Cyber-physical approach toward semiautonomous postprocessing of additive manufactured parts and components

Characteristic features of additive manufacturing (AM) techniques, especially powder bed-based processes, comprise the ability to manufacture highly complex geometries,¹ also known as “complexity for free,”² and the possibility to integrate a variety of functionalities immediately during the manufacturing of a part.³ Pradel et al.,^4,5 on the other hand, relativize this perspective claiming the lack of empirical studies regarding the implications of geometrical complexity on AM build time and cost. Moreover, it must be stated that, regardless the extensive research activities around the world, there are only limited examples that have demonstrated the full knowledge of the additive process chain, which led to metallic products of the highest quality. Examples are the fuel nozzles for the LEAP engine from GE Aviation,⁶ the bionic bracket from Airbus,⁷ and the well-known acetabular cups, where by the end of 2016 about 2% of the global production was manufactured using ARCAM EBM Hardware while more than 50 000 EBM Ti-6Al-4V cups had already been implanted.⁸ Either way, these examples have complex geometries in common (cf. Fig. 1).

Moreover, there is a broad consensus in the literature concerning the economic production of small quantities or even lot size one, on-demand manufacture and pronounced geometric freedom (Table I).

On the other hand, looking at the certification of processes and applications, e.g., in the aircraft industry, there are still great challenges such as instability of the technology, inherent defects and anomalies, and limited precision and resolution.²⁰ Consequently, Frazier stated the need for real-time, closed-loop process controls and sensors in order to ensure quality, consistency, and reproducibility across AM machines.²¹ 

Dowling et al.²² demand an overall equipment effectiveness greater than 70% (scrap rates less than 1000 ppm) referring to a measured level of around 30% in 2015,²³ identifying the reproducibility of parts as the most critical factor. Relevant reasons are the microstructural and mechanical heterogeneity,^24–26 steep residual stress gradients that frequently result in part distortion^27,28 (Fig. 2), and the staircase effect that also drastically dependents on part orientation.²⁹ 

Hence, with regard to postprocessing it can be concluded that variable surface conditions and metallurgical variations obviously are going to influence the tool-workpiece interaction while, on the other hand, the machining strategy itself can influence the deformation of the component. The latter, however, becomes clear if considering remaining residual stresses in combination with a local change in cross section due to subtractive material removal (cf. Fig. 2). Expanding these aspects by geometrical complexity and small batch production, postmachining transforms into a strong cost driver within the fit-for-purpose process chain. In fact, these parts are simply hard-to-clamp and hard-to-probe, which, in addition, scatter around the target geometry. For this reason, it appears consistent that the German National Academy of Sciences Leopoldina, acatech—National Academy of Science and Engineering and the Union of the German Academies of Sciences and Humanities,³¹ has put process chain automation on top of the central future research questions in the field of additive manufacturing (Table II).

Therefore, a cyber-physical approach toward semiautonomous postprocessing of additive manufactured parts and components is presented in the following. This starts with a classification of basic use cases in a simplified cost-geometric complexity diagram (Fig. 3) with reference to Ref. 32. Starting from the point of complexity break-even (region 0), any postmachining increases the cost. The increase in the conventional postmachining cost curve is obviously higher compared to conventional machining, since the inertial AM part geometry is already complex (region 1). A further increase in complexity, again, results in a drastic increase in costs (region 2), thus challenging a potential economic use case. This can transform drastically if conventional postprocessing is enhanced by, e.g., sensing tools and/or a tool spindle, appropriate geometry acquisition hardware in combination with information technology, software, and a common data infrastructure altogether resulting in a cyber-physical system (CPS),^33,34 which substantially reduces the manual effort. The additional investment costs potentially result in a CPS break-even (region 3) followed by a significant reduction in costs due to digitization and automation. The framework for the present approach is seen in region 4 where the resulting CPS enables us to overcome or decrease singular or interacting restrictions from accessibility, clampability, measurement, and calibration efforts while contributing, e.g., as-clamped and as-milled information to an iterative process chain. Semiautonomous in the sense of this work includes the manual insertion of an AM as-built part into an electric clamping device with rough positioning using, e.g., jaws with a geometric negative while, on the other hand, the as-built geometry of the part is acquired in the as-clamped position and further used for path planning and machining. Either way, a renewed increase in costs is assumed with increasing complexity associated with, e.g., a sharp increase in the number of iterations and/or highly complex machining paths. Hence, there is region 5 representing applications where existing restrictions require individual solutions.

