DE102021106430A1 - Defect detection in additive manufacturing - Google Patents

Abstract

The present invention relates to an apparatus and a method for defect detection in additive manufacturing and relates in particular to defect detection during the additive manufacturing process. Disclosed is a method for preventing manufacturing defects during additive manufacturing of a manufactured object, comprising the steps of: providing a detection system in the additive manufacturing device for detecting at least one property of the production object, manufacturing the production object in several production steps, with the detection system determining the at least one property of the production object during production, comparing the at least one property with a target state of the property, determining deviations between the at least one Property and the target state of the property, evaluation of the deviation determined in the deviation determination step, if deviations are determined, determination of an adjustment of at least one process parameter, based on the deviation determined in the deviation determination step to reduce the deviation in the further production process, application of the adjustment determined in the adjustment determination step of the at least one process parameter.

Description

The present invention relates to an apparatus and a method for defect detection in additive manufacturing and relates in particular to defect detection during the additive manufacturing process.

Metallic additive manufacturing (3D printing) describes manufacturing processes for metallic components, in which material is applied layer by layer, thus creating three-dimensional objects. In one of these, selective laser melting (SLM) or also called laser powder bed fusion (LPBF), a metallic powder is melted in layers until it has the desired geometry. This manufacturing process makes it possible to produce complex geometries from one part, for which several components have to be produced and connected to one another in conventional manufacturing, e.g. in the CNC process. In addition, the process is cost-effective, particularly in the production of small series down to batch size 1. Another method for which defect detection can also be used is the so-called Direct Energy Deposition method, in which either a wire or powder is fused to build up complex geometries.

One challenge is the early detection of errors, the reproducibility and the certification of machines and components, which is particularly important in the aviation industry. State-of-the-art methods rely, for example, on computer tomography scans, i.e. the component that has already been produced is scanned using a computer tomograph and examined for errors. On the one hand, the method is expensive, since either the computed tomography scanner is required or time-consuming transport to a provider of computed tomography scans is necessary. Both cases make production more expensive. Other methods from the prior art rely on the optical control by cameras mounted in the printer with temperature measuring units mounted in the pressure chamber in order to monitor the temperature in the melting pool (melting pool). On the one hand, optical methods are error-prone because they depend on the operator, i.e. his experience and ability to concentrate. On the other hand, with this method, as with the CT scan, defects can only be detected ex post, after the manufacturing process has been completed. Metallic alloys used in aviation, for example, sometimes cost more than US$800/kg. If a defect is found after manufacture, the powder is lost, further increasing manufacturing costs depending on the defect rate. Error detection using temperature-based methods are error-prone based on heuristics and are not or only partially automated, i.e. a human decision-maker must ultimately recognize the error.

Therefore, a process is desirable that fully automatically detects and displays errors in-situ, i.e. during the production process, and, if necessary, corrects the process parameters in such a way that the print is either completed successfully or the print is completely or partially interrupted if a manufacturing defect is detected. Such a process can significantly reduce the printing costs and at the same time reduce the production time.

It is the object of the invention to eliminate the disadvantages of the prior art and to provide a device and a method that enable in-situ defect detection in additive manufacturing.

Relevant defect types in additive manufacturing are explained below.

Porosity and Density: Porosity is described by small indentations or holes that have formed in the component and thus reduce the material density in the section under consideration. The reduction in density causes the component to fatigue more quickly and eventually break. The holes are often almost spherical and smaller than 100 µm. Furthermore, the indentations are distributed almost equally over a given area. There are many reasons why porosity occurs, but they also involve gas bubbles trapped in the atomized powder, which are not displaced due to the laser’s insufficient radiation power and thus prevent the powder from fully fusing.

Incomplete Fusion Holes: Individual layers or hatches fuse incompletely and form differently shaped holes. The holes do not have a clear structure as with porosity.

Cracking, delamination and warping: Stress that arises in a component due to rapid heating and cooling can lead to defects such as cracking, ie there are breaks in the component, delamination, ie the manufactured layers separate from one another or warping ( Warpage), ie surface bend, lead. Cracking, delamination and warpage differ from the above defect types in that they spread over a large area, but can be induced by the presence of the above defects. By simulating the additive manufacturing process using finite element methods (FEM), the defects mentioned here can be at least partially predicted, this applies in particular to distortion and temperature propagation. Appropriate countermeasures such as strengthening the supporting structure (support) can thus be taken.

Surface finish (surface finishing): the printing process leaves a certain surface roughness on components, which largely depends on the material, the layer thickness and the printing parameters used. Sandblasting of components in post-processing is standard, but electrochemical post-processing is also used. Inadequate laser fusing can result in optically unsatisfactory surfaces that even post-processing steps cannot remove.

The problem is solved by the teaching of the independent claims.

Advantageous developments are the subject of the dependent claims.

In a first subject matter of the disclosure, a method for detecting manufacturing errors during the additive manufacturing of a manufacturing object is disclosed, which has the steps of providing a camera system on or in a construction space of an additive manufacturing device, determining a camera system path of the camera system during the manufacturing of the manufacturing object, manufacturing the manufacturing object in several production steps, with the camera system following the camera system path during production and recording and storing images of the production object at predetermined positions, determination of an image depth map for at least one sequence of at least two image recordings, determination of a total depth map from all image depth maps for one production step, determination of a 3- dimensional design of the production object from the total depth map.

It is also disclosed that the method also has the steps: comparison of the 3-dimensional configuration with a 3-dimensional target configuration, determination of deviations between the 3-dimensional configuration and the 3-dimensional target configuration, evaluation of the deviation determination step Deviation, if deviations are determined, output of a deviation assessment.

Furthermore, it is disclosed that the camera system is a stereo camera system, which takes two pictures of the production object at the same time, stores them and determines the determination of the local depth map from the two pictures taken at the same time.

Furthermore, it is disclosed that the image depth map determination step, the overall depth map determination step, the 3-dimensional configuration determination step, the adjustment step, the deviation determination step and/or the evaluation step take place after the production of the production object has been completed.

It is also disclosed that the local depth map determination step, the global depth map determination step, the 3-dimensional design determination step, the adjustment step, the deviation determination step and/or the evaluation step take place after the completion of one of the manufacturing steps of the production of the manufacturing object.

Furthermore, it is disclosed that the global depth map of several manufacturing steps is taken into account in the 3-dimensional design determination step.

Furthermore, it is disclosed that the number of manufacturing steps taken into account is n>20 and n<100.

It is also disclosed that in the local depth map determination step, the local depth map is determined by transforming a first recorded image in such a way that it corresponds to a second recorded image and during the transformation the local depth map by a depth map determination algorithm from the class of the monitored or self-monitored machine -Learning procedure is generated.

It is also disclosed that the depth map determination algorithm is carried out by training an artificial neural network using at least one known first recorded image, one known second recorded image and a known local depth map.

It is also disclosed that the speed at which the camera system moves is taken into account during the transformation of the first recorded image into the second recorded image.

It is also disclosed that the position, the change in position, orientation or the change in orientation of the camera system is recorded and the position, the change in position, orientation or the change in orientation of the camera system is recorded as an input variable during the transformation of the first image recording into the second image recording and the depth map is determined therefrom becomes.

Furthermore, it is disclosed that the camera position from the path information, the relative ver displacement of the image recordings, by position sensor information or by information from the rotary encoders of the servomotors of the camera system.

Furthermore, it is disclosed that in the 3-dimensional design determination step, the 3-dimensional design is carried out by using algorithms such as Powercrust, Marching cube, Poisson Surface Reconstruction or Hoppes Reconstruction.

Furthermore, it is disclosed that the camera system can capture light in the infrared range, in particular in the near infrared range, and that it also has the step of capturing a temperature map of the production object during a production step.

Furthermore, it is disclosed that a device for additive manufacturing is set up to carry out a method according to one of the preceding claims.

In a second subject matter of the disclosure, a method for detecting manufacturing errors during additive manufacturing of a manufacturing object is disclosed, which has the steps: providing a detection laser system on or in a construction space of the additive manufacturing device manufacturing the production object in several manufacturing steps by means of a manufacturing laser in the construction space of the additive manufacturing device, Use of a laser in a construction space of an additive manufacturing device to determine properties of a manufacturing object.

Furthermore, it is disclosed that the determined properties are a 3-dimensional configuration and/or a surface temperature.

Furthermore, it is disclosed that the method also has the steps: comparison of the 3-dimensional configuration with a 3-dimensional target configuration, determination of deviations between the 3-dimensional configuration and the 3-dimensional target configuration,

Furthermore, it is disclosed that the method further comprises the steps: evaluation of the deviation determined in the deviation determination step, if a deviation is determined, outputting a deviation evaluation.

Furthermore, it is disclosed that the 3-dimensional configuration takes place by means of pulsed distance measurement, AMCW distance measurement or FMCW distance measurement.

Furthermore, it is disclosed that the detection laser is the manufacturing laser of the additive manufacturing device.

Furthermore, it is disclosed that the detection laser is a laser separate from the manufacturing laser of the additive manufacturing device.

Furthermore, it is disclosed that the method further comprises the steps: providing a partially reflecting mirror which is arranged such that a beam of the laser passes through from the transparent side and a beam reflected from the production object is reflected by the partially reflecting mirror, providing a fixed mirror , wherein the surfaces of the partially reflecting mirror and the fixed mirror are designed in such a way that a beam reflected from the production object is reflected onto a one-dimensional, preferably circular area, and providing at least one sensor on the one-dimensional area that is set up to detect the reflected beam capture.

Furthermore, it is disclosed that the surfaces of the partially reflecting mirror and the fixed mirror are configured quasi-parabolic.

It is also disclosed that the method also has the steps: providing a reflector rotating about a first axis, providing a fixed mirror which is designed in the shape of a segment of a circle perpendicular to the first axis and is arranged in such a way that the rotating reflector is located in the center of the segment of a circle and a has such a surface that a beam traveling radially to the circle segment is reflected parallel to the first axis and is focused onto a building layer, providing a laser receiving device, rotating the rotating reflector about the first axis, directing a laser beam onto the rotating reflector, reflecting the laser beam by the rotating Reflector in the direction of the fixed mirror, Reflecting the laser beam by the fixed mirror in the direction of the production object, Reflecting the laser beam by the production object in the direction of the fixed mirror, Reflecting the reflected laser beam by the fixed mirror in the direction the rotating reflector, reflecting the reflected laser beam by the rotating reflector to a laser receiver, detecting the reflected laser beam by the laser receiving device.

Furthermore, a device for additive manufacturing is disclosed, which is set up to carry out a method according to the preceding method claims.

