Method and apparatus for monitoring a quality of an object of a 3D-print-job series of identical objects

The present invention relates to a method, apparatus as well as a computer-program-product for monitoring a quality of an object of a 3D-print-job series of identical objects, each object built from a multitude of stacked 2D-layers printed by a 3D-printer in additive manufacturing process.

Additive manufacturing or layer based manufacturing is a cutting-edge part production process that starts with a modeling phase by segmenting a 3 -dimensional (3D) object model generated by computer aided design methods into stacked 2 -dimensional (2D) layers of different thickness and

continues with the actual production phase by printing one layer upon a previous layer. In a powder bed fusion additive manufacturing process, one of several existing additive manufacturing processes, two alternating steps are performed. In a recoating step material powder is spread on a flat surface e.g. of a previously processed layer creating a so- called powder bed. In a subsequent exposure step the material powder is exposed and fused to the previous layer.

Depending on the height of the object and the layer

thickness, the number of 2D- layers needed for the object is very big, e.g. thousands of layers, and the time needed to print all the layers can take several hours to days.

Furtheron the printed objects must fulfill a very high quality level depending on the type and deployment of the object, for example in gas turbines.

Though very flexible in creating objects, the manufacturing process is affected by several issues like powder bed issues, insufficient thickness of powder in a powder bed after recoating, overheating during the melt process or

insufficient gas flow needed to blow the dust out of the printer chamber. These issues can result in severe quality defects in the completely manufactured object or it can result in the termination of the print-job, if the issue is identified during the manufacturing process. In both cases the part has to be scrapped and significant print production time and material have been wasted. Especially in print-job series, printing the same objects many times, the

manufacturing process should be optimized to identify objects of low quality as early as possible to prevent high amount of wasted material and time.

It is common practice that operators supervise the additive manufacturing process evaluating sensor data and images provided from different processing steps. This evaluation is usually based on the skill level and experience of the operator. Especially in high print-job series this evaluation is done by different operators resulting in varying

evaluation and adaption results of the additive manufacturing process based on their different skill level and experience.

DE 10 2015 011 013 Al discloses a quality assurance system for generative manufacturing which enables measurements of actual process status variables in combination with a

generative manufacturing process. Such process variables define a possible process room, in which the manufacturing process can be judged as nominal, i.e. showing sufficient quality .

WO 2015/020939A 1 discloses an automatic process control for additive manufacturing comprising at least one camera

generating layer verification images, which are compared with a predictive render of the object. Verification involves edge detection, feature recognition and other computer vision techniques to determine how close the current print is to the predictive render.

Therefore it is the object of the present invention to provide objective criteria for evaluating the quality of the objects during printing to keep a constant quality level in 3D-print-job series taking into account the quality of each single layer and also all layers already printed in a

practical and time efficient way. Additionally lower quality object shall be identified early in the printing process.

This object is solved by subject-matter according to the independent claims. Preferred embodiments are subject to the dependent claims, the description or figures.

The expression "manufacturing process" and "printing process" as well as "manufactured layer" or "printed layer" are used as synonyms throughout this document. According to the first aspect this object is solved

concretely by a method for monitoring a quality of an object of a 3D-print-job series of identical objects, built from a multitude of stacked 2D layers printed in a 3D printer in an additive manufacturing process, comprising the method steps of determining a layer quality indicator of a currently printed layer of an object, comparing the determined layer quality indicator of the currently printed layer with a predetermined lower confidence limit of the corresponding layer, the predetermined lower confidence limit being

calculated depending on layer quality indicators of

previously completely manufactured object complying with predefined quality requirements, and generating a warning signal, if the layer quality indicator of the currently printed layer has a value equal or below the lower quality limit. The layer quality indicator is determined by machine learning means .

With this method the quality of individual layers are

considered already during the manufacturing process. Layers with insufficient layer quality indicator can be identified and measures can be taken early as a warning signal indicates them instantaneously. The criteria indicating a layer as having an essential issue is provided by a layer quality indicator calculated under the same conditions during former print-jobs of identical objects having the same shape, number of layers and preferably being manufactured by the same 3D printer, for which the completely manufactured object was evaluated as compliant with predefined quality requirements.

In a preferred embodiment the lower confidence limit is determined for each object out of a subset of objects of the print-job series, by determining a layer quality indicator for each printed layer of the object, determining an object quality indicator for the completely manufactured object based on the layer quality indicators of all printed layers of the completely manufactured object and calculating a lower confidence limit for each layer depending on the layer quality indicators of those objects having an object quality indicator complying with the predefined quality requirements.