In addition to classic handling tasks, robots can also be applied for light milling tasks such as deburring components. Depending on the complexity of the machining task, an extension of the robot control by a Computerized Numerical Control (CNC) and integration into a Computer-Aided Design (CAD) Computer-Aided Manufacturing (CAM)-CNC process chain is necessary, as it is common for conventional machine tool based milling operations. As shown in Fig. 4(a), the process chain starts with a CAD model of a component, which is loaded into a CAM system. The CAM system supports the planner in selecting the appropriate strategy and machining parameters and performs the calculation of the tool path and the collision detection. Compared to milling with conventional machine tools, robots have decisive disadvantages with regard to their rigidity and positioning accuracy being, therefore, preferably used for processes with low machining forces and/or low quality requirements. Moreover, robot programming is rather time-consuming and requires expert knowledge, which is an additional cost driver, especially in the case of frequent system reconfigurations, as is the case of small and medium lot sizes.³⁵ To meet these challenges, new approaches are required to overcome the limitations of conventional CAD-CAM-CNC process chains. Examples are adaptive tool path planning based on the actual workpiece geometry, which addresses an increase in accuracy and the mastery of process forces³⁶ or paths planning automation approaches based on laser scans.³⁷ Verl et al.,³⁸ on the other hand, provide a comprehensive review on robot milling covering among other topics path planning. The approach presented in this work, however, transforms the formerly inflexible CAD-CAM-CNC chain [Fig. 4(a)] into a semiautonomous iterative loop based on automated geometry acquisition in the as-clamped condition followed by adaptive processing based on automated path planning [Fig. 4(b)].

For this purpose, generic path planning is adapted to consider explicitly the actual as-built and/or as-milled geometry using cooperating laser profile sensors. Moreover, this procedure is repeated until the final geometry is reached. The process forces, on the other hand, are determined with reference to the actual engagement conditions.

Considering the AM process, extensive data are already recorded. This starts with the characterization of the powder material where, e.g., particle shape, size, size distribution, porosity, and chemical composition are examined^39–42 prior to processing. Moreover, the design for AM,^43,44 the AM 3D data export,⁴⁵ build orientation,^44,46 support structure design,^47–49 and slicing operations^50,51 are realized using software solutions being, thus, digitally documented for each individual part or component. The AM process, on the other hand, is frequently analyzed considering the individual machine hardware,^52,53 build envelope workpiece filling factors,⁵⁴ while monitoring process parameters via in-process sensing⁵⁵ based on temporally and/or spatially resolved acoustic,⁵⁶ electromagnetic signatures^57,58 and/or powder bed imaging.^59,60 This means that the process duration (in terms of energy is supply) frequently has an impact on the microstructure and chemical composition (process inherent heat treatment = intrinsic heat treatment), which, unfortunately, is often neglected while attention is required.^26,61 Furthermore, thermal postprocessing in terms of tailored stress relief regimes,^27,62 heat,^63,64 and/or pressure treatments^65,66 are documented in time-and-temperature and/or pressure charts. As a result, there are parts and components as individual as the applied process chain but with a well-documented history. The latter, however, obviously affects postprocessing what in turn needs to be related (Fig. 5).

In fact, cutting forces, tool wear, surface integrity, and dimensional accuracy are influenced significantly by the identified process-determining factors set out in Fig. 5(b). Nevertheless, Montevecchi et al. stated in 2016 with regard to the postmachining of AM parts that in previous works only tool life and surface integrity aspects were investigated while the cutting forces were disregarded.⁶⁷ More recent investigations indicate a distinct correlation of the machining forces with the AM related morphology of the microstructure as well as crystallographic texture.⁶⁸ This is in line with findings from Lizzul et al.⁶⁹ showing that tool life decreases gradually up to 40% depending on the building orientation. Additionally, there is the position dependency of the chip cross section especially if considering disregarded part distortion and/or part calibration errors. This is obviously most pronounced in the case of conventional processing based on the inertial CAD model [Fig. 4(a)]. Hence, it is concluded that the cutting forces have to be recorded temporally and spatially resolved while the particular boundary conditions must be recorded too. The geometric basis for this is the as-build and as-milled geometry acquisition in the as-clamped position as already described. The proposed iterative approach [Fig. 5(a)] allows stepwise before-after comparison enabling the linkage of acquired process data with the derived tool paths and the particular change in geometry. Moreover, the temporally and spatially resolved process data shall be correlated with upstream process-determining factors, e.g., in relation to the AM built-up direction [Fig. 5(b)]. This as a substantial step toward holistic process (chain) understanding. This, on the other hand, requires adequate storage and preservation, sorting, linkage, and visualization of heterogeneous process data [Fig. 5(b)] as prerequisites for further analysis.