Furthermore, it is disclosed that the device further comprises: a laser receiving device, the is set up to receive a laser beam reflected by a workpiece, a calculation unit set up to determine a transit time of the laser from its emission from alignment, intensity, frequency, modulation or reception time and thus to determine the dimensions of the workpiece.

Furthermore, it is disclosed that the device also has: a laser emitter, which is arranged and set up to emit a laser beam, by means of which, in cooperation with the laser receiving device, a runtime of the laser can be determined from its emission and thus the dimensions of the workpiece can be determined.

Furthermore, it is disclosed that the device has: an arrangement of optically effective elements, which is designed in such a way that the laser beam strikes the building layer perpendicularly in the detection area.

It is also disclosed that the device has: a reflector that can be rotated about a first axis, a fixed mirror that is designed in the shape of a segment of a circle perpendicular to the first axis and is arranged in such a way that the rotatable reflector is located in the center of the segment of a circle and has a surface such that a beam running radially to the circle segment is reflected parallel to the first axis and is focused on a building layer.

In a third subject matter of the disclosure, a method for preventing manufacturing defects during additive manufacturing of a manufacturing object is disclosed, which has the steps: providing a detection system in the additive manufacturing device for detecting at least one property of the manufacturing object, manufacturing the manufacturing object in several manufacturing steps, wherein during manufacturing the Recording system that determines at least one property of the production object, comparison of the at least one property with a target state of the property, determination of deviations between the at least one property and the target state of the property, evaluation of the deviation determined in the deviation determination step if deviations are determined, determination of an adjustment of at least one process parameter based on the deviation determined in the deviation determination step to reduce the deviation in the further production process , application of the adjustment determined in the adjustment determination step of the at least one process parameter.

Furthermore, it is disclosed that the at least one process parameter is one or more from the group of installation space temperature, production laser intensity or production speed.

Furthermore, it is disclosed that the detection system is a camera system and/or a projector camera system and/or a laser detection system for detecting a three-dimensional configuration of the production object or for detecting a surface temperature of the production object.

Furthermore, it is disclosed that the adjustment is determined in the adjustment determination step by means of a model-free algorithm.

Furthermore, it is disclosed that the adjustment is determined in the adjustment determination step by means of an artificial neural network for modeling transition probabilities or optimal action sequences. An action is understood as the adjustment of a process parameter as a result of a change in a state variable, for example the temperature in the melt pool.

Furthermore, it is disclosed that the adjustment is determined in the adjustment determination step by means of a model-based algorithm.

It is also disclosed that the adjustment is determined in the adjustment determination step using an artificial neural network for modeling optimized sequences of changes in process parameters based on system modeling for transferring the additive manufacturing device from an actual state to a target state.

Furthermore, it is revealed that for dynamic randomization and domain randomization (Dynamics / Domain Randomization) environmental parameters are varied during a system simulation to force the artificial neural network to higher generalization.

A device for additive manufacturing is disclosed, which is set up to carry out a method according to the preceding method claims.

In a fourth subject matter of the disclosure, a method for detecting manufacturing errors during the additive manufacturing of a manufacturing object is disclosed, which has the steps: providing a camera system on or in a construction space of an additive manufacturing device, determining a camera system path of the camera system during the manufacturing of the manufacturing object in several manufacturing steps, wherein the camera system during production camera system path and takes and saves images of the production object at predetermined positions, determining an image depth map for at least one sequence of at least two images, determining a total depth map from all image depth maps for a production step, determining a 3-dimensional design of the production object from the total depth map, using a laser for scanning a build layer in the build space of an additive manufacturing device to determine properties of a manufactured object.

Furthermore, it is disclosed that the determined properties are a 3-dimensional configuration and/or a surface temperature.

It is also disclosed that the method also has the steps: comparison of the 3-dimensional configuration with a 3-dimensional target configuration, determination of deviations between the 3-dimensional configuration and the 3-dimensional target configuration, evaluation of the deviation determination step Deviation, if deviations are determined, output of a deviation assessment.

It is also disclosed that the method also has the steps: comparison of the 3-dimensional configuration with a 3-dimensional target configuration, determination of deviations between the 3-dimensional configuration and the 3-dimensional target configuration, evaluation of the deviation determination step Deviation, if deviations are determined, output of a deviation evaluation.

Furthermore, it is disclosed that the camera system is a stereo camera system, which takes two pictures of the production object at the same time, stores them and determines the local depth map from the two pictures taken at the same time.

Furthermore, it is disclosed that the determined properties are a 3-dimensional configuration and/or a surface temperature.

It is also disclosed that the method also has the step: calibrating the laser using an image recorded by the camera system, generating a laser disparity map (disparity map) from the data determined with the laser for the same image section as the image recorded by the camera system , generating an image disparity map from the data determined with the camera system, generating a total disparity map from the laser disparity map and the image disparity map. Generate a 3-dimensional design of the manufacturing object based on the overall disparity map.

Furthermore, it is disclosed that the method further comprises the steps: providing a partially reflecting mirror which is arranged such that a beam of the laser passes through from the transparent side and a beam reflected from the production object is reflected by the partially reflecting mirror, providing a fixed mirror , wherein the surfaces of the partially reflecting mirror and the fixed mirror are designed in such a way that a beam reflected from the production object is reflected onto a one-dimensional, preferably circular area, and providing at least one sensor on the one-dimensional area that is set up to detect the reflected beam capture.

Furthermore, it is disclosed that the surfaces of the partially reflecting mirror and the fixed mirror are configured quasi-parabolic.

It is also disclosed that the method also has the steps: providing a reflector rotating about a first axis, providing a fixed mirror which is designed in the shape of a segment of a circle perpendicular to the first axis and is arranged in such a way that the rotating reflector is located in the center of the segment of a circle and a has such a surface that a beam traveling radially to the circle segment is reflected parallel to the first axis and is focused onto a building layer, providing a laser receiving device, rotating the rotating reflector about the first axis, directing a laser beam onto the rotating reflector, reflecting the laser beam by the rotating Reflector toward the fixed mirror, reflecting the laser beam by the fixed mirror toward the manufacturing object, reflecting the laser beam as a reflected laser beam through the manufacturing object toward the fixed mirror, reflecting the reflected laser beam through the fixed mirror towards the rotating reflector, reflecting the reflected laser beam by the rotating reflector to a laser receiver, detecting the reflected laser beam by the laser receiving device.

Furthermore, it is disclosed that the image depth map determination step, the overall depth map determination step, the 3-dimensional configuration determination step, the adjustment step, the deviation determination step and/or the evaluation step take place after the production of the production object has been completed.

Furthermore, it is disclosed that the local depth map determination step, the global depth map determination step, the 3-dimensional design The processing determination step, the adjustment step, the deviation determination step and/or the evaluation step take place after the completion of one of the production steps of the production of the production object.

Furthermore, it is disclosed that the global depth map of several manufacturing steps is taken into account in the 3-dimensional design determination step.

Furthermore, it is disclosed that the number of manufacturing steps taken into account is n>20 and n<100.

Furthermore, it is disclosed that the detection laser is the manufacturing laser of the additive manufacturing device.

Furthermore, it is disclosed that the detection laser is a laser separate from the manufacturing laser of the additive manufacturing apparatus.

Furthermore, a device for additive manufacturing is disclosed, which is set up to carry out a method according to one of the preceding paragraphs.

• 1 zeigt eine Darstellung eines ersten Ausführungsbeispiels einer erfindungsgemäßen Vorrichtung zum Aufbau eines dreidimensionalen Fertigungsobjekts.
• 2 zeigt einen Bauraum der Vorrichtung aus 1.
• 3 zeigt eine Sensoranordnung des ersten Ausführungsbeispiels.
• 4 zeigt eine Darstellung eines zweiten Ausführungsbeispiels der erfindungsgemäßen Vorrichtung zum Aufbau eines dreidimensionalen Fertigungsobjekts.
• 5 zeigt eine Darstellung eines dritten Ausführungsbeispiels der erfindungsgemäßen Vorrichtung zum Aufbau eines dreidimensionalen Fertigungsobjekts.
• 6 zeigt eine Darstellung zur stereoskopischen Tiefenermittlung.
• 7 zeigt eine Punktewolke und ein zugehöriges Dreiecksnetz (Oberflächendarstellung).
• 8 zeigt ein Ablaufdiagramm eines erfindungsgemäßen Verfahrens zur 3D-Rekonstruktion.
• 9 zeigt eine Darstellung einer ersten Abwandlung der erfindungsgemäßen Vorrichtungen zum Aufbau eines dreidimensionalen Fertigungsobjekts gemäß den Ausführungsbeispielen.
• 10 zeigt eine Darstellung einer zweiten Abwandlung der erfindungsgemäßen Vorrichtungen zum Aufbau eines dreidimensionalen Fertigungsobjekts gemäß den Ausführungsbeispielen.
• 11 zeigt eine Schnittdarstellung der zweiten Abwandlung der erfindungsgemäßen Vorrichtungen zum Aufbau eines dreidimensionalen Fertigungsobjekts gemäß den Ausführungsbeispielen.

Exemplary embodiments are described below with reference to the figures.

• 1 shows a representation of a first exemplary embodiment of a device according to the invention for constructing a three-dimensional production object.
• 2 shows an installation space of the device 1 .
• 3 shows a sensor arrangement of the first embodiment.
• 4 shows a representation of a second exemplary embodiment of the device according to the invention for constructing a three-dimensional production object.
• 5 shows a representation of a third exemplary embodiment of the device according to the invention for constructing a three-dimensional production object.
• 6 shows a representation of the stereoscopic depth determination.
• 7 shows a point cloud and an associated triangle mesh (surface representation).
• 8th shows a flow chart of a method according to the invention for 3D reconstruction.
• 9 shows a representation of a first modification of the devices according to the invention for constructing a three-dimensional production object according to the exemplary embodiments.
• 10 shows a representation of a second modification of the devices according to the invention for constructing a three-dimensional production object according to the exemplary embodiments.
• 11 shows a sectional view of the second modification of the devices according to the invention for constructing a three-dimensional production object according to the exemplary embodiments.

FIRST EMBODIMENT

The following is based on 1 a basic structure of a first exemplary embodiment of a device for generating a three-dimensional object (also referred to below as generating device, laser sintering device, LSM or printer) by layer-by-layer solidification of a building material, which is designed as a laser sintering device according to one embodiment, is described. In the case of the generating device 1, layers of the construction material are applied successively on top of one another and the locations in the respective layers corresponding to the object to be manufactured are each selectively solidified before the application of a subsequent layer. In the illustrated embodiment, a powdered build material is used which is solidified at selected locations by exposure to an energy beam. In the illustrated embodiment, the powdery building material is heated locally at the selected points by means of a laser beam, so that it is connected to neighboring components of the building material by sintering or melting.

