Method for additive manufacturing of components

Abstract

The invention relates to a method for additive manufacturing of a component 2, comprising the following steps: - creating a machine code, - creating a generalized anomaly detection model and creating an adaptive anomaly detection model - transmitting the machine code and the detection model to a control device - starting a manufacturing process - monitoring the process using sensors - evaluating sensor signals of the manufacturing process with the aid of the generalized anomaly detection model - training or learning a dedicated anomaly detection model in parallel from the adaptive anomaly detection model using process data of the ongoing manufacturing process - detecting anomalies in component manufacturing with the aid of the dedicated anomaly detection model during the manufacturing process.

Description

[0001] Wire Arc Additive Manufacturing (WAAM) is an example of a method for mass production of metal parts. Another typical method for this is the laser melting method (LMD). Here, the lack of repeatability of the process is a challenge for quality assurance. In order to ensure that the components have a sufficiently high quality, the process must be monitored. In situ monitoring systems enable analysis during the process at the point where it occurs.

[0002] If a defect in the process is then identified, the knowledge gained about the defect has not been used to implement process adjustments. This is because it is not known which defect exists and how to overcome it. Therefore, if an anomaly is detected, the process must be interrupted immediately and the component must be scrapped.

[0003] The technical problem to be solved by the present invention is to provide a method for additive manufacturing of components which has better process control and reduces the scrap rate during production compared to the prior art.

[0004] The above technical problem is solved by a method for additive manufacturing of a component according to claim 1. The method according to the invention comprises the following steps:

[0005] - Create machine code

[0006] - Create generalized anomaly detection models and create adaptive anomaly detection models

[0007] -Transfer machine code and detection model to the control device

[0008] - Start the manufacturing process

[0009] -Use sensors to monitor the process

[0010] -Evaluation of sensor signals for manufacturing processes using generalized anomaly detection models

[0011] - Train or learn a dedicated anomaly detection model in parallel from the adaptive anomaly detection model using process data from an ongoing manufacturing process

[0012] - Detect anomalies in component manufacturing with the help of dedicated anomaly detection models during the manufacturing process.

[0013] The terms used herein are defined as follows:

[0014] Machine code

[0015] Machine code is a sequence of instructions created according to the rules of a high-level programming language, which can be directly executed (without further translation) by a processor of a device, such as a control device with an electronic data processing system. An example of a machine code for an industrial control system such as Sinumerik is G-code (DIN 66025). Another example is a specialized robot programming language such as KRL (KUKA Robot Language) or RAPID.

[0016] Model

[0017] On the one hand, models are used to map parts of reality in order to solve tasks with the help of information processing. Such models are called domain models. For example, they include models for the software to be created, especially models for its code (such as in the form of program flow charts) and data models (such as in the form of neural networks) for describing the structure of the data to be processed from the perspective of operation / professional logic or from the perspective of technical data status.

[0018] abnormal

[0019] The anomaly relates to the component to be manufactured and usually represents a defect or a precursor to a defect, such as an oxide cluster in the component being produced. Such anomalies in the component should be detected by sensors as far as possible during the manufacturing process and likewise compensated or overcome during the manufacturing process.

[0020] Generalized Anomaly Detection Model

[0021] A model, such as a neural network, is adapted to represent training data by learning so that output data can be calculated from input data. The training data is, for example, a data set of sensor data from an already performed manufacturing process that has provided known, such as good, process results in the form of components.

[0022] A generalized anomaly detection model is a model that is trained based on available training data; wherein the purpose of training is to generate a model that has high accuracy for all parameter sets when calculating output data based on input data. A generalized anomaly detection model can be used for different parameter sets, but it has lower computational accuracy compared to a dedicated model. A dedicated model is a model trained on a dataset for a specific parameter set, which is based on an adaptive model. It is conditionally trained on a new parameter set in a few steps and / or in a short time. It can be used specific to a parameter set and due to its specificity has higher accuracy when calculating anomalies based on a dedicated parameter set compared to a generalized anomaly detection model.

[0023] Process data

[0024] These are all the data generated during a process. They include sensor signals and at least partly process parameter sets.