This has the advantage that each single layer indicator relies on the quality of the corresponding layer of objects being rated as "compliant objects", i.e. completely

manufactured objects which comply with predefined quality requirement. The lower confidence limit is a numerical value for example providing a probability that the layer provides a quality which leads to an sufficient overall quality of the completely manufactured object.

In a preferred embodiment at least one of the completely manufactured objects of the subset is inspected by material analytic means to confirm the calculated object quality indicator complying with the predefined quality requirements.

The confirmation of the actual quality of the completely manufactured object by material analytic means provides a reliable basis for classifying the determined sequence of layer quality indicators for all layers of the object to a "compliant" object complying with the predefined quality requirements. As bigger the number of inspected completely manufactured object is, as more reliable are the determined layer quality indicator sequences.

In a preferred embodiment the lower confidence limit is recalculated taking into account the layer quality indicators of each further completed manufactured object with an object quality indicator complying with the predefined quality requirements . This has the advantage that also the determination of the lower confidence limit is based on a growing number of layer quality indicators of compliant completely manufactured objects and can therefore be refined with each additional completely manufactured object complying with the predefined quality requirements.

In a preferred embodiment an upper confidence limit and the lower confidence limit are provided by a standard deviation calculated depending on the layer quality indicator for each layer of completely manufactured objects complying with the predefined quality requirements of the object.

This means that the standard deviation is calculated of all or at least a subset of layer quality indicator sequences of completely manufactured objects complying with the predefined quality requirements, i.e. compliant objects.

The object quality indicator may be determined by machine learning means .

Machine learning means are algorithms based on data

classified by learning algorithms wherein the learning algorithms are based on training data which is already labeled as belonging to one of the classifications assigned in the classification algorithm. Using machine learning means has the advantage that the quality indicators can be

determined automatically without an evaluation by an

operator, preferably after the machine learning algorithm has been successfully performed a training phase. In this

training phase for a subset of at least more than one objects of the print-job series the layer quality indicator is determined for each layer resulting in a probability value for each layer quality. Additionally after finishing the print-job the completely manufactured object is examined with respect to its quality. An completely manufactured object is labeled as "compliant object", if complying with predefined quality requirements or it is labeled as "not compliant" object which does not comply with the predefined quality requirements. The lower and upper confidence limit are determined depending on all layer quality indicator sequences of completely manufactured objects labeled as "compliant" objects. Preferably completely manufactured object of the subset used for training the machine learning algorithm are inspected by material analytic means to ascertain the quality of the completely manufactured object.

In a further embodiment at least one fault detection and prediction value is determined for each layer, each fault detection and prediction value depending on data of different data sources of the present additive manufacturing process and determining the layer quality indicator depending on all determined fault detection and prediction results.

It is the advantage that the layer quality indicator is based on data from different data sources representing different processing conditions and layer conditions. Therefore

different influencing parameters to the manufacturing process are considered. The detection and predication value depends on data of different data sources which can be weighted or interconnected to each other representing the manufacturing process with different aspects. In a preferred embodiment data sources are at least one of a computer aided design model of the object, images of the additive manufacturing process, sensor data of the additive manufacturing process or settings of the additive manufacturing process.

Settings means the values of different components of a 3D- printer or the additive manufacturing process, e.g. settings influencing the amount of powder spread in the powder bed for building the next layer, the temperature setting of e.g. the meltpool, gas pressure settings and so on. In a preferred embodiment a warning signal is generated, if layer quality indicators of subsequent layers show a common trend towards the lower confidence limit.

This has the advantage that already before the lower

confidence limit is reached countermeasures can be started to raise the layer quality of subsequent layers or to stop the complete printing process and scrap the partly printed object as early as possible. In a preferred embodiment the settings of the additive manufacturing process are automatically adjusted based on the generating warning signal .

This allows deriving measures based on a generated warning signal. Settings of the printing process can be changed e.g. automatically based on a warning signal.