As in 1 1, the generating device 1 has an optical system, the components of which are each attached to the components of the machine frame. A construction space 10 is provided in the machine frame.

In the illustrated embodiment, the optics system includes a laser 6, a deflection mirror 7 and a scanner 8. The laser 6 generates a beam 9 which hits the deflection mirror 7 and is deflected by it in the direction of the scanner 8. The scanner 8 is designed in a known manner in such a way that it can direct the incoming beam 9 to any points in a construction layer (current layer) 11 that is located in the construction space 10 . In order to make this possible, an entry window 12 is provided between the scanner 8 and the construction space 10 in an upper partition wall 5 of the construction space 10 , which allows the beam 9 to pass through into the construction space 10 . An F-θ lens 14 is in the path of the beam 9 and directs the beam 9 perpendicular to the building layer 11.

Next, with reference to 2 the installation space of the device is described in the embodiment.

A container 25 open at the top is provided in the installation space 10 . A carrier device 26 for carrying a three-dimensional object to be formed is arranged in the container 25 . The carrier device 26 can be moved back and forth in the vertical direction in the container 25 by means of a drive (not shown). A build plate 13 is provided at the upper end of the carrier device 26 and can be moved vertically with the carrier device 26 . The three-dimensional object is arranged on the building board 13 during its manufacture. The current construction layer 11 (current layer), ie the layer which is currently being manufactured, is defined in the region of the upper edge of the container 25 . The entrance window 12 for the beam 9 directed through the scanner 8 onto the building layer 11 is arranged above the building layer 11 . A coater 27 (recoater) is provided for applying building material to be solidified to the surface of the carrier device 26 or to a previously solidified layer. The coater 27 is indicated by arrows in 2 schematically indicated drive in the horizontal direction over the building layer 11 movable. Metering devices 28 are provided on both sides of the building layer 11, which provide a predetermined quantity of the building material for the coater 27 to apply.

A supply opening 30 is provided on the side of the dosing device 28 . The supply port 30 extends in the direction perpendicular to the plane of FIG 5 over the entire width of the construction layer 11. The feed opening serves to feed construction material, which in the illustrated embodiment is a powder material that can be solidified by radiation, into the construction space.

As in 2 is shown schematically, the installation space in the embodiment is divided into an upper area 40 and a lower area 41 . The upper area 40 forms the actual working area, in which the building material is applied in layers and selectively solidified. The layers extend in the x-direction and y-direction in a horizontal plane. The lower portion 41 receives the container 25 on.

In the embodiment shown, some components are formed by a method of fabricating a three-dimensional element layer by layer by selectively solidifying locations corresponding to the object in the respective layers. In the embodiment, a laser sintering method is used to manufacture them.

When the production device 1 is in operation, the construction material is fed into the construction space 10 via the feed opening 30 and fed to the coater 27 in a predetermined quantity by the metering devices 28 , 29 . The coater 27 applies a layer of the building material to the carrier device 26 or a previously solidified layer, and the laser 6 and the scanner 8 direct the beam 9 to selected positions in the building layer 11 in order to selectively solidify the building material at those locations , corresponding to the three-dimensional object to be formed. The carrier device is then lowered by the thickness of one layer, a new layer is applied and the process is repeated until all layers of the object to be formed are produced.

Based on 3 a sensor arrangement according to the first exemplary embodiment is explained. A camera system 50 is provided in the upper area 40 of the installation space. A first camera module 52 and a second camera module 53 are provided on a carriage 51 . The first camera module 52 and the second camera module 53 are aligned essentially downwards and are arranged non-rotatably relative to one another at a fixed, known distance d. The carriage 51 is attached to a bracket 56 in a rail system. The carriage 51 can be displaced along the horizontal x-axis via a third servomotor 57 . The carrier 56 can be displaced along the horizontal y-axis via rails 58 and servomotors 59 . Due to this arrangement, the carriage 51 and thus the first camera module 52 and the second camera module 53 can be moved freely in the plane spanned by the x and y axes and can be aligned with any point on the workpiece.

A LiDAR system is also implemented by the device for generating a three-dimensional object. In this case, a laser emitter is provided and a measuring laser beam is generated by the laser emitter, which is directed by a deflection device onto the object to be detected—in the present case the construction layer 11 . The deflection device is suitable for directing the laser beam to each point of the building layer 11 whose depth is to be measured. The laser beam is reflected from the surface of the object to be measured, ie from the surface of the construction layer 11 in this case. A laser receiver is also provided, which detects the reflection of the laser beam. From the signal propagation time, ie the time difference between sending and receiving the laser beam and the propagation speed of the laser beam in the corresponding environment medium in and/or outside of the construction space, the distance for each measuring point is determined with knowledge of the arrangement of the laser emitter and the laser receiver and with knowledge of the deflection of the deflection device.

In the present exemplary embodiment, the production laser 6 is used as a measuring laser, with the deflection mirror 7 and the scanner 8 serving as a deflection device. The laser receiver is arranged (outside) on or in the combustion chamber 10 and aligned with the building layer 11 .

SECOND EMBODIMENT

Based on 4 a second embodiment of the invention will be explained. Apart from the differences described below, the second exemplary embodiment is identical to the first exemplary embodiment.

Instead of the camera system 50, a camera system 50' is provided on the carriage 51. The camera system has a camera module 52'. The camera module 52′ can be rotated about the vertical z-axis via a first servomotor 54 and its alignment in the vertical plane can thus be changed. The camera module 52' can be rotated about a horizontal axis, i.e. an axis perpendicular to the z-axis, via a second servomotor 55, and the orientation can thus be changed downwards. With this arrangement, the carriage 51 and thus the camera module 52' can be moved freely in the plane spanned by the x and y axes and the alignment of the first camera module 52' can be aligned with any point on the workpiece.

THIRD EMBODIMENTS

Based on 5 a third embodiment of the invention will be explained. Apart from the differences described below, the third exemplary embodiment is identical to the first exemplary embodiment.

A circular bearing ring 61 is provided in the upper area 40 of the installation space 11 and is connected to the housing in a stationary manner. Within the bearing ring 61, a rotary ring 62 is rotatably mounted. The rotary ring 62 can be freely rotated within the bearing ring 61 via the servomotor 63 . A first camera module 52" is attached to the rotating ring 62. The camera module 52" is oriented radially inwards. In this case, the camera module 52" is inclined downwards in the direction of the construction layer 11 by the angle of inclination α. The angle of inclination α is preselected and fixed. Due to the angle of inclination of the camera module and the rotatability of the rotary ring 62, every area of the construction layer 11 can be recorded by the camera module Further camera modules, for example a second camera module 53", can be attached to the rotary ring. The second camera module 53" is aligned radially inwards and is inclined downwards by an angle of inclination β in the direction of the building layer 11. The angle of inclination β is preselected, fixed and different from the angle of inclination α. Depending on the choice of the angle of inclination, this arrangement allows stereoscopic detection in one pass, faster stereoscopic acquisition with one camera module at a time through parallel acquisition by several cameras or a larger area coverage.

In a modification of this exemplary embodiment, the angles of inclination α and β can each be changed by a servomotor. The area covered by the camera module 52" and the camera module 53" can thus be changed. This allows camera modules with a longer focal length to be used, which increases the resolution but reduces the area captured. Changing the angles of inclination α and β nevertheless ensures that the entire area of the building layer 11 remains detectable.

DEFECT DETECTION THROUGH 3D RECONSTRUCTION IN THE EXEMPLARY EMBODIMENTS:

A first method for in-situ defect detection defines a defect as a deviation (additive or reductive) between the surface of the printed component and the CAD model that goes beyond a specified tolerance. For this purpose, the component is continuously reconstructed in three dimensions during printing and compared with the model, typically in a CAD format such as STL or STEP, in order to detect deviations.

RECONSTRUCTION FROM POINT CLOUDS AND RGB-D IMAGES:

One method for calculating the distance is the so-called LiDAR system, which emits a laser beam and thus (directly or indirectly) measures the time it takes for a laser beam to be reflected back to the system after it has been emitted. It is then possible to determine how far away an object is via transit time and the speed of light.

Stereo cameras are systems consisting of two or more cameras that simultaneously produce two or more than two images of the same scene and determine the depth from them. Alternatively, two images of the same scene are recorded with a movably arranged camera at different times and in different locations, and the depth is determined from them. Since both systems have already determined the depth dimensions, the reconstruction problem is reduced to the segmentation of geometric objects, ie determining which depth point belongs to the object of interest and which depth point belongs to the surface of the unprocessed powder bed and can be ignored.

One option for segmentation is thermal segmentation. Thermal segmentation assumes that a thermogram exists that includes the objects as well as at least part of the surrounding powder bed. In the exemplary embodiments, the camera modules are set up to detect thermal radiation. Based on the thermogram, information about the powder used and its corresponding thermal behavior, the temperature gradient (dT) can be used to determine whether it belongs to the powder bed or to the solid object. The rule is then dT > S, where S is a lower bound that is determined experimentally.

As an alternative to the temperature gradient, the probability (T) can be determined for each pixel as to whether the material covered by this pixel has melted or not. For this purpose, the known physical properties of the material and corresponding simulation methods (finite elements or similar) are used. Then, when the probability is calculated, an appropriate bound (P) is chosen and if T > P then this pixel belongs to the solid object, else to the powder bed. It should be noted here that incomplete fusion holes would also have a low probability. Therefore, in a last step, the surfaces marked as not melted have to be found. The areas with T < P and an area smaller than X square pixels are ignored and therefore belong to the lasered area, X is to be selected appropriately.

In the case of thermal segmentation, it is advantageous to transform the coordinates found into the image coordinates of the depth maps, point cloud and/or the RGB image. Inaccuracies can be avoided if the resolution of the thermal image is sufficiently high relative to the optical image and ideally the resolution of the thermal image corresponds to that of the optical image. This can be achieved, for example, by a sufficiently high resolution of the thermal image acquisition, multiple measurements or measurements with several but offset cameras.

Another option for segmentation is optical segmentation.

The RGB images of the camera modules are used for the optical segmentation. The powder bed and the molten material stand out clearly from each other due to the physical transformation process for most powders. Based on this observation, the production object can be recognized with the help of histogram-based methods. That is, at the point where the gray value of a pixel changes significantly compared to its surrounding pixel, there is an edge. Alternatively, gradient-based methods, for example using the Sobel operator or clustering-based methods, which are known to those skilled in the art, can be used. As before, the hull must be calculated. The use of clustering methods is particularly advantageous here, since on the one hand it is known that there should only be two clusters and on the other hand the CAD model can be used as a template.

Another option for segmentation is geometric segmentation. Here, the segmentation takes place via the depth map or the point cloud.