[0025] Process parameter set

[0026] The process parameter set includes all technical variables used during the process, such as the position and speed of the print head at a certain time, but especially the machine settings, such as energy input (e.g. in the form of current and voltage), welding gun distance from the component, gas flow or arc length correction.

[0027] Adaptive Anomaly Detection Model

[0028] It is a model based on neural network training conditioned on existing training data (i.e. specific to a parameter set); the purpose of the training (e.g. through meta-learning methods, such as model-agnostic meta-learning or the Reptile method) is to generate a model that can be quickly adapted to new process parameter sets, especially within a time period of 0.5 s to 3 s and with less training data.

[0029] Adaptable means that the model can reach a dedicated state with as few additional parameter set-specific training data and / or as few steps and / or as short a time as possible. That is, the dedicated anomaly detection model is generated based on the adaptive model. The adaptive anomaly detection model itself is not directly applicable to anomaly detection. Therefore, it is not well suited because it is designed to adapt as quickly as possible and with little training data, rather than to achieve the best possible detection results.

[0030] Controls

[0031] A control is a computing unit that can execute models and process data. Examples of controls are NC (numerical control systems, such as Sinumerik), edge devices (Industrial Edge or Sinumerik Edge (a computer that connects two networks and can evaluate data)), IPC (industrial PC that is not connected to the second network) or PLC (programmable logic controller, such as SimaticS7)).

[0032] Sensors and sensor signals

[0033] A sensor, also known as a detector, (measured variable or measurement) recorder or (measurement) inductor, is a technical component that can qualitatively or quantitatively detect certain physical or chemical properties (physical properties such as heat, temperature, humidity, pressure, sound field quantity, brightness, acceleration or chemical properties such as pH value, ionic strength, electrochemical potential) and / or material properties of its environment as a measured variable. These quantities are detected by physical, chemical or biological effects and converted into sensor signals that can be further processed, especially electrical signals. Electrical signals are a special form of physical signals. Electrical signals are electrical variables, such as current intensity, voltage or resistance, if they are variable in any way and can therefore record and transmit information. The values ​​of the signal include, for example, DC values, peak values, frequencies, phase shift angles or duty cycles. The values ​​provided by the sensor signal are called sensor data. Mathematical evaluation, that is, further processing of physical sensor data, is also called sensor data.

[0034] Print Head

[0035] The part of the additive manufacturing device that is responsible for the material application is called the print head. In the WAAM method, the print head is the element that guides the ends of the wires to be welded; in the LMD method, the print head is the element that emits the laser beam to melt the powder. The print head is positioned at a certain position during the process by defined features, such as the electrode tip or the wire outlet. This position of the print head is in turn spatially equidistantly associated with the working points in or on the surface of the generated component. The position of the print head and the working points on the component are therefore different, because in the various additive manufacturing methods they do not coincide spatially, but are always related in a controlled manner.

[0036] Location data

[0037] Position data is a set of data whose position is based on a coordinate system defined before the process begins relative to the environment of this defined process.

[0038] Digital Twin (Digitaler Zwilling)

[0039] A digital twin is a digital representation of a tangible or intangible object or process from the real world in the digital world. A digital twin enables a comprehensive data exchange. A digital twin is no longer just data, it consists of a model of the represented object or process and can also contain simulations, algorithms and services that describe, influence the properties or behavior of the represented object or process, or provide related services.

[0040] The present invention has the following advantages over the prior art. In order to be able to cope with a large number of process parameters, conventional algorithms are used, such as the formation of average values ​​in combination with limit values. It is in the case of high-frequency data and complex patterns in the data that information is lost because they are not used for analysis. In order to be able to analyze complex signal characteristics, the above-mentioned adaptive anomaly detection model is considered. For this purpose, a sufficient number of training data sets with parameter sets are collected in the process. With the help of these data, an anomaly detection model specifically for the determined parameters is generated from the adaptive anomaly detection model. If such a model exists, it can be applied to the sensor signal under the condition of this special parameter set for anomaly detection. Adjustment of process parameters causes the dedicated model to become outdated, and adjustment through the adaptive anomaly detection model must be carried out again. Early and reliable finding of anomalies in the produced parts provides great advantages to the manufacturing process, because scrap can be identified earlier and manufacturing resources and manufacturing time can be saved, or in more favorable cases, the detected anomalies can be eliminated in situ.