A second aspect of the invention is realized by an apparatus for monitoring a quality of an object of a 3D-print-job series of identical objects, each object built from a

multitude of stacked 2D-layers printed by a 3D-printer in an additive manufacturing process, comprising at least one first processor configured to determine a layer quality indicator of a currently printed layer of an object, and

comparing the determined layer quality indicator of the currently printed layer with a predetermined lower confidence limit of the layer, the predetermined lower confidence limit being calculated depending on layer quality indicators of previously completely manufactured objects complying with predefined quality requirements, and a signal generating unit configured to generate a warning signal, if the layer quality indicator of the currently printed layer has a value equal or lower than the lower quality limit.

The apparatus provides the advantage that an additive

manufacturing process of a larger print-job series can be monitored automatically requiring layer quality indicator sequences of all layers of only a few completely manufactured objects which comply with the predefined quality

requirements. These predefined quality requirements can be proven by inspecting a completely manufactured object based by material inspection means, especially non-destructive material inspection means.

In a preferred embodiment the apparatus comprises at least a second processor which is configured to determine the lower confidence limit as described for the method. In a preferred embodiment the apparatus comprises an input unit configured to receive a confirmed object quality

indicator of at least one of the completely manufactured objects of the subset from the material analytic means and/or to receive data of at least one of a computer aided design (CAD) model of the object, image data of the additive

manufacturing process, sensor data and settings of the additive manufacturing process.

In a preferred embodiment the signal generating unit is configured to generate a warning signal, if layer quality indicators of subsequent layers show a common trend towards the lower confidence limit.

In a preferred embodiment the apparatus comprises an output unit configured to transmit warning signals to the 3D- printer . A third aspect of the present invention concerns a Computer- program-product with program-code-means stored on a machine readable data carrier to perform all steps of the method, when the program is performed on a programmable computer or a digital signal processor.

Embodiments of the method and apparatus are illustrated in the figures which are described below:

Figure 1 shows a flow diagram of an embodiment of the

inventive method;

Figure 2 shows a schematical diagram of a hierarchy and

systematic determination of the layer quality indictors ;

Figure 3 shows the second embodiment of the inventive method with training phase;

Figure 4 shows an example of layer quality indicators for all layers of an object of the print-job series;

Figure 5 shows an upper and a lower confidence limit for all layers of an object of the print- job series.

Figure 6 shows an embodiment of a layer quality indicator determined for a currently printed object in relation to the upper and lower confidence limit previously determined;

Figure 7 shows an embodiment of all layer quality indicators of a completely manufactured object with a trend analysis evaluation; and

Figure 8 shows an embodiment of the inventive apparatus

connected to a 3D-printer in schematical view. Additive Manufacturing is a production process wherein 3- dimensional (3D) objects are produces by adding and fusing layers of material on top of each other, in contrast to common mechanical production like e.g. milling. The process is generally called additive manufacturing process or 3D- printing process. The production is based on digital 3D- models of the physical object and is therefore computer controlled. This so called computer-aided design, short CAD, is vector-based, since all 3D-objects can be characterized by using lines and points. A digital 3D-model, which is saved in a CAD-file, is converted into thin slices, the so called layers. The layer data are sent to the 3D-printer, which prints layer by layer on top of each other and that

manufactures the 3D-object.

The general idea of the monitoring method, shown in Figure 1, is to determine a layer quality indicator QI(Li), see step SI, which is calculated for each printed layer Li of a given CAD-model of an object automatically during the current printing print-job and compare it to a predefined lower confidence limit LCL (Li) of the corresponding layer Li, see step S2, to allow early measures in the printing process by generating a warning signal, if the current layer quality indicator is equal or below a lower confidence limit, see step S3. The layer quality indicator is determined by the same machine learning functions and is used during the training phase SO. The predetermined lower confidence limit LCL (Li) is calculated, see step SO, depending on layer quality indicators determined for all layers of previously completely manufactured objects complying with predefined quality requirements.

The training process step SO is shown in detail in Figure 3 and explained referring to a object 300 schematically

depicted in Figure 7. During the training step the lower confidence limit LCL is determined for each object out of a subset of preferably more than one objects of the print-job series. For each printed layer Ll,..,Ln of an object of this subset a layer quality indicator QI(L1), ..QI(Ln) is

determined, see step LSI. Subsequently an object quality indicator is determined for the completely manufactured object based on the layer quality indicators QI(L1), ..QI(Ln) of all printed layers (Ll,..,Ln) of the completely

manufactured object, see step LS2. An upper confidence limit UCL (LI ) , .. , UCL (Ln) and a lower confidence limit LCL(Ll),.., LCL(Ln) is calculated for each layer Ll,..,Ln depending on the layer quality indicators of those objects having an object quality indicator complying with predefined quality requirements, see step LS3.