The determination is made via the depth map gradient (dT), i.e. the difference in depth between neighboring pixels. If the depth gradient dT > S, a threshold determined experimentally and roughly equal to a factor of the density ratio between the raw powder and the fused powder, and the layer thickness of a powder layer, then an edge is assumed to exist here. The shell is determined on the basis of the determined points. Otherwise, the same procedure as in thermal segmentation is followed.

In addition to the methods already described, the segmentation can also be carried out/learned using unsupervised or self-supervised methods. Furthermore, the CAD can be used as a prior to further improve the process or to achieve more precise results.

A point network (mesh) must then be calculated based on the point cloud in order to be able to compare the geometry with a CAD file, e.g. in STL format.

One problem with the use of LiDAR is that presumably not all types of defects can be detected, e.g. those areas that have not been lasered or have been insufficiently lasered (incomplete fusion). Under certain circumstances, stereo cameras can compensate for this, since they can also map textures in addition to depth. The working hypothesis in this case is that there are differences in texture between a sound and a sound faulty section in the component. The challenge with stereo cameras is their accuracy, which decreases with the square of the distance to be measured. The further away the workpiece is from the camera, a larger and larger area is covered by a camera pixel.

The invention proposes solutions for the use of a LiDAR and for the use of stereoscopy in defect detection in additive manufacturing. Furthermore, it is proposed to merge the information from LiDAR and stereoscopy in order to achieve the highest possible accuracy in the reconstruction. The sensors represent the data in different structures, stereo cameras provide two images in RGB space, whereas lidar sensors generate point clouds containing the distance to the object and the orientation relative to the sensor. Therefore, it is necessary to use a multi-stage procedure, which in the first step finds a uniform representation of the input data between the different sensor modalities and then calculates the 3D representation of the object from it.

The following is based on 8th the procedure for the 3D reconstruction is described.

At the beginning of the method for reconstructing the 3D shape of the object, in a depth estimation step S10 the accurate depth is estimated, i.e. it is determined how far the object imaged in a pixel of the camera is from the camera. A camera system 50 is installed in the build chamber 10 for this purpose. The camera system 50 is preferably a stereo camera 50. According to the first exemplary embodiment, the camera system can move freely in the x-y direction, i.e. in the horizontal plane, in a translatory manner. A translational movement in the z-direction or the height direction is not provided in the first exemplary embodiment, since a movement in the height direction removes the camera from the working area and the error in the depth estimation of a stereo camera increases quadratically with the distance to the object. Therefore, the camera system 40 should be placed as close as possible (optically or physically) to the printed object. According to the second exemplary embodiment, the camera system 50 can be rotated about the x-axis and the y-axis.

Alternatively, multiple cameras can be attached to increase the scanning speed, e.g. in a row, so that the aggregated field of view covers a larger part of the installation space in either the x-direction or y-direction.

LOCAL DEPTH ESTIMATION:

The depth estimation method begins in a path determination step S20 with the determination of an optimal path for the camera system 50 through the upper installation space 40. The path of the camera system in the x-y plane and, if necessary, the alignment around the z-axis and the horizontal axis definitely. The path depends on various factors such as the progress in the manufacturing process, lighting conditions and the speed required to complete the scan in time.

If the path is determined, the scanning process begins in a scanning step S30. Here, the camera system 50 follows the path determined in the path determination step S20 and takes pictures of the workpiece at predetermined positions.

The image depth map itself is determined in several stages and begins in an image depth map determination step S40 with the calculation of an individual depth map for each image pair. Here, an image A, which was taken from perspective A, is transformed in such a way that it looks like an image B, which was taken from a different perspective than image A. In the transformation step, the depth map is generated by a depth map determination algorithm, which is an algorithm, for example, from the class of self-monitored machine learning methods. Learning takes place by transforming a known image into another known image. The images of a stereo image pair can be used for this, i.e. the algorithm learns to transform the right image into the left image and vice versa. Furthermore, consecutive images of a video sequence can also be used since the scene to be reconstructed is static, i.e. none of the objects to be reconstructed are in motion. In the exemplary embodiment, both are combined and a temporal sequence of stereo pairs (e.g. 3 images) is used for the learning. The transformation error, i.e. the discrepancy between the transformed image and the correct image, is used to improve the algorithm during learning. Several neural networks are used here, which learn the depth estimation. Alternatively, the depth map can be determined using trigonometric algorithms.

DETAILED DESCRIPTION OF DEPTH MAP CALCULATION

The epipolar geometry describes with sufficient accuracy how the depth (z-axis) of the respective imaged areas can be determined from a pair of stereo images. The central challenge associated with this is the korres to find ponding pixels in the stereo image pair (correspondence problem), ie to find out which pixel from the first image depicts the same area of the imaged object as a certain pixel from the second image. With the well-known trigonometric formulas, the depth can be estimated from the disparity, ie the pixel shift between the left and right image.

In order to solve the correspondence problem and thus the depth estimation, the invention teaches:

The aim is to learn a function / that associates a depth map D _t with each image I _t of the stereo image pair, ie D _t = ƒ(I _t ). Another advantage of the approach is that a mono camera can be used instead of a stereo camera. In the application, the estimated depth maps for the left and right images in the stereo image pair are averaged: $$D_t=\frac{D_t^l+D_t^{\mathrm{right}}}{2},$$

The function / is estimated by minimizing the photometric reprojection error. The concept behind this is to reconstruct the other image from the depth map estimated for one image of the stereo pair and to minimize the error that arises in the reconstruction. The disparity is learned implicitly, because in order to learn one image from the other of a stereo pair, the algorithm has to learn how to move the individual pixels in the image plane in order to reconstruct the other image.

$$L_{SSIM}^t\left(I_t,\widehat{I_t}\right)=\alpha \frac{1-SSIM\left(I_t,\widehat{I_t}\right)}{2},$$

$$L_{norm}^t\left(I_t,\widehat{I_t}\right)=\left(1-\alpha \right){\Vert I_t-\widehat{I_t}\Vert }_1,$$

den Rekonstruktionsfehler zwischen dem Originalbild I_t und dem aus der Tiefenkarte rekonstruierten Bild $\widehat{I_t}.$

L_SSIM enthält den strukturellen Ähnlichkeitsindex (SSIM), der die wahrgenommene Ähnlichkeit zwischen zwei Bildern modelliert. Der Wertebereich der Funktion liegt bei [-1,1], wobei 1 erreicht wird, wenn beide Bilder exakt gleich sind. L_norm beschreibt den absoluten Pixelfehler zwischen den beiden Bildern.The optimization criterion consists of several parts. To do this, L _p calculates: $$L_p=\underset{t}{min}\left[L_{SSIM}^t+L_{nO\mathrm{right}m}^t\right],$$

$$L_{SSIM}^t\left(I_t,\widehat{I_t}\right)=a\frac{1-SSIM\left(I_t,\widehat{I_t}\right)}{2},$$

$$L_{nO\mathrm{right}m}^t\left(I_t,\widehat{I_t}\right)=\left(1-a\right){\Vert I_t-\widehat{I_t}\Vert }_1,$$

the reconstruction error between the original image I _t and the image reconstructed from the depth map $\widehat{I_t}.$

L _SSIM contains the Structural Similarity Index (SSIM), which models the perceived similarity between two images. The range of the function is [-1,1], where 1 is reached when both images are exactly the same. L _norm describes the absolute pixel error between the two images.

besteht der Kontextvektor aus $$I_{t\text{'}}=\left\{\left(I_{t-1}^l,I_{t-1}^r\right),I_t^r,\left(I_{t+1}^l,I_{t+1}^r\right)\right\}.$$

Furthermore, the images are indexed with t, since the optimization criterion is still averaged over a context. Part of the context is the image of a stereo pair for which no depth map estimate was determined, as well as the stereo pairs before and after the target image. That means for estimation based on picture $I_t^l$

the context vector consists of $$I_{t\text{'}}=\left\{\left(I_{t-1}^l,I_{t-1}^{\mathrm{right}}\right),I_t^{\mathrm{right}},\left(I_{t+1}^l,I_{t+1}^{\mathrm{right}}\right)\right\}.$$

$$d_t^{*}=\frac{d_t}{d_t^{\varnothing }},$$

sorgt dafür das die Übergänge zwischen den Tiefenschätzungen zweier benachbarter Pixel glatt sind. Damit lassen sich abrupte Änderungen in der Disparität (Gradient) zwischen zwei benachbarten Pixeln, die vermutlich Artefakte sind, vermeiden. Die dahinterstehende Annahme ist, dass es nur wenige Ausnahmefälle gibt, in denen ein großer Gradient zu rechtfertigen ist. Zu bemerken ist, dass d_t die inverse Tiefe ist und $d_t^{\varnothing }$

die durchschnittliche inverse Tiefe, womit $d_t^{*}$

die Mittelwert-Normalisierte inverse Tiefe ist.The second component L _s of the optimization criterion: $$L_s={\Vert {\partial }_x{\mathrm{i.e}}_t^{*}\Vert }_2{e}^{-{\Vert {\partial }_x{I}_t\Vert }_2}+{\Vert {\partial }_y{\mathrm{i.e}}_t^{*}\Vert }_2{e}^{-{\Vert {\partial }_y{I}_t\Vert }_2},$$

$${\mathrm{i.e}}_t^{*}=\frac{{\mathrm{i.e}}_t}{{\mathrm{i.e}}_t^{\varnothing }},$$

ensures that the transitions between the depth estimates of two neighboring pixels are smooth. This avoids abrupt changes in the disparity (gradient) between two neighboring pixels, which are probably artifacts. The underlying assumption is that there are only a few exceptional cases where a large gradient can be justified. Note that d _t is the inverse depth and ${\mathrm{i.e}}_t^{\varnothing }$

the average inverse depth, with what ${\mathrm{i.e}}_t^{*}$

is the mean-normalized inverse depth.

The estimation of depth maps based only on the described optimization criteria leads to dimensionless maps. In order to be able to carry out the error detection by means of a comparison between the reconstructed component and the CAD template, it is necessary to represent the depth in a unit of measure, e.g. mm. As previously described, an essential part of the low estimation is the relative configuration of the camera between two images of a stereo pair already given by the calibration process. It is taught how to determine the relative translational change in the camera configuration between two consecutive RGB images (I _s ,I _t ) by also measuring the speed at which the camera moves between two images and using it as a learning signal: $$L_v=\left|{\Vert t_{s\to t}\Vert }_2-\left|v\right|\Delta T_{s\to t}\right|.$$

The measured speed v is multiplied by the time ΔT _s→t that has elapsed between the two images (I _s ,I _t ) in order to determine the actual translatory change. The optimization criterion is then calculated as the difference between the distance actually covered and the estimated distance ||t _s→t || ₂ where t represents the translational component (Δx, Δy) of the estimated configuration change.

