection based on experience, without each individual layer defect having to be exactly known in advance with regard to its exact appearance in at least one image of the material layer. In contrast to the methods mentioned at the beginning, an algorithm from the field of machine learning is used here specifically to inspect the powder bed with the material layer made of particulate material, not or not only to the behavior and properties of the device used for additive manufacturing and /or in relation to workpiece layers that have already been produced. The new process and the corresponding device make a very efficient contribution to avoiding defective workpieces and workpiece layers as early as possible. The error vector with the individual error probabilities for the various layer errors enables a qualitative and—at least in some exemplary embodiments—even a quantitative statement with regard to the quality properties of the workpiece produced in a very efficient manner.

In some embodiments, depending on the error vector, the selective solidification of the particulate material may be stopped or deferred until the material layer of the particulate material has a desired homogeneity through appropriate post-processing. In some exemplary embodiments, the production process can be terminated prematurely if a defect-free workpiece cannot be expected due to layer defects that are detected with a high degree of probability. Since different layer defects can occur at the same time or at different times, depending on a specific process sequence, locally distant from each other or in the immediate vicinity or to different extents, an inspection of the powder bed based on the machine learning method using the trained statistical learning model is particularly suitable. If the process parameters change, the static learning model can be “post-trained” in an efficient manner. The above task is therefore completely solved.

In a preferred embodiment of the invention, the material layer is illuminated in step c) from a plurality of different directions and a large number of images of the material layer are recorded, with each image from the large number of images showing the material layer with a different illumination direction, and with the individual Properties are determined using the plurality of images.

The configuration is particularly advantageous for inspecting a powder bed of metallic particulate material. However, it can equally be used to inspect a powder bed of plastic material or mixed materials. The multitude of images show the material layer with different light reflections and different shadows. Therefore, the individual properties of the material layer can be recorded more reliably and in more detail. It is advantageous if the multiplicity of images are recorded with a single camera which is arranged in a fixed position relative to the production platform and/or the material layer. This enables quick image acquisition and easy assignment of the various illumination images to one another. In particular, the configuration facilitates a semantic differentiation of different layer defects from one another, since it allows a detailed inspection of the material surface.

In a further refinement, the large number of images are supplied together as input data to the previously trained statistical learning model.

In this embodiment, each image from the plurality of images can form an inspection channel. The statistical learning model can process the image information from the different illumination images together and can therefore correlate the different image information. For example, when illuminated from a first direction, an edge in the material layer can be visible that is not visible when illuminated from a different direction. On the other hand, a reflection that occurs in one direction of illumination can resemble a non-existent film defect. The configuration advantageously contributes to recognizing as many layer defects as possible that are actually present and, moreover, to distinguishing between layer defects that are actually present and layer defects that are only apparently present. Established statistical learning models from the field of machine learning, some of which are described in more detail below, regularly enable the joint processing of several input data sets, such as the separate RGB color channels of a multicolored image. The embodiment takes advantage of this capability for correlated analysis of the plurality of multiple images with different directions of illumination. As a result, this design is relatively easy to implement and allows for fairly rapid inspection of the material surface.

In a further refinement, in step c) a height map of the material layer is determined using the at least one image, the height map being supplied to the trained statistical learning model as an input data record.

In preferred embodiments, the height map is determined using a plurality of images each showing the layer of material with a different direction of illumination. Preferably, the height map is determined as a 2.5D height map using a method as described in the aforementioned DE 10 2017 108 874 A1 and the same priority U.S. 2020/158499 A1 which is incorporated herein by reference. Alternatively or additionally, the height map could be determined in further exemplary embodiments according to the triangulation principle, for example according to the principle of strip light projection. The configuration has the advantage that the statistical learning model receives currently measured height information. As a result, the scope of the training data and the training time required in advance for the statistical learning model can be reduced. In addition, very critical layer defects, such as missing material or deep grooves as a result of a damaged layer formation tool, can be recognized very quickly using the height map.

