Die bekanntesten Einsatzgebiete von künstlicher Intelligenz (KI) sind zurzeit Bild- und Sprachverarbeitungen. Im Gegensatz dazu liegt der Fokus hier auf dem Lernen komplexer Zusammenhänge in Produktionsprozessen: untersucht wird ein KI-Anwendungsfall, in dem die Qualität des Endproduktes basierend auf Prozessdaten, die während der Produktion erhoben werden, vorhergesagt wird. Mit Hilfe Neuronaler Netze wird die Belastbarkeit von additiv gefertigten Bauteilen in Echtzeit bewertet. Dazu werden Imperfektionen, wie z. B. Lufteinschlüsse im Bauteil, berücksichtigt, da diese einen erheblichen Einfluss auf die Qualität des Bauteils haben. Das heißt, basierend auf Prozessdaten des additiven Fertigungsprozesses werden potenzielle Imperfektionen im Bauteil detektiert und in eine dreidimensionale Repräsentation des Bauteils eingefügt. Mit Hilfe des trainierten Neuronalen Netzes werden dann Festigkeitskennwerte für das Bauteil bestimmt, die Aufschluss über die Qualität geben.
Zum Trainieren eines Neuronalen Netzes ist ein großer Trainingsdatensatz notwendig. Diese Daten für diesen Anwendungsfall experimentell zu generieren würde bedeuten, dass sehr viele Bauteile mit Imperfektionen gefertigt und anschließend geprüft werden müssten. Um diesen zeitaufwendigen und unwirtschaftlichen Prozess zu umgehen, werden Finite-Elemente-Simulationen zur Erstellung der Datenbasis genutzt. Das heißt, es werden Finite-Elemente-Modelle des Bauteils erzeugt, in die Imperfektionen künstlich eingebracht werden. Mit numerischen Simulationen können anschließend die Festigkeitskennwerte des Bauteils ermittelt werden. Da diese Simulationen zeitaufwändig sind, ist eine Echtzeit-Anwendung nicht möglich. Stattdessen trainieren wir mit den Ergebnissen der Simulationen ein Neuronales Netz, dass die Simulationsergebnisse vorhersagen soll. Ein solches Neuronales Netz kann dann während des Produktionsprozesses genutzt werden, um in Echtzeit die Auswirkungen der detektierten Imperfektionen auf die Qualität zu beurteilen. Dadurch können mangelhafte Bauteile aussortiert und eine gleichbleibende Qualität gewährleistet werden. Es ist sogar möglich, den Druck mangelhafter Bauteile abzubrechen, was Zeit und Ressourcen spart.
1 Introduction
Machine learning and artificial intelligence (AI) applications have evolved significantly in the past few years. One field of application is finding relations in large datasets and learning the underlying unknown complex mathematical functions. In this paper, we present an AI-based approach that enables quality assurance in an additive manufacturing (AM) process. AM, or 3D printing, is a category of production processes where components are created layer wise. It enables the production of highly complex components of any geometry and even the design of internal structures such as cooling channels or injection nozzles. This design freedom allows to produce lightweight and cost-effective components, which makes this technology very attractive for the aerospace industry. Since AM is a relatively new technology, there is much less experience compared to traditional manufacturing processes, e.g., in terms of the impact of disturbances in the production process on the component quality. Consequently, in order to be allowed to install an AM component, e.g., in an aircraft, an approval process must be passed consisting of a lot of destructive and non-destructive tests. More precisely, destructive tests are required to verify general material properties and the strength of a component. Further, non-destructive testing is required to examine each component individually and, if necessary, to initiate post-processing steps for quality assurance. For example, a computer tomography (CT) scan is used to detect imperfections. Unfortunately, these needed tests and post-processing steps are very time-consuming and expensive. To reduce this disadvantage of the AM processes, we will show in this paper how production data obtained during the printing process can be used to assure component quality which will decrease the demand of tests and make the approval process more efficient.