Since a stereo camera is mounted in the proposed setup and not only translational changes can be measured by the encoders on the motors that move the camera, but also rotational changes between two consecutive stereo pairs (S _s ,S _t ) can be measured, the following is taught: $$L_{pOse}^s={\Vert L_v^l+L_v^{\mathrm{right}}\Vert }_2+{\Vert L_{\mathrm{right}Ot}\Vert }_2.$$

sicher, dass beide die translatorische Veränderung zwischen den zeitlich aufeinanderfolgenden Bildern beider Stereopaare $\left(I_s^l,I_t^l\right)$

und $\left(I_s^r,I_t^r\right)$

gleich sind, da die beiden Objektive der Stereokamera sich mit gleicher Geschwindigkeit translatorisch bewegen müssen. Die rotatorische Geschwindigkeit kann dabei unterschiedlich sein für beide Objektive. Weiterhin erfasst L_rot den zeitlichen Fehler der Rotation zwischen den gleichen Bildern eines Stereopaars $\left(I_s^l,I_t^l\right)$

und $\left(I_s^r,I_t^r\right)$

sowie dem räumlichen Fehler in der Rotation zwischen den Bildern eines Stereopaars (I^l,I^r) . Die Rotationsachse ist dabei durch die Montage der Stereokamera auf dem beweglichen Element gegeben. Mit dem formulierten Optimierungskriterium wird gleichzeitig auch die Maßeinheit für die Tiefe induziert.In doing so ${\Vert L_v^l+L_v^{\mathrm{right}}\Vert }_2$

sure that both the translational change between the temporally consecutive images of both stereo pairs $\left(I_s^l,I_t^l\right)$

and $\left(I_s^{\mathrm{right}},I_t^{\mathrm{right}}\right)$

are the same, since the two lenses of the stereo camera must move translationally at the same speed. The rotational speed can be different for both lenses. Furthermore, L _rot captures the temporal error of the rotation between the same images of a stereo pair $\left(I_s^l,I_t^l\right)$

and $\left(I_s^{\mathrm{right}},I_t^{\mathrm{right}}\right)$

and the spatial error in the rotation between the images of a stereo pair (I ^l ,I ^r ). The axis of rotation is given by mounting the stereo camera on the movable element. With the formulated optimization criterion, the unit of measurement for the depth is also induced at the same time.

This results in the full optimization criterion as: $$L={\mathrm{<figure-callout\; id="1"\; label="\lambda "\; filenames="DE102021106430A1\_0043.png"\; state="\{\{state\}\}">\lambda </figure-callout>}}_1\mu L_p+{\lambda }_2{L}_s+{\lambda }_3{L}_{pOse}^s.$$

$$\mu =\left[\underset{t\text{'}}{min}rpe\left(I_t,\widehat{I_{t\text{'}}}\right)<\underset{t\text{'}}{min}rpe\left(I_t,I_{t\text{'}}\right)\right],$$

Die Binäritat der zweiten Gleichung entsteht durch die Verwendung der Iverson Klammern [·].The factor µ E {0,1} is a binary mask that masks stationary pixels. The method described is based on the assumption that the scene for which a depth map is estimated is moving, either because the camera is actively moving or because a majority of the objects in the scene are moving. If the pixel values do not change between two consecutive images, the performance drops sharply. Therefore, it is proposed to filter stationary pixels in the optimization criterion through a binary mask, ie $$\mathrm{right}pe\left(I_t,\widehat{I_t}\right)=L_{SSIM}^t\left(I_t,\widehat{I_t}\right)+L_{nO\mathrm{right}m}^t\left(I_t,\widehat{I_t}\right).$$

$$\mu =\left[\underset{t\text{'}}{min}\mathrm{right}pe\left(I_t,\widehat{I_{t\text{'}}}\right)<\underset{t\text{'}}{min}\mathrm{right}pe\left(I_t,I_{t\text{'}}\right)\right],$$

The binary nature of the second equation arises from the use of the Iverson brackets [·].

zu berechnen.There are two distinct networks, the depth map D _t and the reconstruction $\widehat{I_t}$

to calculate.

In order to estimate the depth map, the so-called PackNet architecture is used, which consists of several convolution layers for the encoder as well as the decoder. Each layer in the encoder is followed by a packing operation that compresses the learned features and an unpacking operation that decompresses the features to produce the depth map. The depth map of the final slice is then used to reconstruct images of the stereo pair omitted from the depth map or the same stereo image but with a time difference of [-1,1].

Furthermore, it is necessary to determine the relative change in the camera configuration in order to enable the reconstruction of an image from a depth map. For this purpose, the relative change in configuration of the target image I _t relative to the context is estimated. To estimate the configuration change, the network architecture encoder is taken from V. Guizilini, R. Ambrus, S. Pillai, A. Raventos, and A.Gaidon, "3DPacking for Self-Supervised Monocular Depth Estimation," arXiv:1905.02693 [cs] , Mar. 2020 and this was extended by a recurrent layer and a 1×1 convolution layer, which reduces the dimension of the output to 6, i.e. 3 translational dimensions (Δx, Δy, Δz) and 3 rotational dimensions (ϕx, ϕy, ϕz). The recurrent layer makes it possible to use larger time horizons for estimating the change in configuration in order to get more stable estimates.

A distinction must be made between left-right reconstruction and temporal reconstruction. In the left-right reconstruction, the relative configuration change between the left and right object of the stereo camera is known from the camera calibration. Using the intrinsic camera parameters and the depth map, the correspondence between a pixel position in the left image and the right image can be established by translation and rotation. The pixel values are then taken from the original image for the reconstruction using bilinear sampling kernels (sampling): $${\widehat{I}}_i^c={\displaystyle \sum _{n=0}^H{\displaystyle \sum _{m=0}^W{I}_{n.m}^c}}max\left(0.1-\left|x_i^s-m\right|\right)max\left(0.1-\left|y_i^s-n\right|\right).$$

The index t has been omitted for the sake of clarity.

ergeben sich aus der Rücktransformation der Pixelkoordinaten $\left(x_t^s,y_t^s\right)$

im Zielbild Î mit den intrinsischen Kameraparametern und der entsprechenden Transformationsmatrix A_θ: $$\left(\begin{array}{c}{x}_i^s\\ y_i^s\end{array}\right)=A_{\theta }\left(\begin{array}{c}{x}_t^s\\ y_t^s\end{array}\right).$$

The pixel coordinates $\left(x_i^s,y_i^s\right)$

result from the inverse transformation of the pixel coordinates $\left(x_t^s,y_t^s\right)$

in the target image Î with the intrinsic camera parameters and the corresponding transformation matrix A _θ : $$\left(\begin{array}{c}{x}_i^s\\ y_i^s\end{array}\right)=A_{\theta }\left(\begin{array}{c}{x}_t^s\\ y_t^s\end{array}\right).$$

In order to increase the accuracy of the method for estimating the depth map, the resolution of the depth maps that result in the intermediate layers of the decoder can be increased by a nearest-neighbor interpolation to the resolution at the decoder output. The depth maps scaled in this way are additionally reconstructed according to the method described above and taken into account in the optimization criterion.

The optimization criterion is averaged over all reconstructions, pixels and the batch and can be trained using a standard method such as a gradient descent method. A distinction must be made between the initialization phase and the transfer phase.

In the initialization phase, it is assumed that the model's coefficients are estimated for the first time based on a training set that has been previously collected. The model is then ready for the first application. Each time the model is applied, additional data is collected, which in turn can be used to improve the model. This happens in the transfer phase, in which the already trained model is improved with new data. It should be noted that a model change is only accepted if it performs better on a test data set (see below) than the original model.

To test the accuracy of the model, it is suggested that a set of different molds be made and placed in a fully calibrated rig. The measuring stand is then driven off.

The method described generates a sequence of depth maps for a scan S=D ₀ ,...,D _t ,...,D _T . In order to generate a complete point cloud from the individual scans, it is necessary to locate the measured depths in an absolute xyz reference system.

The spatial resolution of the depth map can be obtained with reference to 6 be improved as follows:

wobei d die Disparität x_l - x_r ist. Man berücksichtige ein passendes Paar der Disparität d die einer Tiefe Z entspricht. Wird ein Fehler von einem Pixel gemacht (d+1 anstelle d), wird die Tiefe zu Z'. Um den Tiefenfehler ΔZ auszuwerten, soll dieser über den Disparitätsfehler dargestellt werden. Unter Heranziehung der Ableitung von Z als Funktion von d, erhält man: $$\left|\frac{\Delta Z}{\Delta d}\right|=\frac{Bf}{d^2}=\frac{Z^2}{Bf}.$$

With B as baseline/lens distance, Z as depth, / as focal length and x _l ,x _r as distances between the projected point and the optical axis, the equation for depth Z can be found as follows: $$Z=\frac{Bf}{\mathrm{i.e}},$$

where d is the disparity x _l - x _r . Consider a matching pair of disparity d corresponding to a depth z. If an error of one pixel is made (d+1 instead of d), the depth becomes Z'. In order to evaluate the depth error ΔZ, this should be represented using the disparity error. Taking the derivative of Z as a function of d, we get: $$\left|\frac{\Delta Z}{\Delta \mathrm{i.e}}\right|=\frac{Bf}{{\mathrm{i.e}}^2}=\frac{Z^2}{Bf}.$$

For a disparity of one pixel, the error increases quadratically with the depth. If subpixel disparity is assumed, Δd should be considered, ie $$\frac{Z^2}{Bf}\Delta \mathrm{i.e}.$$

Resolution improves as baseline/objective distance increases. However, care should be taken when choosing a large baseline. This is because matching becomes more difficult as the baseline increases (at fixed depth) due to increased distortion/deviation between left and right image. This resembles the usual detection/location trade-off.

Resolution improves with focal length. This follows from the fact that, for a given image size, the pixel density of the image plane also increases with increasing focal length. Therefore, the disparity resolution is higher. Although this stems from a simplified stereo model, the three principles apply to all stereo systems.

The depth map determined in this way is initially dimensionless. In order to introduce the units of measurement necessary for later determination of the object geometry, the speed at which a camera moves is taken into account during the transformation. In other words, in addition to the image recordings, the movement speed of the camera system 40 is an input parameter for the depth map determination algorithm. This allows the depth map determination algorithm to learn the correct unit of measure. In addition, not only the translational speed of the camera system 40 is taken into account, but preferably also the rotational speed of the changes in orientation of the camera system 40. On the one hand, the depth estimation for an individual image can be improved, on the other hand, the accuracy of the subsequent combination of all local ones Depth maps to a global depth map for the entire construction space.