[0041] It is expedient here to store the adaptive anomaly detection model trained with respect to the process parameter set as a dedicated anomaly detection model and to use the dedicated anomaly detection model when the process parameter set is used again. This shortens the training time of the adaptive anomaly detection model, during which only a generalized anomaly detection model is otherwise provided.

[0042] Furthermore, it is advantageous if the adaptive anomaly detection model can be used when starting a second manufacturing process with a second set of process parameters and adapted with the aid of new training data to create a further dedicated model.

[0043] In another advantageous design of the present invention, the manufacturing method further comprises the following steps:

[0044] - Building a digital twin 10 of the produced component 2 in parallel during the process from the sensor data including the position data of the detected anomaly 8

[0045] - Predicting the position of the print head 14 at a certain time by machine code 12

[0046] -Analysis of the work area around the location in terms of existing anomalies through digital twins

[0047] -Adapt process parameters to eliminate anomalies when reaching the working area.

[0048] The above-mentioned anomaly detection can detect anomalies as early as possible and deal with them in the parts being produced by adjusting process parameters at the location or around the anomaly, wherein the process strongly depends on the current environment (especially the melting zone in the parts being produced). For example, if oxidation occurs as an anomaly in the third layer, this will also lead to abnormal process characteristics in the subsequent layers. The same applies to adjacent welds. The above-mentioned anomaly detection first provides the current abnormal data points. In the design of the method, the background at the corresponding time point is taken into account and added to the simulation and implementation of compensation measures. Therefore, oxidation or shape deviations have an impact on subsequent layers and the data collected here. Now, the method has fully taken into account the impact of defects on multiple layers, so that the defects of multiple layers are related to each other. In this way, the existing information is used to obtain better results in the data analysis in the field of in-situ monitoring without loss. The time and space background, which is very important in the additive manufacturing process, is effectively utilized by the solution during data evaluation and defect overcoming measures can be initiated in situ.

[0049] Particularly suitable additive manufacturing methods in which the method can be advantageously implemented are wire arc additive manufacturing (WAAD) and laser metal deposition (LMD).

[0050] For this purpose, it is particularly advantageous to incorporate the results of anomaly detection by means of an anomaly detection model into the digital twin.

[0051] In this context, it is expedient if the digital twin is designed in the form of a point cloud and stores an evaluation of the process state for each point of the cloud.

[0052] The evaluation criterion is, for example, the type of anomaly (such as oxidation or porosity). Anomalies are therefore a precursor to defects, which can also be referred to as latent defects.

[0053] An outlier value is preferably determined for each point, the calculation of which advantageously includes information about neighboring points in the working area, ie, process data and sensor data.

[0054] As the working area, the spatial extent of the liquid phase that is dominant at the observation time point is preferably used. The shape of the double ellipsoid can be well understood mathematically in the evaluation, which has proven to be advantageous for the determination of the working area and its virtual transmission into the digital twin. It is suitable here that the working area changes continuously with the shape and size of the liquid phase, i.e. the local melt pool. The shape and size of the liquid phase can be detected, for example, by a welding camera or derived from the component temperature and the currently introduced energy. A special form of a double ellipsoid is an ellipsoid that is symmetrical with respect to the semi-major axis. A sphere is a special form of an ellipsoid.

[0055] Furthermore, it is expedient to carry out an adaptation of process parameters for eliminating anomalies at unusual locations or in an environment, for example to burn off oxides by locally increasing the heat input and / or to slow down the movement speed of the print head in order to melt and thus eliminate porous structures.

[0056] The decision whether and when to adapt the process parameters is based on a prediction of where the print head will move within a defined time period based on the machine code. The environment of the position thus calculated is then searched thoroughly for anomalies (e.g. by querying an octree with a double ellipsoid). If anomalies are present around the position, countermeasures are initiated (e.g. increasing the heat input, the speed of movement).

[0057] Particularly suitable additive manufacturing methods in which the method can be advantageously implemented are wire arc additive manufacturing (WAAD) and laser metal deposition (LMD).