2 Technology Background and State of the Art
In this paper, we specifically refer to the AM process laser powder bed fusion (LPBF), where components are produced layer wise by melting a fine-grained metal powder with a high-power laser beam: A three-dimensional component design is divided into several layers of a specific thickness in z-direction. For each layer, a uniform powder bed is created by the recoating process. The laser beam is directed over this powder bed with special scanning mirrors to scan the component areas in the current layer, where it melts the powder into a solid metal. The quality of AM components depends mainly on the metal powder used and on the printing parameters such as laser power, hedge distance, laser speed or layer thickness. However, local defects can occur even with well-chosen parameter settings, cf. [1]. The resulting imperfections have been shown to cause a negative change in the mechanical structural properties of the final component compared to the properties of an optimal component. For example, porosity contained in the AM component, caused by keyholing or lack of fusion, affects dynamic and static properties and can cause anisotropic material behavior, see [2, 3]. Furthermore, the surface roughness of AM components particularly affects the fatigue properties, cf. [4]. Consequently, comprehensive quality assurance during the production process is essential for the outcome of the AM process. To achieve this, monitoring systems are used to collect information about the manufacturing process from which production-related component defects can be derived. The challenge is to determine the influence of defects on the properties of the manufactured components. For this purpose, we present an approach to evaluate the influence of defects to the component quality in real-time using AI.
Anzeige
A crucial point in AM using LPBF is the stable melting of the metal powder. As shown, e.g., in [5], the temperature in the melt pool has a significant influence on the microstructure of the AM components. For this reason, various monitoring systems have been developed to observe the melt pool and several commercial suppliers of 3D printers offer monitoring systems for the melt pool or the thermal exposure, such as EOS, SLM Solutions, TruPrint, Renishaw and Sigma Additive Solutions.
We consider in this paper the optical tomography described in [6]. With this monitoring approach, the radiation during the LPBF process is recorded by an off-axial camera system, i.e., the camera does not follow the optical path of the fusion laser. The light omitted during the melting process consist of three components: reflected laser radiation, plasma emission and thermal radiation. A suitable band-pass filter is used to record only the thermal radiation. The layer-wise recorded data is visualized color-coded to show hot-spots and cold-spots. An analysis of AM test specimens has shown a correlation between the extreme temperature of spots and material failures.
Process data generation, i.e., the monitoring of the printing process, has already found its way into the industry, whereas the automatic detection of imperfections and defects is still the subject of current research. To provide an overview of this ongoing research activities, we state in the following a few examples where AI methodologies have been applied to identify anomalies in the AM process.
In [7], an application of supervised machine learning to detect defects is demonstrated. The approach is based on images taken for each layer with different lighting settings. Features for each location are extracted from these images and classified into anomalous and nominal. In this way, defects in the AM components can be detected and localized. A support vector machine is used for the classification. To generate the required reference data for supervised learning, anomalies of an AM component are detected from CT scans and the associated monitoring images are labelled.
Anzeige
In [8], three machine learning approaches are applied to detect hot spots based on high-speed videos of the printing process. Aim is, to classify between normal and defect-related brightness evolution. For this purpose, features were extracted from the videos, e.g., the mean gradient of the brightness in successive images, the maximum mean brightness drop between two consecutive frames, and the shape and size of the bright region. Then, these features are used to classify the video snippets by a support vector machine or by a neural network. Additionally, k-means is applied to cluster the development of the mean brightness. All three methods have shown an accuracy of more than 90 %.
However, the research results just listed are always only a detection of defects or anomalies. An evaluation of their impact on the component’s quality is missing. Our vision is to evaluate the component’s quality in real time based on the detected defects applying machine learning approaches. For this purpose, we need for many components information about their defects and their quality in terms of structural properties. To determine the structural properties of a component, it must be proven by destroying tests. For example, in tensile testing the component is slowly extended until it breaks. With it, e.g., the maximal force that the components can withstand is determined. However, this experimental approach requires to produce and destroy a large set of components to generate a representative database for training AI methods. This is very expensive and time-consuming and practical not feasible. To reduce this data generation effort, it was investigated, in a more general application context, whether numerical simulations can be used to create a representative training database. In particular, in [9], it is shown how deep neural networks can be used to estimate biaxial stress-strain curves of sheet metals. Based on synthetic generated crystallographic texture data, biaxial stress-strain curves were obtained from crystal plasticity-based numerical biaxial tensile tests obtained with finite elements (FE) simulations. With these numerical test data, deep neural networks were trained. It is shown that their predictions are close to the numerical computed test results. However, it is important to create a well-grounded data base to applicate the approach to real sheet metals.