DETERMINATION OF UNCERTAINTY:

The accuracy of the depth estimation is significantly influenced by two types of errors:

The accuracy of the estimate decreases quadratically with the distance to the object (systematic error). This is because a single pixel covers a larger and larger spatial area in the world the further away that area is from the camera. Since the depth is only ever estimated for a single pixel, this means that depth changes that exist within a pixel are only averaged.

Imperfections in the model used for depth estimation affect accuracy.

The first error can be reduced by reducing the distance between the camera and the object, increasing the focal length of the lens, increasing the viewing distance between the stereo camera lenses, and reducing the distance traveled by the camera between two stereo pairs to reduce the overlap maximize. The overlap makes it possible to minimize the error in the global depth map generation step.

On the one hand, the error introduced by the inadequacy of the model can be reduced with more data and a better model. In addition, the model error can be included in later calculations by assigning the corresponding uncertainty of the model with regard to the estimation to each estimated value for the depth. The Bayesian interpretation of probability is used here. Furthermore, the path planning with which the depth estimation started can be adjusted in such a way that areas with high uncertainty are approached again.

GLOBAL DEPTH ESTIMATION:

In the image depth map determination step S40, a multiplicity of overlapping depth maps are created, which are converted into a uniform depth map for the entire construction space in an overall depth map determination step S50. For this purpose, a camera pose (position and orientation of the camera) is estimated for each individual image depth map in order to determine and globally optimize the relative relationship between the determined depth points. The aim is to align the points of the individual depth maps with each other in such a way that the most plausible global depth map is created. In the exemplary embodiment, a number of methods are preferably combined in order to achieve increased accuracy. An estimate of the camera pose for each individual depth map has already been determined in the image depth map determination step S40. However, this estimate may still be inaccurate. To improve this, the camera is localized on the one hand based on the information from the rotary encoders of the motors that change the camera pose, from the inertial measurement unit (IMU) attached to the camera or an interior localization technology (e.g. RFID, magnetometer, etc.). Furthermore, the camera can be localized based on the image information from the image recordings. Individual features in the overlapping areas of the images are calculated (SIFT features). The relative poses of the images to one another can thus be determined and merged with the relative poses from the image depth map determination. The correspondence between the individual features in the image recordings is then filtered in order to reduce erroneous assignments. The depth maps are then divided into segments, each with n consecutive images, and for each of these segments the camera pose aligns all images with one another. There is an overlap of one image between the segments, i.e. the last and the first image of two consecutive segments are always the same. All segments are then aligned with one another by calculating segment-specific features and determining and optimizing their correspondence with other segments. There are thus two camera pose determinations - one based on the sensors and one based on the image information - which are merged into a global pose in a final step in order to then align the image depth maps accordingly and merge them into a single overall depth map.

In a configuration determination step S60, a three-dimensional configuration of the workpiece is determined. Here, the total depth map is converted into a point cloud by choosing a suitable coordinate system and calculating the x-y-z coordinates for each individual depth point. Furthermore, the individual points are enriched with secondary information that is necessary for further calculations, e.g. uncertainty and surface normals. If necessary, the point cloud is built up layer by layer. At the end of the print, a point cloud is obtained that includes all printed objects and all defects.

There are several possibilities at which point in time a surface reconstruction makes sense. These are briefly described below.

Generation of the entire point cloud and subsequent meshing, i.e. until the print is finished. Then all printed objects and defects can be calculated and extracted.

SURFACE MESHING AFTER RECONSTRUCTION OF A LAYER (DEEP MAP).

Meshing of the point cloud from the last n layers (e.g. 20 < n < 100) and subsequent defect detection in order to identify possible errors for the current layer and correct them with appropriate laser correction. The number of layers n can possibly also be controlled adaptively.

SURFACE RECONSTRUCTION AND NETWORKING (MESHING):

The aim is to use the point cloud obtained to determine a surface representation of the desired objects. This means that adjacent points are connected with edges and 3 edges form a triangle, which can be used to describe a surface in three-dimensional space. One of the most widespread formats is the so-called STL format, which is also used for additive manufacturing and the additive manufacturing device, among other things.

• - Powercrust
• - Marching cube
• - Poisson Surface Reconstruction
• - Hoppes Reconstruction

In 7 a point cloud and the associated triangular network (mesh) are shown. The transition can be implemented using a variety of different algorithms. The following algorithms are mentioned here as an example:

• - power crust
• - Marching cube
• - Poisson surface reconstruction
• - Hoppes Reconstruction

Delaunay triangulation algorithms create a 3-dimensional object, i.e. this is filled with tetrahedrons (volume based), and are therefore only suitable to a limited extent as there is some overhead.

It is important that the mesh has a valid representation. This means that there are no intersecting triangles in particular. Furthermore, it would also be desirable, but not essential, not to have so-called hanging-nodes.

Alternatively, one could carry out the surface reconstruction and the meshing after each lasered layer and thus try to localize defects in each layer. This procedure is more complex and finding and classifying defects at the local level is more difficult.

The following is a description of the object extraction. Once the meshing is complete, there is a representation of the entire construction space, i.e. areas next to the workpiece have also been reconstructed. In order to find errors by comparing the reconstructed object and the CAD model, it is necessary to extract the workpiece. For this purpose, the CAD model is used as a prior and the sub-area in the reconstructed pressure space that best matches the model is found.

MATCHING AND LOCALIZATION (GLOBALTO LOCAL) OF DEFECTS:

• - Verzug (distortion)
• - Poröse Strukturen,Risse, Löcher etc.

Each extracted area can now be compared to the model and checked for errors using the original CAD geometry. All recognized structures that are now inside a geometry, in the case of certain defects such as warping, possibly also structures that are outside of a geometry, are to be seen as potential sources of error and may have to be examined more closely, e.g. by calculating the volume or the associated bounding box to determine the size of the defect. At this point, a distinction must also be made between the following defect types:

• - distortion
• - Porous structures, cracks, holes etc.

A graph is generated from the mesh; a commercial program library is sufficient for this. All individual, connected graphs form a separate object. In addition, the printed object or the type of defect can be determined with the bounding boxes and volume (size comparison). With this procedure, the defect extraction and classification can be efficiently implemented. In addition, a distinction can be made here between a macro and micro scale and thus appropriate countermeasures can be initiated, i.e. merging the defect (increased performance, time, parameters, etc.) or a print termination of the corresponding component.

In order to be able to show the customer, i.e. the user of the additive manufacturing device, the defects, they are displayed on a data display device. Ultimately, this software should be embedded in the entire software or be able to be used via an API. In a first step, a defect file is created, which can be visualized with ParaView, for example.

The individual defect types are displayed in different colors so that they can be distinguished from one another quickly and easily.

DEFECT DETECTION THROUGH IMAGE PROCESSING:

In contrast to the reconstruction-based method from the first exemplary embodiment, the method described below makes fewer assumptions about the type of defect, only that it manifests itself robustly in the selected sensor, for example RGB or NIR/IR camera. The principle of the method is based on the observation that it is often easy to generate a large amount of "good" sensor data, i.e. in which there is no defect. In this way, a generative model that explicitly or implicitly represents the (spatial) distribution of the “good” data can be learned. A defect is then defined as a data point to which the learned distribution assigns a low probability, i.e. so that the statement is made that the observed data point is probably not a "good" sensor datum.

1. 1 Die Zuordnung einer niedrigen Wahrscheinlichkeit heißt nur, dass eine Anomalie vorliegen kann, nicht aber, dass diese Anomalie auch zu einem Defekt führt bzw. geführt hat.
2. 2 Die Art des Defekts ist nicht gegeben, bspw. liegen Porosität oder unvollständige Verschmelzung vor.
3. 3 Die Lage des Defektes ist nicht unmittelbar ablesbar, was allerdings abhängig vom generativen Modell ist.

The process has the following challenges:

1. 1 The assignment of a low probability only means that an anomaly can exist, but not that this anomaly also leads or has led to a defect.
2. 2 The type of defect is not given, e.g. there is porosity or incomplete fusion.
3. 3 The location of the defect cannot be read immediately, but this depends on the generative model.

It is therefore advantageous to combine the method described with an additional step, for example the method described above or human intervention in the printing process when defects are detected.

TEMPERATURE FROM NIR/IR IMAGES:

In this exemplary embodiment, the camera system 40 is capable of detecting (near) infrared (NIR/IR). In the exemplary embodiment, a further individual camera is provided for this. In order to make the error detection more robust, the temperature on the surface of the component can be continuously measured during printing using a (near) infrared (NIR/IR) camera. This allows spatial and temporal temperature gradients to be determined, which provide information about the melting process and thus also about possible errors. Furthermore, the temperature is a defect characteristic, which is often at the beginning of the causal chain for the development of a defect, so that a defect can be identified earlier than, for example, the depth-based method, since the defect is already present here.

A challenge faced by NIR/IR cameras is their spatial resolution, which is lower than RGB cameras, i.e. a pixel in an RGB image cannot directly correspond to a pixel in an IR image. Conversely, a pixel in the IR image is only the temperature average from the implicit temperature values that would be determined if the NIR/IR camera had the same resolution as the IR camera. This can mean that small defects cannot be detected (early). Furthermore, the resolution is often also lower than the diameter of the laser beam on the surface of the component, with the same effects as before.

Temperature information is combined with the methods described above to enable robust fault detection.

EXTENSION OF THE MELT LASER IN THE SLM PROCESS TO MEASURE DISTANCE AND TEMPERATURE:

For defect detection using 3D reconstruction, it is necessary to accommodate appropriate sensors on (outside) or in the installation space of the additive manufacturing device without having to change the design too much in order to keep costs within an acceptable range. A setup of optical and NIR/IR camera as well as LiDAR contradicts the postulate. One possible solution is to use the melting laser to measure the component to be printed and to determine the temperature on the surface. The idea uses the functionality of a LiDAR. The usual measurement principles according to which LiDAR systems measure distances are shown below.

PULSED DISTANCE MEASUREMENT:

dabei bezeichnet c die Lichtgeschwindigkeit und t_0f die Laufzeit des Lichts.In the simplest case, a laser pulse is directed at the object to be measured and the time difference between the exit of the pulse and the arrival of the reflection of the pulse at the object is measured. Since the speed of light is constant, depending on the surrounding medium, the distance between the object and the laser can be determined from it. The geometry of the component can then be determined on the basis of the process parameters, e.g. layer index, and defects can be determined as described above. The distance d to the component is given as follows: $$\mathrm{i.e}=\frac{c}{2}{t}_{Of},$$

where c denotes the speed of light and t _0f the transit time of light.