Central elements of the hardware concept are an industrial robot equipped with a milling spindle (HSK-E63 interface), a turntable, and an electronic vice (Fig. 6).

In addition, a tactile measuring probe including a calibration setup equipped with four carbide-measuring balls (no. 6 in Fig. 6) with exactly known position and size is used as a tactile and optical reference. Moreover, there are three cooperating blue (405 nm) laser scanners fixed to a three-axis linear system, which is mounted to the rigid base of the turntable (Fig. 6). Furthermore, a master control (type: Generic Motion Control—ARNC0 from B&R Industrie-Elektronik GmbH) is used to control the hardware elements (CNC based). Nevertheless, the scanners are mounted in a staggered arrangement (Fig. 7) in order to capture three sides at once. The entire part or component is geometrically determined by iterative scanning in combination with defined rotation of the clamped part.

The measurement is initialized by the B&R master control, which activates the measuring computer, equipped with a Beckhoff twincat software system via the OPC-UA data exchange standard. twincat, on the other hand, controls the three-axis linear system, the rotatory axis of the turntable, and the scanners. The scanners are integrated via an open interface using a DLL program library without the need for a separate control unit. All movable axes are equipped with a direct path measuring system (magnetic tape sensor principle). The measurement of the master scanner is triggered by the direct position measuring systems while the master scanner triggers the slave scanners. Nevertheless, the procedure for the as-built geometry determination in the clamped state (cf. Fig. 4) is presented in Table III. Each scanner emits a blue laser line that is reflected by the surface exposed and detected again by the scanners via scanner-integrated cameras. The cameras, arranged in a triangulation angle, transform the incident radiation in a linearized height profile. In addition to the part, the stationary carbide-measuring balls of the calibration setup are scanned. This is the basis for simple tactile calibration of the machining robot in the workpiece coordinate system (CS).

Using the final data files (Table III), the external shape of the measurement object is reconstructed line by line using homogeneous transformation matrices (TM) [Eq. (1)] in which T represents the displacement and R the rotation between two coordinate systems. To obtain the joined point cloud of the three scanners, all single point clouds must be transformed into a common coordinate system. To achieve this, the parameters of the transformation chain must be determined,

In order to combine the single point clouds of the scanners to a global point cloud and thus to reconstruct the outer shape of the measuring object, it is necessary to determine the extrinsic parameters of the three scanners. For this purpose, it is not sufficient to apply Iterative Closest Point (ICP) Algorithm, for which many powerful implementations are available.^70–72 The reason for this is that in the case of moving scanners, the single point cloud must first be assembled with the help of the stored encoder values of the three-axis system, since the different scanned profile lines of each scanner only have x and z coordinates. In order to reconstruct the scanned surface of the measuring object, each profile line must be transformed separately into a fixed coordinate system. In the present example, the coordinate system M is used (Fig. 8). In order to determine the extrinsic parameters of the scanners, a transformation chain is introduced from the workpiece (W) to the rotary table (D), the three-axis system (M), the scan head (P), and the individual scanners (S_i). Also, the corresponding coordinate systems are introduced (Fig. 8).

The transformation chain is closed via the position vector of points on the workpiece surface in the tool coordinate system (r_w) and the coordinates of the position vector of the scanned workpiece surface (r_s). Transformation relations ⁱT_i exist between the coordinate systems. To determine the parameters of the transformations (three translations dx, dy, and dz and three rotations α, β, and γ), a workpiece of suitable complexity for different rotation angles of the turntable is scanned. Afterward, a least squares optimization is used to adjust the parameters of the transformation chain to fit the scanned surface into an Standard Triangle Language (STL) model of the calibration standard (Fig. 9).⁷³ The use of the transformation chain has the advantage that, in addition to the extrinsic parameters of the scanner, accuracy measures can also be determined, as described in Ref. 74.