In a further embodiment, in step c) a multiplicity of error vectors are determined, each of which represents a multiplicity of individual error probabilities, with each error vector from the multiplicity of error vectors representing the individual error probabilities in relation to a selected pixel region in the at least one image. In this configuration, the selected pixel areas are at least partially different from one another, so that the error vectors are representative for pixel areas of the at least one image that are different from one another.

It is particularly advantageous if each error vector from the large number of error vectors represents the individual error probabilities in relation to another individual pixel in the at least one image.

In these refinements, the error vectors indicate the probabilities for defined layer errors in locally delimited areas and in particular at the pixel level. The configurations make it possible to determine the extent and/or shape of any layer defects that may be present and/or their spatial progression. In particular, error vectors that relate to individual pixels in the at least one image enable the dimensions of detected layer errors to be determined very precisely, so that, for example, a pore size in the layer sequence of the workpiece can be estimated. The design contributes to a particularly efficient implementation of the new method and the corresponding device, since layer defects that do not exceed a specific size, configuration or shape can be specifically tolerated depending on the individual requirements of the workpiece produced. An advantageous classification of individual layer defects and workpiece defects that may result therefrom is also facilitated with this configuration.

In a further embodiment, morphological and/or dimensional properties of a defined defect in the material layer are determined using the multiplicity of defect vectors.

Morphological properties contain information about the structure and/or type of a defined defect. Dimensional properties include information about extent and/or shape. Determining the morphological and/or dimensional properties makes it easier to classify any layer defects and to make a targeted decision as to whether the workpiece produced can meet defined specifications in terms of strength, durability, shape and/or dimensions. The configuration is therefore particularly advantageous for an efficient process analysis.

In another embodiment, the pre-trained statistical learning model includes a convolutional neural network (CNU). The trained statistical learning model particularly preferably includes a convolutional neural network based on the U-net model.

A convolutional neural network uses the mathematical operation "convolution" to analyze an input data set. Filter matrices are convolved with the at least one image in sections in several steps. The result of this convolution is a data set from which the presence or absence of a feature represented by the filter can be estimated. A number of such convolution operations are preferably carried out sequentially one after the other. A separate filter matrix is advantageously used for each relevant layer defect. For example, a separate filter matrix can be provided for one or more of the following layer defects: Grooves in the surface of the material layer made of the particulate material, local accumulations of particulate material on the defined surface, non-uniform grain sizes of the particulate material, holes or depressions in the surface of the Material layer of the particulate material, bonding or fusion of material particles. The multiplicity of convolution operations lead together with further operations of the convolutional neural network, such as in particular unification operations (Pooling) and delinearization (using a Rectifying Linear Unit, ReLU) to the error vector. As has been shown on the basis of investigations, such a convolutional neural network enables an error vector which is representative of the layer errors mentioned to be determined in a very efficient manner.

A convolutional neural network based on the U-net model supplements the determination of the error vector in further subsequent convolution steps with information from the first convolution steps (so-called upscaling) and thus provides a very accurate segmentation of at least one image into different layer error areas and error-free image background. Such folding networks are traditionally used in the medical field. The investigations have shown that such networks are very well suited, especially when inspecting the material layer made of particulate material, to detect layer defects very early and accurately.

In a further configuration, the at least one image of the material layer is normalized using a reference image, with the reference image showing a homogeneous, diffusely reflecting surface.

This refinement increases the accuracy of the detection of any layer defects, since the previously trained statistical learning model receives one or more input images that are independent of the individual lighting. In this way, disadvantageous effects due to manufacturing tolerances in the lighting, signs of aging, different relative positions of individual lighting elements in the working space of the device relative to the camera, and other things can be reduced. In some embodiments, the homogeneous, diffusely reflecting surface of the reference image can be a white sheet of paper. The reflection properties of the diffusely reflecting surface preferably correspond to a Lambert radiator.

In a further refinement, the at least one error vector is stored as a historical error vector together with a time stamp which identifies the material layer, with the further defined workpiece layers being generated as a function of the historical error vector.