In the following, we will transfer this idea to the AM production process. We present an approach that uses numerical simulations to generate a representative training database for AI methods. In detail, we developed an enhanced FE simulation that considers the material properties of AM components and allows to evaluate the influence of imperfections on the component quality. Note that the FE simulations itself could not be applied in a real-time quality assurance process because of the large computational time needed for one simulation.
3 Approach
The monitoring systems currently in use help to manually detect anomalies in the melting process which indicate the formation of imperfections in AM components. As mentioned above, several commercial 3D printers have integrated monitoring systems to record the melt pool, or the temperatures of the heat affected zone. These systems are an important enabler to detect imperfections. To gain greater benefit from monitoring, the detected imperfections must be analyzed in terms of their potential impact on component quality. The approach presented here provides such an evaluation.
Currently, the application of numerical simulation methods, like the FE method, is being investigated to determine structural properties of defect-afflicted components. In general, numerical simulation methods are based on a discretization of the component to numerically solve the partial differential equations that describe the behavior of a component under structural load. However, to obtain reliable results, a fine discretization is required. For this reason, these numerical methods are computationally expensive and cannot be used for a real-time service.
Another investigated method for monitoring and quality assurance of AM is machine learning. One problem of the application of AI methodologies is that usually a large number of data sets is required to train an AI. For AM applications, this means that many parts must be printed, especially parts with defects, to obtain a representative data basis. In other words, to train an AI that learns the relationship between defect position and structural properties, information about the component condition as a result of the manufacturing process (e.g., presence of imperfections) as well as the mechanical properties of the components are required. To obtain the mechanical properties experimentally, destructive tests must be carried out. Therefore, specimens or components have to be manufactured and manually tested one by one. This effort, i.e., manufacturing and testing a large number of components to obtain a sufficient data set to apply an AI method, is not economically feasible.
In this paper, we combine these two approaches. The use of component-related and validated numerical simulation methods offers an alternative to the expensive production and testing step. In our approach, the training data is generated by means of FE simulations which compute the mechanical properties of components with imperfections. The data obtained will be used to train a neural network that can predict mechanical properties in real time based on the detected imperfection.
This approach is illustrated in Fig. 1 and consists of two steps: training a neural network to estimate the quality of AM components (blue colored) and an online application of the trained model during the printing process to predict component strength (orange colored).
Fig. 1
Concept for AI based quality assurance in AM
×
3.1 Training of an AI Model
As it can be seen in Fig. 1, developing a neural network for quality prediction of AM components consists of three sub steps: virtual testing, generating training data and training itself. Furthermore, experimental tests are used to verify the results of the virtual tests.
Numerical simulation methods, i.e. FE simulations, are applied for virtual testing of a large set of components with various imperfections scenarios. In detail, the design of the component is discretized by defining finite elements, e.g., hexahedral volume elements. Based on this discretization, the deformation and stress under load are calculated by the numerical solution of partial differential equations. For this purpose, the material characteristics must be known, which can be obtained from experimental tests. One possibility to compute the effect of pores in the simulation is to define a FE model with pores by generating very small finite elements around the pore. However, since the pores in AM components are very small, with diameters around 10 µm [12], a very fine discretization is needed to represent pores which would lead to an extremely high computational effort. To reduce the overall computational effort, a multi-scale method was used [10, 11]. For this purpose, representative volume elements are defined for a small piece of material with a pore. The corresponding material characteristics are computed by preliminary simulations based on very small finite elements. These representative volume elements are integrated in the FE model of the construction with larger finite elements and represent the imperfections. Subsequently, numerical simulations are used to predict structural properties and failure behavior of the virtually generated AM components with imperfections. However, due to the inhomogeneous material properties the computational effort is still high. With our developed FE model generator, the introduction of the imperfections as well as the generation of the simulation models can be easily realized. This offers the possibility to generate component models with arbitrary and specifically introduced imperfection states and to simulate their effects on the structural properties of the component.