The additive manufacturing device must be expanded to include a time measuring unit and a photodetector.

AMPLITUDE-MODULATED CONTINUOUS WAVE DISTANCE MEASUREMENT:

wobei ΔΦ die Phasenverschiebung ΔΦ und ƒ_m die Frequenz des Signals beschreibt, dass durch die Amplituden des reflektierten Signals gebildet wird.In contrast to the pulsed method, the laser in the AMCW method emits a constant wave. During emission, the radiant power of the laser is modulated with a reference signal. The phase shift of the reflection and the reference signal is then determined. The phase shift determined is a function of the distance, which can be calculated from this. $$\mathrm{i.e}=\frac{c}{2}\frac{\Delta \Phi }{2\pi f_M},$$

where ΔΦ describes the phase shift ΔΦ and ƒ _m the frequency of the signal that is formed by the amplitudes of the reflected signal.

A challenge associated with this process is the continuous change in the radiation power and thus the melting process on the powder surface. Common process parameters are therefore probably obsolete and have to be found again.

In order to implement the method described, the additive manufacturing device must be expanded to include components that can modulate the laser beam and determine the phase shift.

The last method, Frequency-modulated Continuous Wave (FMCW), on the other hand, already achieves resolutions in the 150 µm range.

FREQUENCY-MODULATED CONTINUOUS WAVE DISTANCE MEASUREMENT

In contrast to the AMCW method, the frequencies of the carrier signal are modulated in the FMCW method, ie the wavelength of the laser is modulated depending on a reference signal. The distance d to an object is then a function of the frequency shift of the echo signal and can be represented as follows $$\mathrm{i.e}=\frac{c}{2}\frac{\left|\Delta f\right|}{2\mathrm{i.e}f/\mathrm{i.e}t}.$$

where |Δƒ| denotes the frequency difference and df/dt the frequency shift per unit time.

A challenge that arises with this method is changing the frequency of the laser beam and thus the wavelength. The wavelength affects the absorption of energy at the material surface, so the process parameters for a print may have to be adjusted.

The additive manufacturing device must be expanded to include components similar to those used in the AMCW process.

Furthermore, the reflectivity of a surface changes depending on its temperature. If the reflectivity drops, the intensity with which the reflected laser pulse arrives at the photodetector also changes. This can be described as follows $$\Delta R=\frac{\partial R}{\partial T}\Delta T+\frac{\partial R}{\partial N}\Delta N.$$

In the equation above, R describes the reflection factor and T describes the temperature. The second part of the equation describes the changes in the reflection factor as a function of the change in the number density of the charge carriers ∂N caused by the optical excitation.

STRUCTURAL ADJUSTMENT OF THE DEVICE FOR CREATING A THREE-DIMENSIONAL OBJECT

The functional expansion of the melting laser is particularly effective when structural changes are made to the additive manufacturing device. Since the laser is very sensitive to movement, the beam is directed onto the surface using mirrors and lenses. This means that the angle of incidence and the angle of reflection change depending on the position. Furthermore, situations can arise in which the reflection goes directly back into the focusing lens, i.e. when the angle of incidence of the light is close to 90 degrees. As described above, the additive manufacturing device must be expanded to include photodetectors, among other things. Since the reflected laser beam no longer has a fixed position in the installation space due to the continuous change in the angle of reflection, the additive manufacturing device must be designed in such a way that the individual elements are arranged in such a way that the light transit time and light intensity measurements are carried out precisely.

MODIFICATION OF EMBODIMENTS, REPLACEMENT OF F-θ LINS 14

Depending on the configuration and arrangement of the F-θ lens, it can be problematic to also use the production laser 6 as a measuring laser, since the refraction at the F-θ lens can lead to beam distortions that must be taken into account when calculating the laser beam propagation time and can lead to inaccuracies. In particular, refraction distortions that are difficult to compensate for can occur as a result of the F-θ lens 14 after the reflection of the laser beam 9 when it passes through the F-θ lens 14 again.

In order to eliminate this problem, it is proposed to replace the F-θ lens 14 in a first modification of the embodiments. This is illustrated below 9 described.

A fixed mirror 15 is provided in place of the F-θ lens. The fixed mirror 15 is horizontally in the shape of a segment of a circle, with the scanner 8 being arranged in the center of the circle. The mirror surface of the fixed mirror 15 is inclined downwards in such a way that a beam 9 arriving horizontally from the scanner 8 is reflected vertically downwards onto the building layer 11 . The scanner 8 has a mirror which rotates about the vertical axis (z-axis) in such a way that the beam 9 sweeps over the entire horizontal extension of the fixed mirror 15 and is reflected vertically downwards by the fixed mirror over this area. This produces a one-dimensional area in the shape of a segment of a circle on the construction layer 11, on which the beam 9 impinges. In order to be able to process a flat area on the construction layer 11, the laser 6, the mirror 7, the scanner 8 and the fixed mirror 15 are moved over the construction space in a constant arrangement relative to one another in the x-direction. A portion of the beam 9 reflected by the building layer 13, the reflected beam 9', is reflected back onto the scanner 8 via the fixed mirror. Since the mirror of the scanner 8 has already rotated further about the z-axis by a sufficiently precisely predeterminable value when the reflected beam 9' arrives, the reflected beam is not reflected directly to the laser 6, but hits it there slightly offset. At this point, a laser receiver 16 is provided, which is fixed relative to the laser 6, the mirror 7, the scanner 8 and the fixed mirror 15 and moves together with them along the x-axis.

Because of this arrangement, the beam 9 is always aligned perpendicularly to the building layer 11 and the path length of the beam 9 and the reflected beam 9' varies only by exactly the distance of height differences on the building layer 11 perform accurate depth determination using LiDAR.

Alternatively, it is proposed to modify the F-θ lens 14 in a second modification of the embodiments. This is illustrated below 10 and 11 described.

A modified F-θ lens 14' is designed to be provided with a partially specular surface 141. FIG. An annular fixed mirror 15' is also provided. Two laser sensors 16' are attached to a ring 17 in such a way that they can move on the ring. Both the partially reflecting surface 141 and the ring-shaped fixed mirror 15' are configured quasi-parabolic, so that the reflected beam 9' is always reflected onto the position of the orbital ring 17. The sensors 16' are moved along the revolving ring 17 in such a way that the reflected beam 9' can be detected either by one or the other sensor 16'. Since the reflection can suddenly change from one side to the other when crossing over the center of the image, it is advantageous to arrange the sensors 16' opposite one another on both sides.

The arrangement according to the second modification does not permit as high an accuracy as the arrangement according to the first modification because the beam alignment is not always vertical. However, since here with the sensors 16′ considerably fewer components can be moved relative to the installation space, the complexity is lower.

HYBRID APPROACH TO DEFECT DETECTION

In the first step, the lidar sensor is calibrated online, i.e. with each image. With the calibration parameters it is now possible to create a disparity map from the lidar data. This is where the calibration parameters come into play, because they allow the disparity map to be calculated for the same image section that the stereo camera sees (lidar scans in radius X, i.e. the point cloud must be aligned with the stereo image). In the next step, a global disparity map is created from the two disparity maps, which is much more precise than the individual maps. Both steps, calibration and fusion can be implemented with neural networks.

DYNAMIC, IN-SITU ADJUSTMENT OF PROCESS PARAMETERS:

On the one hand, defects are caused by factors that already exist before printing, e.g. the geometry is not suitable for printing or there were quality defects during the production of the powder. On the other hand, defects also occur during of the print because process parameters such as the scanning speed of the laser do not match the current state of the print. In existing environments of additive manufacturing devices, the process parameters are determined and set before the manufacturing process. The determination can be made by simulating the manufacturing process and/or by creating samples. However, in order to effectively avoid defects, it is necessary to dynamically change the process parameters during runtime, in response to the defect probability existing at a point in the process. For this purpose, the manufacturing process is formulated as a Markov decision problem, which can be solved with an algorithm from the class of reinforcement learning.

SELF-LEARNING PARAMETER ADJUSTMENT:

The algorithm (policy) continuously receives information about the manufacturing process, e.g. temperature in the powder bed, and decides whether the process parameters need to be changed with the aim of reducing the probability of defects.

On the one hand, there are model-free approaches that do without learning an explicit model of the dynamics of the additive manufacturing process and try to assign the best change in the process parameters to each state, e.g. using an artificial neural network. In the context of reinforcement learning, we speak of the dynamics of the environment (environment), which is modeled by the transition probability distribution of the states p(s',r|s,a). s, s' ∈ S are the states, a ∈ A is a change in the process parameters (actions), r ∈ R provides information about how good the change in the process parameters was (rewards), i.e. it reduced the defect probability. In this context, the additive manufacturing device, the manufacturing process and the component can be viewed as the environment.

Alternatively, a model-based approach can be used. In the model-based approaches, the environment is explicitly modeled in order to plan the best change in the process parameters. Planning means that there is an actual state and a target state. The planning algorithm then tries to find the best sequence of changes in the process parameters (policy) with the help of the model of the environment. The best sequence in this context is the one that leads to the lowest defect probability.

The advantage of model-based approaches is that they require less data (sample efficiency), since many relationships are (locally) linear in reality. This means that only a few data points from the local environment are needed to learn the connection.

Model-free approaches, on the other hand, need significantly more examples so that they only learn the connection between changing the process parameters and the success (reward). However, the described disadvantage of model-free approaches disappears if precise simulations of the additive manufacturing process can be used or if the number of printers in use whose data is accessible is large, because planning in model-based approaches is essential more complex and is more constrained by simplifying assumptions and approximations than is the case with model-free approaches. The challenge in model-free approaches that have been trained in a simulation is the successful transfer to the hardware. Even with highly accurate simulations, there is a so-called reality gap, i.e. the learned model recommends suboptimal changes to the process parameters, since the simulation model differs too much from reality in some or all dimensions.

It is taught to combine a model-free and a model-based approach, first learned in simulation and then refined on a real additive manufacturing fixture with few manufacturing examples. Furthermore, the use of methods that reduce the reality gap, such as dynamics or domain randomization and measurement of the physical parameters are taught. For dynamics and domain randomization, critical parameters of the environment are repeatedly varied during the simulation in order to force the policy to be learned to be more generalized. Furthermore, the use of already collected data on the manufacturing process is taught to accelerate learning (learning from demonstrations).

A challenge is the heterogeneity of the individual additive manufacturing devices, even if the same process is used. To counter the problem, a collective learning approach is proposed.