In some exemplary embodiments, it is therefore possible in the repeated steps c), i.e. when inspecting subsequent material layers, to check whether a layer defect detected in the current material layer is persistent over a number of layers. The design helps to reduce "false alarms" and to discontinue and/or modify the selective solidification of the particulate material of a current layer of material only when actually necessary to achieve a required part quality. For example, isolated layer defects that are limited to one material layer or apparent layer defects that are actually not present and appear as such due to light reflections, for example, can be eliminated in an efficient manner. In some embodiments, the current error vector can be compared with one or more historical error vectors in a deterministic manner after the inspection using the statistical learning model, i.e. as part of a post-processing. For example, a potential layer failure can be ignored if the historical failure vectors do not indicate the same layer failure of previous layers of material.

In a further embodiment, the at least one image of the material layer is stored as a historical image together with a time stamp that identifies the material layer, with the further defined workpiece layers being generated as a function of the historical image.

In this embodiment, which can be used as an alternative or in addition to the aforementioned embodiment, the history of detected layer defects is provided using the recorded images. The design makes it possible to supply one or more historical images together with a respective current image to the trained learning model as a common multi-channel input data record. The statistical learning model can then advantageously carry out the inspection of the material surface taking into account the history and recognize temporal correlations. In preferred embodiments, the at least one current image and one or more historical images can each form a channel of a convolutional neural network. In further exemplary embodiments, the statistical learning model can have a short-term memory, such as that implemented by an LSTM (Long Short Term Memory) network. Accordingly, in some embodiments, the statistical learning model may be an LSTM convolutional neural network.

This refinement can be particularly advantageously combined with the determination of a height map of the material layer, with a convolutional neural network being used, which performs the convolution operations three-dimensionally. For example, the first two dimensions can be the spatial pixel information along the x and y axes of the height map and the third dimension of the convolution operations can be time, using the current elevation map and one or more historical elevation maps. For example, the input data set can be a tensor whose dimensions correspond to the width and height of the height maps and the number of historical and current height maps. The configuration enables a very advantageous implementation of the new method and the corresponding device with monitoring of the manufacturing process for a large number of different workpieces and process sequences.

In a further embodiment, the individual properties of the material layers from the repeated steps c) are each stored together with a time stamp that identifies the respective material layers, the workpiece being released for use depending on the stored individual properties.

In this embodiment, the process-accompanying inspection of the material layers is advantageously used in order to assess the suitability of the manufactured workpiece for its intended use after the manufacturing process has ended. The stored individual properties can advantageously also be used to document a quality assurance process. The design contributes in a very efficient manner to achieving high product quality in an additive manufacturing process.

It goes without saying that the features mentioned above and those still to be explained below can be used not only in the combination specified in each case, but also in other combinations or on their own, without departing from the scope of the present invention.

• 1 eine schematische Darstellung eines Ausführungsbeispiels der neuen Vorrichtung,
• 2 ein Flussdiagramm zur Erläuterung eines Ausführungsbeispiels des neuen Verfahrens,
• 3 ein Flussdiagramm zur Erläuterung der Inspektion der Materialoberfläche gemäß Ausführungsbeispielen des Verfahrens aus 2, und
• 4 eine vereinfachte Darstellung zur Erläuterung der Funktionsweise eines faltenden neuronalen Netzwerks, das in Ausführungsbeispielen des neuen Verfahrens und der neuen Vorrichtung vorteilhaft zum Einsatz kommen kann.

Embodiments of the invention are shown in the drawing and are explained in more detail in the following description. Show it:

• 1 a schematic representation of an embodiment of the new device,
• 2 a flowchart to explain an embodiment of the new method,
• 3 a flowchart to explain the inspection of the material surface according to exemplary embodiments of the method 2 , and
• 4 a simplified representation to explain the functioning of a convolutional neural network, which can be advantageously used in exemplary embodiments of the new method and the new device.

In 1 an exemplary embodiment of the new device is denoted in its entirety by the reference number 10 . The device 10 has a manufacturing platform 12 on which a workpiece 14 is produced additively according to an exemplary embodiment of the new method. The workpiece 14 is built up in layers from bottom to top from a material stack in sequential steps. The reference number 16 indicates a currently uppermost workpiece contour or uppermost workpiece layer. An uppermost layer of material from which the workpiece layer 16 is produced is denoted by the reference number 18 .