[0058] Another aspect of the present invention is an additive manufacturing device for carrying out the method according to one of the method claims, comprising a robot arm, a control device and a print head as well as sensors for detecting process parameters.

[0059] Further embodiments and further features of the present invention are explained in more detail with reference to the following figures.

[0060] in:

[0061] Figure 1 A schematic diagram showing an apparatus for arc wire additive manufacturing,

[0062] Figure 2 shows the time course when training anomaly detection model during the process,

[0063] Figure 3 shows the effect of the training phase of anomaly detection models on the quality of inferences,

[0064] Figure 4 A diagram showing a point cloud of a generated digital twin of a component,

[0065] Figure 5 The working area in the form of a double ellipsoid for determining anomalies is shown.

[0066] Figure 1A schematic diagram of an additive manufacturing method for producing a component 2 is given. This additive manufacturing method is designed, for example, as wire arc additive manufacturing (WAAM). What is described below can also be applied similarly to other additive manufacturing methods, such as laser metal deposition. In this case, a print head 14 with an electrode 26 is directed to a specific position on the component 2 by a robot arm 24. In addition, a welding wire 28 is provided, which is melted at the corresponding position 12 of the print head 14 through the electrode 26. In addition, a control device 4 and an energy supply device 34 are provided, by which energy is supplied to the electrode 26 and the welding wire as well as the robot arm feed. The entire process is controlled by the control device 4, which will be described below. In particular, the process is monitored by a plurality of sensors. It is shown only by way of example. Figure 1 A sensor 6 in the form of a welding camera 30 is provided.

[0067] By means of the welding camera 30, for example, values ​​such as the geometry of the liquid phase 20 can be determined, which can also be Figure 4 Other sensor data such as voltage and current in the process, temperature of the liquid phase 20, feed of the print head 14 and feed speed of the welding wire.

[0068] For each position 12 of the print head 14, which position is clearly associated with a position in the component 2, position data in the form of a point in a coordinate system and a time stamp are preferably recorded and transmitted to the control device 4. In the control device 4 or a computer connected thereto, a digital twin 10 is created simultaneously during the printing of the component 2, such as Figure 4 Here, the digital twin 10 comprises a point cloud 16, wherein the print head 14 or its position 12 is also schematically indicated by a triangular point. Here, for each point 17, an outlier value is calculated from the sensor data determined from the point cloud 10 and transmitted to the control device.

[0069] In order to evaluate the outliers (which will be described later), we first define a working area 18 ( Figure 5 ), which is represented in the digital twin by an ellipse, which in the real component 2 essentially includes a liquid phase 20 in which the material of the welding wire has just melted on the surface of the component 2. A double ellipsoid is particularly suitable when considering the liquid phase. The working area is particularly important if detected anomalies are to be correlated, anomaly densities are to be determined or compensation strategies are to be implemented.

[0070] When applying the welding wire 28, anomalies 8 may occur in the liquid phase 20, which may be a precursor to a defect. For example, a defect may be an oxidation, which is caused by the presence of impurities on the welding wire 28 or by the introduction of impurities during the process. Air bubbles in the form of pores may also be anomalies and may become defects. When evaluating anomalies, not only the point 17 considered in each case is evaluated, but also the adjacent point 17' which is present in the working area 18 and is transmitted to the digital twin 10 in the ellipse 22. Figure 5 ,5 5a, 5b and 5c. For each point 17 or 17', an anomaly value is calculated and then taken into account when the print head 14 approaches the anomaly 8 determined in this way again. The print head 14 can be spatially tangential to the anomaly 8 from multiple sides, so that the anomaly can be compensated by changing the process parameters in the vicinity of the anomaly 8, thereby eliminating defects that may be caused by the anomaly before the defect occurs. Compared with conventional additive manufacturing methods, this method is limited in that the occurrence of defects is suppressed by the method and the rejection of the component 2 is greatly reduced.