The method ends in step S4, if all layers LI, .. , Ln of the object 300 are printed and layer quality indicators QI(L1), ..QI(Ln) are determined and evaluated. The method ends in step S5, if the print-job of the object is stopped before all layers Ll,..,Ln are printed, e.g. as a consequence of a previous warning signal transmitted to the 3D-printer. The printing process of each further object of a 3D-print-job series corresponding to the same CAD-model can now be

evaluated in comparison to the previously determined

sequences of layer quality indicators determined either in a dedicated training phase or determined continuously taking into account further sequence of layer quality indicators of a completely manufactured compliant object complying to predefined quality requirements of the object.

The layer quality indicator is determined in a two step approach, as shown in Figure 2, by machine learning

techniques, e.g. random forests, neural networks, Markov- Models or Gaussian classification. At least one, preferably several, detection and prediction values DPI, DP2 , DP3 are calculated by a machine learning method based on different data sources. Data sources are e.g. the CAD-model of the object DS1, images from a power bed camera D2 , time line data DS3 of sensors e.g. detecting a temperature and pressure of a melt pool or further process parameters D4 of the printing process. The data DS1, DS2 , DS3 and DS4 of different data sources are also called multimodal data. Examples for

multimodalities are powder-bed images, Meltpool images, time series sensor data and CAD data. The detection and prediction value is an indicator that a specific layer causes issues for the overall printed object. The usage of different multimodal data increases the reliability for detection of different type of issues or errors during the manufacturing process. In a second step the detection and prediction results DPI, DP2 , DP3 are aggregated into one value which is the layer quality indicator QI .

The layer quality indicator is a numerical value providing a probability for a layer contributing to a completely

manufactured object complying with a predefined required quality. The numerical value can provide a percentage in the range between 0 and 100 or can be between the value 0 to 1 or normalized to any other interval. The layer quality indicator QI(Li) can be aggregated depending on the single detection and prediction values DPI (Li) , DP2 (Li) , DP3 (Li) of the layer Li, e.g. by taking a minimum value of all detection and prediction values DPI (Li) , DP2 (Li) , DP3 (Li) or by multiplying weighted detection prediction values or an addition of weighted detection and prediction results, e.g. providing an aggregated value of all detection and prediction values

DPI (Li) , DP2(Li) , DP3(Li) .

A detection and prediction value DP (Li) as well as the layer quality indicator QI(Li) shown in Figure 2 as a graph are defined by the equation:

DP _Li = f (DSl _L1 , DS2 _Li , ... , DSA _Ln )

The layer quality indicator of a layer Li is

QI (Li) = f iDPl _Li , ... , DP3 _Li ) . The detection and prediction values DPI, DP3 and the layer quality indicator QI are calculated for each layer of an object of the print-job series. Figure 4 shows a sequence 10 of layer quality indicators for all layers Ll,..,Ln one object 300 of the 3D-print-job series. For each of the layers Ll,..,Ln one layer quality indicator QI(L1), ..QI(Ln) is determined and shown as a function of the layers L of an object. As an example the layer quality indicator QI(Li) of a Layer Li is marked in Figure 4.

A graph of layer quality indicator QI(L1), ..QI(Ln) of all layers ordered by consecutive layers, also called a sequence of layer quality indicators and shown in figure 4, can be evaluated for the first printed object of a given CAD-model and every following object in new print-job of this print-job series. During the training phase a subset of objects of the print-job series these layer quality indicator is determined for each printed layer of an object. For the completely manufactured object an object quality indicator is determined based on the layer quality indicator of all printed layers after a print-job is finished. Additionally for the

completely manufactured objects of the subset an evaluation of the quality of the object is performed preferably by nondestructive inspection means to evaluate if the object needs the required features and therefore the required quality. If this is the case the sequence of layer quality indicator 10 as shown in figure 4 is used for calculating an upper and lower confidence limit.

For all these layer quality indicator sequences 10 of objects complying with the quality requirements a standard deviation of the quality indicator is calculated. E.g. the upper and lower quality limit is evaluated by adding or subtracting one or several times the standard deviation to the mean value of all considered layer quality indicator graphs for each individual layer. If a determined layer quality indicator QI(Li) for layer Li lies in a value range between the upper confidence layer UCL and the lower confidence limit LCL, the quality of the considered layer leads with a certain

confidence to an completely manufactured object complying with the quality requirements.