After completion of a respective workpiece layer 16, a new material layer 18 made of a particulate material 20, such as a metallic material and/or a plastic material, is distributed on the layer stack with the aid of a doctor blade 22. For this purpose, the production platform 12 is typically lowered in the direction of the arrow 24 by the height of the next material layer and the particulate material 20 is removed from a reservoir 26 and distributed over the existing layer stack with the aid of the squeegee 22 .

A structuring tool is shown here in simplified form at reference number 28 . In some exemplary embodiments, the structuring tool 28 generates a laser beam 30 and moves it relative to the production platform 12 and the material layer 18 to be structured. The material particles are selectively melted and/or partially melted with the laser beam 30 so that they solidify as they cool.

In other embodiments, patterning tool 28 may generate an electron beam to pattern a workpiece layer on manufacturing platform 12 . In some embodiments, the apparatus 10 may include more than one patterning tool 28, such as using two or more laser and/or electron beams to create a workpiece layer.

The structuring tool 28, sometimes simply referred to below as a writing laser, is connected to an evaluation and control unit, hereinafter referred to as a controller 32 for short, which controls the movement of the laser beam 30 along the material surface. The controller 32 has an interface 34 here, via which a data set 36 can be read in, which defines the workpiece 14 to be produced in a multiplicity of layers arranged one on top of the other. The controller 32 controls the movement of the laser beam 30 relative to the material stack as a function of the data set 36, the laser beam 30 describing a trajectory in each workpiece layer 16 to be produced, which results from data record 36. In some exemplary embodiments, the controller 32 is implemented using one or more commercially available personal computers running an operating system, such as Microsoft Windows, MacOS or Linux, and one or more control programs with which exemplary embodiments of the new method are implemented. In some exemplary embodiments, the controller 32 can be implemented as a soft PLC on a commercially available PC. Alternatively or additionally, the controller 32 can be implemented using dedicated control hardware with one or more ASICs, FPGAs, microcontrollers, microprocessors or comparable logic circuits.

The device 10 also has a measuring arrangement 38, 40 which is set up to inspect the surface of the layer stack. The measuring arrangement here includes an illumination arrangement 38 and a camera 40, each of which is connected to the controller 32 (or to a separate controller for the measuring arrangement, not shown here). In this preferred exemplary embodiment, the camera 40 is set up to record a plurality of images of the surface of the material stack, with the surface being illuminated in different directions. For example, the lighting arrangement 38 here includes a multiplicity of lighting modules 38a - 38f which are arranged at different positions relative to the production platform 12 . Alternatively or additionally, the lighting arrangement 38 can be movable relative to the production platform 12 in order to illuminate the material surface from different directions. For example, the manufacturing platform 12 could be arranged on a turntable.

In 1 are three lighting modules 38a, 38b, 38c arranged side by side. The lighting modules 38a, 38b, 38c are thus able to generate three lighting directions 42a, 42b (not shown here), 42c that run largely parallel. Three further lighting modules 38d, 38e, 38f are arranged here transversely to the lighting modules 38a, 38b, 38c and parallel to one another on a second side of the production platform 12. The lighting modules 38d, 38e, 38f can generate three further lighting directions 42d (not shown here), 42e, 42f. In some exemplary embodiments, the device 10 has six further lighting modules (not shown here), of which three can be arranged opposite the lighting modules 38a-38c and another three can be arranged opposite the lighting modules 38d-38f. In exemplary embodiments, the device 10 is able to illuminate the material surface from at least three different main directions. In some exemplary embodiments, the main directions can each contain three partial lighting directions offset parallel to one another, as is shown in 1 is shown. This enables a very advantageous determination of a 2.5D height map of the material surface according to a method as in the initially mentioned DE 10 2017 108 874 A1 or the same priority U.S. 2020/158499 A1 is described. Both documents are incorporated herein by reference.