The algorithm used to fit the scanned point cloud into the STL model is similar to the point-to-plane type of the ICP algorithm (Fig. 10). However, in each optimization step, the transformation chain must be determined individually for each profile line of each scanner.

As already stated, the proposed procedure for automated path planning is based on the actual geometry either in the initial or intermediate state and uses the so-called adaptive planning templates. The latter are updated in every iteration until the part or component reaches the final state (Table IV).

Steps 2–5 are repeated until the target geometry is reached. This takes place in an autonomous process loop summed covering the tasks in Tables III and IV.

Microstructural heterogeneity and residual stresses due to AM processing were discussed in Sec. II of this paper. Furthermore, it was elaborated that this influences mechanical postprocessing while, in turn, mechanical postprocessing can also influence the deformation of the part or component. From this, it was derived in the first step that path planning must be based on the as-build geometry, which is preferably acquired in the as-clamped condition. On this basis, it was concluded that the cutting forces shall be linked to the engagement conditions and, in addition, preferably related to the process-determining factors from, e.g., additive manufacturing and/or heat treatment [Fig. 5(b)]. This shall be emphasized using a cuboidal Ti-Al-4V sample as an illustration (Fig. 11). There is consensus that, e.g., Ti-6Al-4V shows frequently an as-build morphology that consists of columnar, prior-beta-grains, which grow across many layers in counter-direction to the heat flux.^77–79 It is also known that the mechanical properties have frequently noticeable differences if comparing the characteristics in build-up direction and, e.g., perpendicular to it.^21,80,81 This, in turn, has an impact on the machining forces. Another influencing factor with regard to the machining forces results from unconsidered shape deviation, which frequently results in position-dependent milling cross sections (Fig. 11).

If expanding the simplified example of the cuboidal sample by including complex free-form surfaces with spatial distortion, it becomes obvious why previous works often neglected the machining forces in case of postmachining of AM parts or components.⁶⁷ Consequently, it is concluded that it is important to record machining forces and engagement conditions while relating both to the microstructural characteristics of the AM part or component (Fig. 11). This is, one the one hand, to study the implications from subtractive machining of (inherent) graded AM microstructures in depth and, on the other hand, for tool displacement compensation and/or to control machining forces and bending moments in order to avoid potential damage from postprocessing. Either way, there are many ways to determine the machining forces while multicomponent dynamometers based on piezoceramic sensors represent the most accurate commercially available solution.⁸³ Nevertheless, Salehi et al. indicate that the use of these is limited to laboratory scales due to geometric restrictions, mounting constraints, high investment costs, and dynamic measuring aspects.⁸³ Other approaches like the evaluation of drive torques or the integration of force sensors into the robot structure are already implemented but have the disadvantage of long measuring chains with noticeable measuring offsets to the interaction zone.⁸⁴ Consequently, in this approach, sensors were integrated directly in the cutting tool. Drossel et al. state in this regard that the use of commercial sensors results in an unfavorably large packaging space, while piezoceramic materials, on the other hand, offer a great potential for miniaturization and thus an advantageous alternative in terms of tool integration.⁸⁵ In fact, the integration of these types of sensors and actuators enables vibration control, condition monitoring, as well as measurement and control of dynamic forces.⁸⁵ Either way, the applied sensor concept is based on tungsten carbide sensor plates mounted on the milling tool behind the insert (Fig. 12). The use of tungsten carbide minimizes the influence of the sensor on the system stiffness. The tool-integrated sensor system, named Fraunhofer IWU SensoTool, consists of the milling tool, the gateway, the antennas, as well as the tool-integrated piezoceramic sensor. Either way, a detailed description of the sensor design can be found in Ref. 84.

The gateway is the central element of the wireless data and power transmission. Based on RFID technology, it transmits energy from the spindle-mounted antenna to the tool-mounted antenna and serves as a data interface between tool and machine. The main components of the tool-integrated sensor system comprise the piezoceramic sensor and the tool-integrated electronics.⁸⁵ The effectiveness, however, was proven in a comparative measurement with a three-component dynamometer type 9255B from Kistler (Fig. 13).

Either way, a three-axis measuring tool is used which is equipped with three replaceable cutting inserts and three sensors per insert.