[0071] An important basis for defect elimination is the early identification of defects by appropriate evaluation of process parameters and process data during the manufacturing process. This is done by anomaly detection models. The process progresses as follows when using anomaly detection models: Figure 2 Here, the three trajectories of the process along the time axis 38 are represented by three arrows 40, 42, 44. Here, the left side shows the course of the actual process 40, the states of the anomaly detection 42 associated therewith run parallel thereto in the middle, and the outer right side shows the time course of the training phase 44 of the anomaly detection model during the course of the process. The thin arrows between the individual boxes in the figure represent the data flow 46.

[0072] The process progression is designed here as follows. First, process data in the form of training data TI is generated, for example, from a previous process. It is input into a generalized anomaly detection model Mg for monitoring the process (Ps) at the beginning. The generalized anomaly detection model Mg is used in the first stage of the process to identify anomalies. As mentioned above, the generalized anomaly detection model Mg is a model trained based on available training data, wherein the purpose of the training is to generate a model that has a higher accuracy for all parameter sets when calculating output data based on input data. The generalized anomaly detection model Mg can be used for different parameter sets, but it has a lower calculation accuracy compared to the dedicated model MS1.

[0073] To illustrate how the generalized anomaly detection model Mg works, Figure 3 Two different graphs are plotted along the time axis in FIG. The normalized voltage or the difference between the actual and predicted normalized voltage is recorded on the y-axis (hence the axis is unitless). Figure 3 a shows here the prediction of process data by the generalized anomaly detection model 48, and therefore the following is a prediction of anomalies, which substantially but specifically approximately agrees with the real, measured data 50. Graph 52 reflects the error distance between curves 48 and 50.

[0074] As the process continues and possible anomalies in component 2 are predicted and thus detected by the generalized anomaly detection model Mg, the adaptive anomaly detection model Ma( Figure 2) through multiple iteration cycles to form a dedicated anomaly detection model Ms1. The time for forming the dedicated anomaly detection model Ms1 lasts between 0.5 seconds and 5 seconds according to process control, usually between 1 second and 3 seconds. If the formation of the dedicated anomaly detection model Ms1 by the adaptive anomaly detection model Ma is completed, the generalized anomaly detection model Mg is replaced and additional anomaly detection is performed by the anomaly detection model Ms1 dedicated to the process. This is the second stage of the process, i.e., the main stage. Figure 3 As can be seen in b, according to Figure 3 The modeled data 48 ′ fits the measured data 50 significantly better than in the case of the generalized anomaly detection model of FIG. 1 a.

[0075] Therefore, the time for running the second phase of the process 40 using the generalized anomaly detection model Mg is relatively short, and the majority of the process continues using the dedicated anomaly detection model Ms1 iteratively created with the help of the adaptive anomaly detection model Ma. This dedicated anomaly detection model Ms1 can be stored in the database and can be used again in the manufacture of parts with the same set of process parameters without having to start the process using the generalized anomaly detection model Mg in this case.

[0076] However, the further process Ps2 marked by the process arrow 40' starts again first using the generalized anomaly detection model using the new parameter set, wherein, according to the arrow 44, the further dedicated anomaly detection model Ms2 is trained in parallel with the adaptive anomaly detection model Ma, and the data from the first training interval is provided to the adaptive anomaly detection model Ma via the data flow 46.

[0077] The following is a detailed description of other aspects Figure 1-Figure 3 The method described. For integrating the time context, a (neural network) model is used, which evaluates the current (data) point 17 in terms of an outlier based on past values ​​(e.g. the past 0.5 seconds). The time period used as the time context here changes dynamically depending on the current welding speed (i.e. the feed speed of the print head 14) and the length of the weld pool. For example, the length of the weld pool can be evaluated by the welding camera 30 in combination with a computer vision algorithm, such as the Canny Edge Detection method. This ensures that all process values ​​relevant to the current liquid weld pool are always included in the calculation of the outlier.

[0078] To analyze the data, you can use regression models (forecasting), such as the seasonal ARIMA model or neural networks based on Conv1D or RNN / LSTM elements, or models that reconstruct the data with the help of autoencoding.

[0079] In addition to RNN / LSTM with temporal memory, batch CNN structures can also be used to predict data.