To train a neural network, the FE models with imperfections must be provided as training data in combination with the virtual structural test results obtained from the simulations. To transform the FE models into input data for a supervised machine learning algorithm, the hexahedral volume elements have to be linked to pixels of a 3D image of the component design. This step is needed, because the elements of the FE model have different sizes and shapes. Furthermore, they are described by their corners, which is not intuitive for finding relations between geometrical located defects and structural properties. In the generated 3D image, the pixels are associated with material properties to represent the FE model. For elements affected by defects, the material properties of the corresponding representative volume element are used for the corresponding pixels. A supervised learning approach is used, where the 3D images extracted from the FE models with imperfections are the inputs and the calculated structural properties of the FE simulation are the outputs. With these input and output data, a convolutional neural network was trained to represent the influence of imperfections to the structural properties.
In further research, experimental testing will be used to validate the virtual testing methodology. Note, that this validation requires much less experimental tests than generating training data for an AI only with those tests.
3.2 Online Application of the AI to Predict the Impact of Imperfections in Additive Manufactured Components
As indicated by the orange boxes in Fig. 1, after training a neural network, it can be used in the online quality monitoring. This step is also divided into three sub steps: Recording monitoring data, detection of anomalies by analyzing monitoring data, and estimation of the impact of the detected imperfections by applying the trained neural network to predict component’s quality in terms of structural properties.
During the printing process, monitoring data is recorded, e.g., layer-wise monitoring images, which provide information about the temperatures in the heat affected zone, as shown in [6]. The recorded monitoring images have to be analyzed to detect anomalies in the printing process which are indications for imperfections like pores. Currently, different imaging methods are implemented in the 3D printers, but the monitoring data is often not analyzed automatically in real-time. Therefore, an image analysis tool needs to be developed to detect anomalies in the provided monitoring images. This can be either an AI, e.g., another neural network for image classification, or an image processing algorithm. After the automated detection, the anomalies must be localized in the printed components.
Based on the construction, a 3D image of the component is created and the detected imperfections are localized on the image. Based on this input data, the trained neural network predicts the structural properties. If the prediction does not meet the requirements, the AM component can be marked as insufficient or even the printing process can be aborted, which saves time and energy.
4 Feasibility Study
An initial feasibility study with tensile specimens indicated that the developed concept can be successfully implemented. First, training data for the machine learning step was generated using an FE model of a flat tensile specimen, as shown in Fig. 2. The hexahedral elements of the FE model are illustrated by the segmentation of the specimen. As it can be seen in Fig. 2, the elements in the examination area, where the specimen is thinner and the failure will occur, are significantly smaller than in the shoulders, where the specimen is gripped by the testing machine. Each element of the FE model is afflicted with material properties, which are amongst others Elastic Modulus, breaking point and Poisson’s ratio. The elements with pores have different material properties because a pore reduces the strength. The example of Fig. 2 shows a FE model with two sizes of pores coloured in red and green.
Fig. 2
FE model of a tensile specimen with imperfections (red and green)
×
In the FE simulation, a virtual tensile test is done whereby the material properties of the elements affect the load distribution during the test and with it the strength of the whole specimen. One result of the virtual test is the Elastic Modulus for the specimen which shows the resistance to being elastically deformed in case of stress. For our initial feasibility study, we have used only this result of the FE simulation.
To train a neural network for the relation between pores in a specimen and Elastic Modulus of this specimen, a representative database is required. To generate the training data, several scenarios with different kinds of pores and distributions of pores are assumed, i.e., a small set of large pores, several small pores, or pores only in the shoulders of the test specimen. For these different scenarios, FE models with randomly distributed pores are generated and virtually tested. With it, the influence of the size and the distribution of the pores on the strength of the specimen is represented in the training data.
As mentioned above, for the neural network, the FE model is transformed into a 3D image of the specimen, as shown in Fig. 3, to serve as input data. However, each pixel represents the physical characterization of the corresponding element of the FE model. In Fig. 3, the dark blue coloured pixels show the material properties of perfect printed material, whereas the yellow-coloured pixels represent areas with imperfections and consequently with weaker material properties. In this way, the size difference between the finite elements can be taken into account and the localisation of pores as well as the closeness to other pores is represented.