As an alternative or in addition to the in 1 In the arrangement shown in the illustration, device 10 could have a ring light with a large number of light sources that can be controlled individually and/or in segments, with the ring light (not shown) being arranged, for example, above production platform 12 and preferably around material stack 22 in order to illuminate material surface 18 allow from different directions.

In the following, with additional reference to the 2 until 4 Exemplary embodiments of the new method are explained using one or more control programs on the device in accordance with 1 can be implemented. According to step 50, a data record 36 is read into the controller 32, which defines the workpiece 14 in a multiplicity of workpiece layers 16 arranged one on top of the other. As an alternative or in addition to this, the controller 32 could first receive a data set via the interface 34 that defines the workpiece to be produced “as a whole”, such as a CAD data set, and based on this determine the multiplicity of workpiece layers 16 arranged one on top of the other. In this case, too, the controller 32 ultimately receives a data record which defines the workpiece 14 in a multiplicity of workpiece layers 16 arranged one on top of the other. According to step 52 a material layer 18 is produced on the layer stack 22 . For example, the controller 32 can remove particulate material 20 from the reservoir 26 with the aid of the squeegee 22 and distribute it on the layer stack. The distribution of the particulate material 20 should be as uniform and homogeneous as possible.

According to step 54, the surface of the material layer 18 is then inspected using the measuring device 38, 40 to detect any anomalies, such as in particular grooves, holes, depressions, waves, accumulations of material, density variations and/or particle inhomogeneities (e.g. different grain sizes and/or clumping ) in the material layer 18 can be seen. If the surface of the new material layer 18 corresponds to all desired criteria, the method branches according to step 56 to step 58, according to which a new workpiece layer 16 is produced in the uppermost material layer 18 using the writing laser 28 . The writing laser 28 melts material particles along the defined trajectory selectively and connects the melted or partially melted particles with each other in this way.

If the surface of the new layer of material 18 does not meet the desired criteria or does not do so sufficiently, the method can return to step 52 according to loop 60 in order to rework the surface of the layer of material 18 or to create it completely from scratch. According to step 62, steps 52-58 are repeated until the workpiece 14 corresponding to the data set 36 is completed.

In some exemplary embodiments, a freshly produced workpiece layer 16 can be inspected in a targeted manner using the measuring device 38, 40, which is indicated by reference number 64. Depending on this, a subsequent workpiece layer can then be modified in order to correct a deviation in shape or size, for example. According to step 66, the workpiece produced can be released for an intended use based on the history of the inspections from the repeated steps 52 and/or 64.

3 FIG. 12 shows an exemplary embodiment for the inspection of the material layer 18 according to step 54 and, if necessary, step 64 2 . The surface of the material layer 18 is preferably illuminated here with short-wave light. According to step 68, the material surface is illuminated here from a first direction. The first direction can, for example, be the illumination direction 42a according to FIG 1 be. According to step 70, the camera 40 is used to record a first image I of the illuminated material surface. Then, according to step 72, the material surface is illuminated from a further direction, for example from the illumination direction 42c according to FIG 1 . According to step 74, a further image J _k of the material surface is recorded with the aid of the camera 40, while the material surface is illuminated from the further direction 42c. According to step 76, a decision is made as to whether additional images of the material surface should be recorded with illumination from additional directions, for example from the illumination directions 42e, 42f according to FIG 1 . In some exemplary embodiments, the first illumination 68 and the further illuminations 72 can be generated simultaneously using different wavelengths/light colors and the different images can be separated from one another using the different wavelengths/light colors.

According to step 78, in some preferred exemplary embodiments, the recorded images are normalized using a reference image in order to obtain a uniform illumination level regardless of the arrangement of the illumination modules in the working space of the device 10 and any manufacturing tolerances. In some exemplary embodiments, the reference image can show a white sheet of paper that was captured by the camera 40 . Alternatively or additionally, in some exemplary embodiments, a 2.5D height map of the material surface 18 can be determined using the recorded and possibly normalized images, as has already been described above with reference to FIG DE 10 2017 108 874 A1 or the U.S. 2020/158499 A1 was mentioned. According to step 80, the elevation map and/or the normalized images are each provided with a time stamp that is representative of the material layer 18 currently recorded. In some embodiments, the timestamp may be a sequential number that is incremented with each new material layer 18, independent of the current time.