COLLECTIVE LEARNING

Additive manufacturing devices from different manufacturers differ in many ways, even when they use the same manufacturing process. This has an impact on the dynamics of the environment in the manufacturing process. It is difficult to simulate every single additive manufacturing device in advance and adjust the parameters of the algorithm accordingly. Therefore, two methods are proposed to Use data from each additive manufacturing device to improve the algorithms. Methods come from the field of Privacy Preserving Machine Learning, which make it possible to protect sensitive data of individual customers without impairing learning.

OPEN EXCHANGE (CENTRAL LEARNING)

Each additive manufacturing device on which our software is installed continuously transmits process parameters and the states of its environment to a central server. In this case, confidential information such as the geometry of the component is modified so that the data can still be used to learn the algorithm, but the original geometry or the confidential information can no longer be determined. Based on the data and the procedure described above, the algorithm is improved and the improvements are transmitted back to the additive manufacturing device.

SEMI-OPEN EXCHANGE (FEDERATED LEARNING)

Each additive manufacturing device on which our software is installed has an additional processing unit that allows the model to be improved using the data generated in the machine. For this purpose, the printer already has a generic model when it is delivered, which is then adapted to the specific properties of the printer over the runtime through the learning process. After such a local model has been improved with the data from the additive manufacturing device, the changes in the model parameters are transmitted to a central server, which improves a global model with the individual updates. The global model is then sent back to the additive manufacturing device to replace the old model.

The invention has been described by means of exemplary embodiments. The embodiments are only illustrative in nature and do not limit the invention as defined by the claims. Within the scope of the claims, configurations that can deviate from the exemplary embodiment are possible, recognizable to a person skilled in the art.

Thus, in the exemplary embodiment, the manufacturing laser of the additive manufacturing device is used to detect the surface temperature or the shape of the workpiece. Alternatively, a separate detection device with a separate detection laser can also be provided.

In the exemplary embodiment, a camera system consisting of two image recording devices that record images at the same time is used. Alternatively, the camera system can consist of an image recording device that records images offset in time and space.

In the modified exemplary embodiment, the camera system has an additional camera for recordings in the NIR/IR spectrum. Alternatively, one of the stereo cameras or the single camera can be used for this function if it has the ability to capture the NIR/IR spectrum.

In the exemplary embodiment, the camera pose is localized, among other things, on the basis of information from rotary encoders of the motors that change the camera pose, from an inertial measurement unit (IMU) attached to the camera, or an interior localization technique (e.g. RFID, magnetometer, etc.). Alternatively or additionally, the camera pose can also be determined using a further camera and a marking on one of the housing walls. Such an embodiment is based on the 4 described. The embodiment after 4 corresponds to the first exemplary embodiment and differs from it in the points listed below. The camera system 50 has a further camera module 60 . This is connected in a rotationally fixed manner to the first camera module 52 and the second camera module 53 and is arranged on the carriage 51 . In contrast to the first camera module 51 and the second camera module 53, the third camera module 60 is oriented upwards. The upper housing cover 43 usually consists of a glass pane. A grid 44 that can be seen from below is applied to this. The third camera module 60 can capture the grid 44 and use it to determine its position on the xy plane and its orientation.

In the first, second and third exemplary embodiment, the camera systems are provided within the construction space 40 . However, they can also be provided outside of the installation space 40 . In this case, the installation space must be equipped with appropriate windows. This has the advantage that the camera systems are not exposed to the thermal stress that prevails in the installation space. The camera systems can be arranged above or on one or more of the four sides of the installation space.

In the exemplary embodiment, the device according to the invention for constructing a three-dimensional production object is shown as a laser sintering device. However, the invention is also applicable to Direct Energy Deposition manufacturing devices gene applicable. Since there is no longer a powder bed, the object must now be separated from the background during segmentation. Due to the fact that the background is not equipped with the same regularity as in the powder bed process, it must be taken into account when segmenting that the apparatus that applies the material (powder or wire) (MAA) (e.g. a robotic arm) differs from the production object and the Production object to be separated from the background. Structural changes could be that the MAA are painted in a uniform color that is not adopted by the material. Furthermore, the background can be shielded and colored in such a way that it can be easily segmented, e.g. by printing in a separate cell. With this method, depending on the structure of the production device, the sensors do not have to be attached in the upper part of the installation space, but can either be attached directly to the MAA or on the side in / on the printer or in an independent, mobile device such as a robot arm . It is important that cavities and other areas that are only visible from a few perspectives can be fully reached, with attachment close to the MAA being advantageous. Furthermore, the provision of different components around the sensor is advantageous in order to form protection against the heat or the provision of different filter systems in order to reduce the high radiation intensity (light intensity) but also the influence of UV radiation / IR radiation on the image (RGB ) to reduce. The advantage of additional devices results from the fact that there are few / no phases in which the radiation intensity generated by the MAA is low, such as in the powder bed process when the recoater is applying new powder. In the case of local depth estimation, if the camera is attached directly to the MAA so that it specifies the path, the path planning is limited to determining the trigger moments for the image recordings and - if the camera modules can be aligned independently - planning the camera alignment .

In the first exemplary embodiment, the camera system 50 has the first camera module 52 and the second camera module 53 in order to carry out stereoscopic depth detection using two camera images. As an alternative to this, depth detection is possible using the grazing light method. In this case, the second camera module 53 is replaced by a projector which projects a light pattern (structured light) onto the building layer 11 . The first camera module 52, which is offset from the second camera module 53 and thus also from the projector and is aligned with the building layer 11, takes images of the building layer 11 with the light pattern according to the camera system path. From the perspective of the first camera module 52, a change in the depth of the construction level 11 results in distortions of the light pattern. From the distortions of the light pattern recorded in the recordings of the first camera module 52, the changes in depth of the building layer 11 can be determined in a known manner on the basis of epipolar geometry and a point cloud can be generated according to the first exemplary embodiment. However, the high accuracy in the micrometer range achieved by the grazing light method is offset by higher costs due to the need for one or more projectors.

If the projector is provided in addition to the camera systems 50 of the exemplary embodiments, the depth determination by the grazing light method can also be carried out in addition to the stereoscopic depth determination.

Alternatively or as a substitute, a depth determination can also be carried out using radar. For this purpose, a radar sensor is provided on (outside) or in the upper area of the installation space 10 . The radar sensor has a radar emitter and a radar receiver. The radar emitter emits electromagnetic waves in the radio frequency range as a radar signal, which are reflected by the building layer 11 . The reflections of the radar signal are received by the radar receiver. The distance is determined from the signal propagation time, i.e. the time difference between sending and receiving the radar signal and the propagation speed of the radar signal in the corresponding surrounding medium in and/or outside the installation space. In this respect, the procedure is similar to lidar depth determination. The results from the radar depth determination can be combined/fused with the other depth determination methods, or replace individual other depth determination methods.

A camera system path is a sequence of positions, orientations or image acquisition initiations of a camera system with one or more camera modules.

An additive manufacturing device within the meaning of this document is any material-applying manufacturing, for example by LSM, DED (direct energy deposition) or MAA.

A partially reflecting mirror or a partially reflecting surface is one that reflects the radiation irradiated from one side and at least partially irradiates the radiation from the other side, at least more than when irradiating one side.

In the description, the terms "and", "or", and "either ... or" are used as a conjunction in the Meaning of the logical conjunction (mathematical AND), used as a junctor in the sense of the logical adjunction (mathematical OR, often "and/or") or as a junctor in the sense of logical contravalence (mathematical exclusive OR).

The method steps specified in the description or the claims only serve to list the required method steps. They imply an order only where the order is explicitly stated or deemed necessary by those skilled in the art. In particular, the listing does not imply an exhaustive enumeration.

Reference List

1

Creation device, device for creating a three-dimensional object, laser sintering device, LSM, printer

5

upper partition

6

Lasers, production lasers, measuring lasers, detection lasers

7

deflection mirror

8th

scanner, reflector

9

beam, laser beam

9'

reflected beam

10

installation space

11

build layer, current build layer

12

entry window

13

build plate

14

F-θ lens

141

partially reflecting surface

15

fixed mirror

15'

ring-shaped fixed mirror

16

Laser receiver, laser receiving device

16'

laser sensor

17

circulation ring

25

container

26

carrier device

27

Coater, recoater

28

dosing device

30

feed opening

40

upper area of the installation space 10

41

lower area of the installation space 10

50

camera system

50'

Camera system (2nd embodiment)

51

Sleds

52

first camera module

52'

first camera module (2nd exemplary embodiment)

52"

first camera module (3rd embodiment)

53

second camera module

53"

second camera module (3rd embodiment)

a

Tilt angle of the first camera module 52"

β

Tilt angle of the second camera module 53"

54

first servomotor

56

Carrier of the rail system

57

third servomotor

58

rails

59

actuator

61

bearing ring

62

rotating ring

63

actuator

dT

temperature gradient

T

Probability whether pixel covers melted or unmelted material

Claims (9)

Method for preventing manufacturing defects during additive manufacturing of a manufacturing object, comprising the steps of: providing a detection system (50, 50', 50", 6, 8, 14, 15, 15', 16, 16') in the additive manufacturing device (1 ) for detecting at least one property of the production object, production of the production object in several production steps, during production the detection system (50, 50', 50", 6, 8, 14, 15, 15', 16, 16') the at least one Property of the production object is determined, comparison of the at least one property with a target state of the property, determination of deviations between the at least one property and the target state of the property, evaluation of the deviation determined in the deviation determination step if deviations are determined, Determination of an adjustment of at least one process parameter, based on the deviation determined in the deviation determination step to reduce the deviation in the further production process, application of the adjustment of the at least one process parameter determined in the adjustment determination step. The method according to the preceding claim, wherein the at least one process parameter is one or more from the group of installation space temperature, production laser intensity or production speed. The method according to any one of the preceding claims, wherein the detection system (50, 50', 50", 6, 8, 14, 15, 15', 16, 16') is a camera system (50, 50', 50") and/or a projector camera system and/or a laser detection system (6, 8, 14, 15, 15', 16, 16') for detecting a three-dimensional configuration of the production object or for detecting a surface temperature of the production object. The method according to any one of the preceding claims, wherein the adjustment is determined in the adjustment determination step using a model-free algorithm. The method according to the preceding claim, wherein the adjustment is determined in the adjustment determination step using an artificial neural network for modeling transition probabilities. The method according to any one of the preceding claims, wherein the adjustment is determined in the adjustment determination step using a model-based algorithm. The method according to the preceding claim, wherein the adjustment is determined in the adjustment determination step using an artificial neural network for modeling optimized sequences of changes in process parameters based on system modeling for transferring the additive manufacturing device from an actual state to a target state. The method according to the preceding claim, wherein for dynamics and domain randomization environmental parameters are varied during a system simulation in order to force the artificial neural network to higher generalization. Device for additive manufacturing, which is set up to carry out a method according to the preceding method claims.

Images