Fig. 3
3D-image representation of FE model
×
These 3D images and the simulation results, i.e., the Elastic Modulus, which measures the stiffness of a material, are the training data. Thereby, the images are the input, and the Elastic Modulus is the output, because the neural network should learn, how pores in the AM component affect this physical property of the component. Due to the input is given as 3D-image, we used 3D convolutional layers for our neural network. After testing several training architectures and hyperparameters, a suitable neural network was obtained. The final architecture of the neural network consists of several successive units of one 3D-convolutional layer, one max-pooling layer and one batch normalization layer; a layer to flatten the obtained feature tensor; and a dense layer with 5 nodes as last hidden layer. With it, about 4000 trainable parameters must be determined. Furthermore, our investigations showed that normalization of the input data to [0,1] was more suitable than a standardization. For the training, we had about 350 simulation results. To enlarge the dataset, the specimens are rotated and mirrored, which will not affect the simulation results but results in more training data.
Figure 4 visualizes the performance of the trained neural network. On the left, the prediction and the given output data are compared. Therefore, the data set is sorted by the output value, i. e. the Elastic Modulus computed by the FE simulation (black line); the orange dots show the corresponding predictions. As it can be seen, the predictions are close to the given output values and the general influence of the imperfections seems to be represented. On the right of Fig. 4, the distribution of the relative error, i.e., the difference between the prediction and the given output value in relation to the given output value, is shown. The relative error is generally small. In detail, the median of the relative error is −0.001 % which is close to zero. And for 95 % of the data points, the relative error was less than 0.21 %. The maximal relative error was 0.83 %.
Fig. 4
Visualization of neural network’s performance
×
5 Conclusions
In this paper, we presented an approach to apply AI in quality assurance of an AM process. Thereby, complex FE simulations were used to evaluate the effect of imperfections (i.e., pores) on structural properties of an AM component. To make this important information available in real time, the two disciplines FE simulation and machine learning have been combined. More specifically, a neural network is trained to represent the relations between imperfections that occur during production and the structural properties of the component. This neural network is used as predictor to estimate online the impact of the detected anomalies on the robustness of the component.
In the future, the data base will be expanded to include more complex component structures. Here, too, validated simulation tests will be used to determine the influence of pore frequencies and positions in the component on its quality. With the increasing data base, it is expected that not only correlations between pores and quality in concrete components can be predicted. With more complex neural network structures, more general relationships between, e.g., distances of pores to outer walls, pore sizes and wall thicknesses should also be learned.
We see great potential in this approach for improving the use of AM components. The experience gap compared to conventional manufacturing processes is reduced by data-driven approaches. Furthermore, since the approach can be validated by real material and component testing, it is possible to reduce the testing effort for approval in the future. In this case, significant material and cost savings can be expected.
The AM process itself can also be improved using this approach. For example, parameters can be adjusted or an AM process can be stopped if the AI model predicts that the AM component would not pass the approval process.
Acknowledgements
We acknowledge the financial support from the German Federal Ministry for Economic Affairs and Climate Action for the projects AMCOCS (Additive Manufactured Component Certification Services) and CertiFlight (End-to-end digital quality assurance for innovative approval processes based additive manufacturing technologies).
Open Access Dieses Kapitel wird unter der Creative Commons Namensnennung 4.0 International Lizenz (http://creativecommons.org/licenses/by/4.0/deed.de) veröffentlicht, welche die Nutzung, Vervielfältigung, Bearbeitung, Verbreitung und Wiedergabe in jeglichem Medium und Format erlaubt, sofern Sie den/die ursprünglichen Autor(en) und die Quelle ordnungsgemäß nennen, einen Link zur Creative Commons Lizenz beifügen und angeben, ob Änderungen vorgenommen wurden.
Die in diesem Kapitel enthaltenen Bilder und sonstiges Drittmaterial unterliegen ebenfalls der genannten Creative Commons Lizenz, sofern sich aus der Abbildungslegende nichts anderes ergibt. Sofern das betreffende Material nicht unter der genannten Creative Commons Lizenz steht und die betreffende Handlung nicht nach gesetzlichen Vorschriften erlaubt ist, ist für die oben aufgeführten Weiterverwendungen des Materials die Einwilligung des jeweiligen Rechteinhabers einzuholen.