According to steps 82, 84, 86, 88, the height map and/or the normalized images from step 78 are supplied to a previously trained statistical learning model, which is preferably implemented here as a convolutional neural network. The basic functionality of such a network is 4 implied.

According to 4 one or more images 90a, 90b of the material surface 18 are fed to the statistical learning model as an input data set. In some embodiments, image 90a includes a current height map of the top layer of material 18 and image 90b includes the height map of a previous, underlying layer of material. Several such historical elevation maps of previous layers of material in the sequence of layers can be supplied to the statistical learning model together with the current elevation map. For example, height maps of two, three, four, or five previous layers of material may be included in the stack of input images 90a, 90b.

In other exemplary embodiments, the stack of input images 90a, 90b can contain current--preferably normalized--images of the uppermost material layer 18, each with a different direction of illumination. Accordingly, determination of a height map can be omitted in these exemplary embodiments. In further exemplary embodiments, the stack of input images 90a, 90b can contain current and historical images, each with different directions of illumination.

The input images 90a, 90b are now each convolved with one of a plurality of filter masks 92a, 92b. The convolution operations result in an image stack 94 in which, for each input image 90a, 90b and each filter mask 92, 92b, the respective convolution result is included. In the preferred embodiments, stack 94 also includes a step known to those skilled in the art as the ReLU step. With the ReLU step (step 84 according to 3 ) the stack 94 is deliberately de-linearized. Then takes place according to 3 a so-called pooling step, which can be implemented in particular as so-called max-pooling. With the pooling step 86, a stack 96 is generated in which redundant image information is reduced. Steps 82, 84, 86 can be repeated multiple times, as indicated by reference numeral 88 in 3 is indicated, with the stack from the previous steps serving as the input data set for the further convolution and pooling steps. In some embodiments, between three and ten convolution steps 82, ReLU steps 84, and pooling steps 86 can be cascaded one after the other.

With reference to 3 one or more error vectors 100 can be determined in a further step 98, in particular using the so-called softmax function, which supplies a probability distribution as a function of the previous stack 96'. Each error vector 100 therefore contains a multiplicity of individual error probabilities 102a, 102b, each individual error probability 102a, 102b being an individual indicator of whether a defined layer error from the multiplicity of layer errors mentioned above is present in the inspected material layer 18.

The possible layer errors correlate with the filter matrices 92a, 92b, which in turn can be determined in advance using the training data provided. In preferred exemplary embodiments, the training data contain height maps and/or normalized images of material surfaces that have one or more of the layer defects mentioned above and, in addition, at least one height map and/or images of a material surface that is free of defects. In order to generate such training data, a manufacturing process can be disrupted in a targeted manner, for example by overfilling with powder material, introducing grooves or waves in the material surface, mechanical shocks and others. In addition, images or height maps recorded in this way can be mirrored or manipulated using image processing methods in order to obtain a wide range of training data. The training data are supplied to the convolutional neural network 112 in advance in several training cycles. The resulting error vectors are analyzed for each training data set supplied. As part of a so-called backpropagation algorithm, the parameters of the filter masks 92a, 92b are modified until the resulting error vectors correctly represent the layer errors in the training data. Then the statistical learning model is sufficiently trained.

• - Ronneberger, O.; Fischer, P.; Brox, T., „U-Net: Convolutional Networks for Biomedical Image Segmentation“, Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015
• - Milletari F, Navab N, Ahmadi SA. V-net: „Fully convolutional neural networks for volumetric medical image segmentation“; In 2016 Fourth International Conference on 3D Vision (3DV) 2016 Oct 25 (pp. 565- 571). IEEE.
• - 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, 2017 Nov 22.
• - Nabavi, Seyed & Rochan, Mrigank & Yang, & Wang,. (2018). „Future Semantic Segmentation with Convolutional LSTM“
• - Xingjian SH, Chen Z, Wang H, Yeung DY, Wong WK, Woo WC. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. InAdvances in neural information processing systems 2015 (pp. 802-810).