[0080] Temporal structures in the data, such as frequency changes or irregular patterns, are therefore also incorporated into the process evaluation. Process data with constant process parameters must be used for training in order to achieve a high degree of accuracy.

[0081] Preferably, each location reached by a print head 14 (position 12 of a print head 14 ) is given a time stamp which provides information about when it was reached.

[0082] The spatial context contains relevant information across layers, such as information from the previous layer or adjacent welds. For this purpose, as described above, a digital map is created simultaneously during the manufacturing process, i.e. a digital twin 10 of the component 2, in which all relevant information is stored. The digital twin 10 includes a point cloud, in which an evaluation of the process state (such as normal / abnormal, outliers, type of anomaly, etc.) and a timestamp are stored for each point. The point cloud can be stored in a spatial structure, so that spatial structures such as a double ellipsoid or a sphere can be queried more efficiently. Such data structures are, for example, octrees, r-trees or kd-trees, which can quickly find points in space due to their tree-like structure.

[0083] In order to integrate the spatial background into the calculation of the outliers of the current working point 19 (which is spatially equidistant from the print head), the space around the current working point 19 is included in the data analysis. For this purpose, a three-dimensional working area 18 is first defined around the working point 19, from which the spatial correlation of the data points 17, 17' is determined. The working area 18 here includes the weld pool, i.e. the area occupied by the liquid phase 20 during additive manufacturing. For simplicity, a sphere or a more general ellipsoid or a more general ellipsoid such as Figure 3 The double ellipsoid shown.

[0084] The axis length of the ellipsoid can be adaptively adapted to the geometry of the weld pool. This can be adjusted dynamically based on sensor data, such as data from a welding camera 30 (extraction of the weld pool geometry), by means of process parameters such as the process speed (higher welding speeds lead to thinner, longer ellipsoids) or based on previously performed thermal simulations (temperature curves, size and shape of the weld pool). Furthermore, the timestamps of the included points 17 or simulation results can be added to draw conclusions about the existing cooling behavior and, from this, the size of the weld pool.

[0085] Based on the defined working area 18, points 17 established in the process history are selected and enriched with meta-information, such as anomaly probability values, values ​​for defect criticality or defect types. For this purpose, data structures such as octrees can be used, which allow spatial indexing and thus an optimized spatial search for relevant points. Each point contains information about local process anomalies or defects, so that a spatially selected point cloud is obtained, which contains relevant information in the immediate environment of the current working point 19. The information obtained is now used on the one hand to improve the anomaly detection and on the other hand to adapt the process.

[0086] The result of the evaluation of the anomaly by the model (including the time context) is the anomaly value, which indicates the probability of the anomaly and thus also the probability of the defect. Here, the spatial context is now included in two different ways: If an anomaly has been identified, it can be mathematically and logically linked to other anomalies via the spatial context information. The spatial distribution of the anomalies is thus visible, and potential defects are also visible. Here, the above-mentioned double ellipsoid maps the related space in which the defects are linked to each other. If, for example, an oxidation occurs in the first layer, problems with the connection arise in the subsequent layer. These two defects are linked to each other, so that the spatial distribution of the abnormal areas in the component can now be better estimated.

[0087] If, for example, multiple potential anomalies with low anomaly values ​​are found in a high-load area of ​​a component, each of which does not lead to scrap alone, but together leads to component failure, then the anomaly density can also be determined to estimate whether a critical accumulation of anomalies has been reached in the component area. Here, the background space for the anomaly density does not necessarily include only the space of the weld pool (liquid phase 20), but can also be larger depending on the component 2 (for example, in the area of ​​thin strips, which may also become undesirable predetermined fracture locations due to anomalies). In this way, the density of anomalies in the component 2 can be determined in situ. If the anomaly density exceeds the limit value, the process can be stopped as early as possible, and the costs caused by scrap can be reduced.

[0088] Defects can also be classified with the help of time and space context. For this purpose, a neural network is used for classification, which takes the preprocessed process value (the distance between the predicted and actual value in the prediction model or the reconstruction error in the model for reconstruction) as an input variable. The sig moid function is used in the last layer to assign a probability value to each defect. If, for example, it is known that the previous layer has cooled significantly due to the large difference in the time stamps, an adhesion defect can be inferred with a higher probability. In addition, if, for example, oxidation exists in the adjacent layer, there is a higher probability that the defect is caused by the inclusion of dirt.