The more objects of the 3D-print-job series are used for training the learning function to evaluate the detection and prediction results as higher is the probability that the detection and prediction value DPI, DP2 , DP3 is correct.

Also after the training phase the upper and lower confidence limit, UCL and LCL, can be refined by evaluating a mean layer quality on all layer quality indicator sequences so far determined complying with the predefined quality

requirements. This means that each layer quality indicator sequence of a completely manufactured object within

determined object quality indicator complying to the quality requirements is included into the sample of considered layer quality indicator sequences. Accordingly the upper confidence limit UCL and the lower confidence limit LCL can be

recalculated depending on the new number of layer quality indicator sequences complying to the predefined quality requirements . During a current printing process of a layer Li the layer quality indicator QI(Li) is determined according to step SI as shown in figure 1 and figure 3. In a further step S2 the determined layer quality indicator QI(Li) of the currently printed layer is compared with a predetermined lower

confidence layer LCL (Li) of the corresponding layer Li. If the layer quality indicator QI(Li) is above the lower

confidence limit LCL the value is stored and the same process is performed for the next layer. If the layer quality

indicator is equal or below the lower confidence limit a warning signal is generated, see step S3, and transmitted to the additive manufacturing process. Such a sequence of quality indicators 11 is shown in figure 6. Here the layer quality indicator QI (LI , .. , Li - 1 ) of the layers Ll,..,Li-l, marked with reference sign 11, show the value between the upper and the lower confidence layer UCL, LCL . At layer Li the layer quality indicator QI(li) is lower than the lower confidence layer LCL (Li) and accordingly a warning signal is generated.

In case the determined layer quality indicator Li shows a value above the upper confidence layer a warning signal can also be generated. In this case the warning signal can be used to adapt the settings of the manufacturing process e.g. by lowering the amount of material powder for this layer as this high confident layer indicates that the present settings for the layer can be optimized e.g. to save material. Further a trend analysis can be performed taking into account not only the layer quality indicator QI(Li) of the currently printed layer Li, but also a subset of preceding layer quality indicators 12 of preceding layers, see figure 7. If the layer quality indicators of the subset of layers already show a gradient value below a certain value a warning signal will also be generated.

The warning value can trigger different actions. E.g.

parameter settings of the additive manufacturing process can be adjusted. Especially early warnings from a trend analysis may lead to an adaption of the setting of the 3D-printer. This can lead to a healing of an issue occurred in layer Li by subsequent layers printed with adapted settings. Therefore a self-healing process without any interaction of an operator can be achieved. The layer quality is continuously monitored in this way until the last layer of the object is printed and a completely manufactured object is created. For each of these completely manufactured objects the sequence of layer quality indicator values is determined and stored.

Besides an automatic adaption of settings, the warning signal can be sent to a supervision monitor to inform an operator, that the print process for each a boundary conditions if no interaction will be started.

A monitoring apparatus performing the described method is shown in figure 8. The monitoring apparatus 100 comprises a first processor 110 configured to determine a layer quality indicator QI of a currently printed layer. The first

processor 110 is also configured to compare the determined layer quality indicator QI(Li) of the currently printed layer Li with a predetermined lower confidence limit LCL of the corresponding layer evaluated during the training phase SO . The monitoring apparatus comprises further at least a second processor 130 configured to perform the training phase, which is to perform the steps LSI, LS2 and LS3 of the training phase as serial as shown in figure 3.

Further on a signal generating unit 120 is included and configured to generate the warning signal . Further on an input unit 140 is configured to receive data from the

different data sources, e.g. a camera 310 or a sensor 320 of a 3D-printer 200. The input unit 140 is also configured to receive data from a material analytic means 400 which provide measured information on the object quality indicator based on preferably a non-destructive material analytic method. The material inspection means 400 may provide information to the monitoring apparatus 100 via the 3D-printer 200 or it might be connected directly to monitoring apparatus 100. The 3D- printer 200 shows schematically printed layer LI,.., Ln, which form a completely manufactured object 300.

All methods and method steps can be implemented by

corresponding means which are adapted for performing the respective method steps. All functions provided by particular means can be a method step of the method.

The scope of protection is given by the claims and not restricted by features discussed and the description as shown in the figures. The invention is not limited to the described example. The invention also comprises all combinations of any of the described or depicted features.