In addition to the above statements, reference is made here to some publications which describe the implementation of an advantageous convolutional neural network and which are incorporated here by reference, viz

• - Ronneberger, O.; Fischer, P.; Brox, T., "U-Net: Convolutional Networks for Biomedical Image Segmentation," Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015
• - Milletari F, Navab N, Ahmadi SA. V-net: "Fully convolutional neural networks for volumetric medical image segmentation"; In 2016 Fourth International Conference on 3D Vision (3DV) 2016 Oct 25 (pp. 565-571). IEEE.
• - 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, 2017 Nov 22.
• - Nabavi, Seyed & Rochan, Mrigank & Yang, & Wang,. (2018). "Future Semantic Segmentation with Convolutional LSTM"
• - Xingjian SH, Chen Z, Wang H, Yeung DY, Wong WK, Woo WC. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. InAdvances in neural information processing systems 2015 (pp. 802-810).

Referring again to 3 the error vectors are determined according to step 98 in some embodiments for each pixel of the height map and/or for each pixel of the normalized images, resulting in the probability of the presence of layer errors at the pixel level. In other exemplary embodiments, a single error vector can be determined for the material layer 18 to be inspected, or partial areas can be defined within the material layer 18, for each of which an individual error vector is determined using the convolutional neural network.

According to step 104, in some exemplary embodiments, the error vectors are each provided with a time stamp that identifies the material layer currently being inspected. The error vectors provided with the time stamp can be stored as historical error vectors in a memory 106 (see 1 ) get saved. According to step 108, the respective current error vectors can be compared with the historical error vectors from the memory 106 in order to identify layer errors across several layers of material are persistent. Conversely, layer defects that are only detected in a current material layer and that do not appear in subsequent material layers can be discarded as "outliers". In some exemplary embodiments, the error vectors can be compared according to step 108 as part of a deterministic comparison.

As already noted 4 indicated, historical layer error information can be supplied to the statistical learning model in further exemplary embodiments as a stack of height maps or input images 90a, 90b, so that the resulting error vectors 100 already represent the individual error probabilities, taking into account the persistence of any layer errors. According to step 110, the individual properties of the material layer 18, in particular the presence or absence of defined layer defects, are determined using the defect probabilities 102a, 102b of the defect vectors. In some embodiments, in step 110 the extent and/or shape of the detected layer defects are determined by comparing the defect vectors/probabilities of defects for different pixels or pixel regions. The inspected material surface is advantageously segmented and classified in step 110 into defect-free sub-areas and defective sub-areas, with the defective sub-areas being differentiated from one another depending on the type of layer defect detected in each case. Based on these properties, step 56 is made 2 decided whether and how the process for manufacturing the workpiece is to be continued.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• US 2015/0061170 A1 [0003, 0005]
• DE 102016201289 A1 [0006]
• EP 3459715 A1 [0007]
• WO 2015/020939 A1 [0008]
• DE 102017108874 A1 [0009, 0022, 0048, 0055]
• US2020158499 A1 [0009, 0022, 0048, 0055]

Claims (13)