[0089] Reference numerals list

[0090] 2 Components

[0091] 4 Control device

[0092] 6 Sensors

[0093] 8 Abnormal

[0094] 10 Digital Twins

[0095] 12 Print head position

[0096] 14 Print head

[0097] 16 Point Cloud

[0098] 17, 17' point

[0099] 18 Work Area

[0100] 19 working points

[0101] 20 Liquid Phase

[0102] 22 Ellipsoid

[0103] 24 Robotic Arm

[0104] 26 Electrodes

[0105] 28 Welding wire

[0106] 30 Welding Camera

[0107] 32 Workbench

[0108] 34 Energy Supply Device

[0109] 36 Table feed path

[0110] P S Process Startup

[0111] 38 Process Timeline

[0112] 40 Process

[0113] 42 Anomaly Detection

[0114] 44 Separation

[0115] 46 Data Stream

[0116] T1 training data

[0117] Mg Generalized Anomaly Detection Model (ADM)

[0118] MA Adaptive Anomaly Detection Model

[0119] MS1 Dedicated Anomaly Detection Model

[0120] 48 Modeling Data

[0121] 50 Measured data

[0122] 52 Error distance

Claims

1. A method for additive manufacturing of a component (2), comprising the following steps: - create machine code, - Create a generalized anomaly detection model by training the representation learned from the training data, where The training data is derived from a dataset of sensor data from completed manufacturing processes with known process outcomes. - training the representation using training data, wherein output data are calculated for this purpose from input data, - and create an adaptive anomaly detection model - training the adaptive anomaly detection model specifically to a parameter set using available training data, -Transmitting the machine code and the generalized anomaly detection model and the adaptive anomaly detection model to the control device - Start the manufacturing process - Use sensors to monitor the process -Evaluation of sensor signals for manufacturing processes using generalized anomaly detection models - Train a dedicated anomaly detection model in parallel from the adaptive anomaly detection model using process data from an ongoing manufacturing process - detecting anomalies in component manufacturing with the help of a dedicated anomaly detection model during the manufacturing process, and - storing the adaptive anomaly detection model trained for the process parameter set as a dedicated anomaly detection model and using the dedicated anomaly detection model when the process parameter set is applied again, and - a dedicated anomaly detection model is used when starting a second manufacturing process with a second set of process parameters and is trained by the adaptive anomaly detection model, and wherein, - During the process, a digital twin (10) of the produced component (2) is constructed in parallel from the sensor data including the position data of the detected anomalies (8) - predicting the position (12) of the print head (14) at a certain time by means of a machine code, -Analysis of the work area around the location in terms of existing anomalies through digital twins -Adapt process parameters to eliminate anomalies when reaching the working area.

2. The method according to claim 1, characterized in that Anomalies identified by the anomaly detection model are added to the digital twin.

3. The method according to claim 1 or 2, characterized in that Each position (12) reached by the print head (14) is assigned a time stamp.

4. The method according to claim 1, characterized in that The digital twin (10) includes a point cloud (16), and for each point, an outlier is determined by one of a generalized anomaly detection model and an adaptive anomaly detection model, and a process state is stored for the outlier.

5. The method according to claim 4, characterized in that In order to determine the outlier, the data of the working area (18) including the adjacent points (17), (17') are also included.

6. The method according to claim 5, characterized in that The working region (18) has a spatial extent that is in a liquid phase (20) at an observation time point.

7. The method according to claim 4, characterized in that The point cloud is presented in a spatially structured data structure in the form of an octree.

8. The method according to claim 4, characterized in that The working area (18) is represented by a double ellipsoid (22).

9. The method according to claim 1, characterized in that Adaptation of process parameters to overcome anomaly (8) includes higher heat input.

10. The method according to claim 1, characterized in that Adaptation of process parameters to overcome anomaly (8) includes lower print head speed.

11. The method according to claim 1, characterized in that The additive manufacturing method is arc wire feeding additive manufacturing.

Images