Method for the additive manufacturing of a workpiece (14), with the steps: a) receiving (50) a data set (36) which defines the workpiece (14) in a multiplicity of workpiece layers arranged one on top of the other, b) generating (52) a material layer ( 18) having a defined surface formed from a particulate material (20), c) taking (70, 74) at least one image of the material layer (18) and inspecting the material layer (18) using the at least one image in order to determine individual properties (110 ) of the material layer (18), d) selective solidification (58) of the particulate material (20) on the defined surface using a structuring tool (28) using the data set (36) and depending on the individual properties of the material layer (18 ), wherein a defined workpiece layer (16) is produced from the plurality of workpiece layers arranged one on top of the other from the material layer (18), and e) repeat len of steps b) to d), wherein further defined workpiece layers are generated from the plurality of workpiece layers arranged one on top of the other, characterized in that the at least one image of the material layer (18) in step c) using a previously trained statistical learning model (112) is inspected, with the aid of the previously trained statistical learning model (112) being used to determine at least one error vector (100) which represents a multiplicity of individual error probabilities (102a, 102b), each individual error probability from the multiplicity of individual error probabilities (102a, 102b) is an individual indicator of whether a defined layer defect from a large number of possible layer defects is present in the material layer (18), and step d) is carried out as a function of the at least one defect vector (100). procedure after claim 1 , characterized in that the material layer (18) is illuminated in step c) from a plurality of mutually different directions (42a, 42c, 42e, 42f) and that a plurality of images of the material layer (18) is recorded, each image from the plurality of images showing the layer of material with a different direction of illumination (42a, 42c, 42e, 42f), and wherein the individual properties are determined using the plurality of images. procedure after claim 2 , characterized in that the plurality of images (90a, 90b) are supplied together as input data to the previously trained statistical learning model (112). Procedure according to one of Claims 1 until 3 , characterized in that in step c) a height map of the material layer (18) is determined using the at least one image, the height map being supplied to the trained statistical learning model (112) as an input data set. Procedure according to one of Claims 1 until 4 , characterized in that in step c) a multiplicity of error vectors are determined, each representing a multiplicity of individual error probabilities, each error vector from the multiplicity of error vectors representing the individual error probabilities in relation to a selected pixel region in the at least one image. procedure after claim 5 , characterized in that each error vector from the plurality of error vectors represents the individual error probabilities with respect to another individual pixel in the at least one image. Procedure according to one of Claims 5 or 6 , characterized in that morphological and/or dimensional properties of a defined defect in the material layer are determined using the plurality of defect vectors. Procedure according to one of Claims 1 until 7 , characterized in that the previously trained statistical learning model includes a convolutional neural network (112), in particular a convolutional neural U-net network. Procedure according to one of Claims 1 until 8th , characterized in that the at least one image of the material layer (18) is normalized using a reference image, the reference image showing a homogeneous, diffusely reflecting surface. Procedure according to one of Claims 1 until 9 , characterized in that the at least one error vector (100) is stored together with a time stamp that identifies the material layer (18) as a historical error vector (104), the further defined workpiece layers being generated as a function of the historical error vector. Procedure according to one of Claims 1 until 10 , characterized in that the at least one image of the material layer (18) together with a time stamp identifying the material layer decorated, is stored as a historical image (80), the further defined workpiece layers being generated as a function of the historical image. Procedure according to one of Claims 1 until 11 , characterized in that the individual properties of the material layers from the repeated steps c) are each stored together with a time stamp identifying the respective material layers, the workpiece being released for use (66) depending on the stored individual properties. Device for the additive manufacturing of a workpiece (14), with - a memory for receiving a data set (36) which defines the workpiece (14) in a multiplicity of workpiece layers arranged one on top of the other, - a manufacturing platform (12), - a layer formation tool (22) , - a structuring tool (28), - a camera (40), which is aimed at the production platform, and - an evaluation and control unit (32), which is set up for a material layer (18) with a defined surface of a particulate to produce material (20) using the layer formation tool (22) on the production platform (12), also to record at least one image of the material layer (18) using the camera (40), and also to remove the particulate material (20) on the defined surface using the To solidify structuring tool (28) selectively, wherein from the material layer (18) a defined workpiece layer (16) from the plurality of successive ang assigned workpiece layers is generated, characterized in that the evaluation and control unit (32) is also set up to inspect the at least one image of the material layer (18) using a previously trained statistical learning model (112), wherein using the previously trained statistical learning model (112) at least one error vector (10) is determined, which represents a multiplicity of individual error probabilities (102a, 102b), each individual error probability from the multiplicity of individual error probabilities (102a, 102b) being an individual indicator of whether a defined Layer defects from a large number of possible layer defects are present in the material layer (18), and the evaluation and control unit (32) controls the layer formation tool (22) and the structuring tool (28) as a function of the error vector (100).

Images