The presented study explicitly takes into account central restrictions coming from additively manufactured free-form parts and components with particular regard to hard-to-clamp and hard-to-probe constraints. After rough positioning in an electric vice using 3D printed jaws, designed as geometric negative of the part to be machined, geometry acquisition is performed in the as-clamped condition. The latter by the use of cooperating blue laser (504 nm) profile scanners, which means there is no need for chalk sprays. Based on the as-clamped geometry, the so-called adaptive planning templates are derived, which represent the particular path based on the latest target-actual comparison. In this way, part deformation that has been induced by a previous processing step is taken into account. Moreover, the developed solution iterates automatically until the target geometry is reached, thus enabling semiautonomous processing. In addition, force measurement equipment is integrated while the corresponding data are linked with the scan-based description of the engagement conditions. This is seen an essential part of a holistic process data model, which links process-determining factors from additive manufacturing and postmachining. In fact, the implication of heterogeneous microstructure and/or part distortion on the machining forces is a key example. Conversely, the mastery of the process forces should be mentioned in order to avoid damage caused by postprocessing.

• The so-called “complexity for free,” attributed to AM, results in hard-to-clamp and hard-to-probe parts and components, which implies major challenges for mechanical postprocessing.

• Flexible clamping, geometry acquisition in the as-clamped position, as well as automated path planning are regarded as key elements to overcome the existing economic hurdles in AM postprocessing.

• The iterative approach serves to reduce the machining forces and enables the compensation of component deformation induced by remaining residual stresses and the cross-sectional weakening during mechanical processing.

• Iterative machining reduces the influence of the particular machining strategy on the final work result while the additional effort is countered by automation.

• Force measurement in combination and scan-based description of the engagement conditions is a by-product of the present approach but highly valuable for the evaluation of process forces under particular consideration of an heterogeneous microstructure and shape deviation.

• The CPS enabled postmachining process data must be merged with the material and process data of the AM process in order to form a holistic process data model as a basis for deep analysis.

The present work describes the implementation of a flexible manufacturing system developed toward the semiautonomous postprocessing of additive manufactured parts and components. After successful hardware and software integration, an extensive experimental test campaign is going to be performed in the next step. The test campaign is used to determine cycle times, achievable accuracies, number of iterations and process limits as well as interface requirements in terms of cross-process data fusion.

The presented study is part of the Fraunhofer Lighthouse Project “futureAM—Next Generation Additive Manufacturing” funded internally by the Fraunhofer-Gesellschaft e.V.

A. Seidel studied mechanical engineering at the Technische Universität Dresden. He is also a qualified welding engineer and has a degree in steel and light-metal engineering. In 2020, André joined the Division of Cyber-physical Production Systems (CPPS) at Fraunhofer Institute for Machine Tools and Forming Technology (IWU), focusing now on the data-driven investigation of causal relationships in complex coupled processes.

C. Gollee studied mechanical engineering at the Technische Universität Dresden with a focus on simulation methods. In 2019, he joined the Division of Cyber-physical Production Systems (CPPS) at the Fraunhofer Institute for Machine Tools and Forming Technology (IWU) and is mainly focused on the simulation-based improvement of accuracy in the machining of additively manufactured components with industrial robots.

T. Schnellhardt studied mechanical engineering with the specialization production engineering at the Technische Universität Dresden. In 2019, he joined the Division of Cyber-physical Production Systems at the Fraunhofer Institute for Machine Tools and Forming Technology (IWU). His research includes the integration of milling applications into autonomous process chains with the focus on path planning.

M. Hammer studied Mechanical Engineering at the Technische Universität Dresden and works in the field of adaptive postprocessing at the Fraunhofer Institute for Machine Tools and Forming Technology (IWU),

J. Dassing studies Mechatronics at the Technische Universität Dresden working as a student engineer in the Department of Cyber-physical Production Systems (CPPS) at the Fraunhofer Institute for Machine Tools and Forming Technology (IWU).

R. Vogt studied Mechanical Engineering at the Technische Universität Dresden and works as an application engineer in the Department of Cyber-physical Production Systems (CPPS) at the Fraunhofer Institute for Machine Tools and Forming Technology (IWU).

T. Wiese studied mechanical engineering at the Technische Universität Dresden. He deals with control engineering of machines, plants, and robots. In 2018, Torben joined the Fraunhofer Institute for Machine Tools and Forming Technology (IWU), focusing on adaptive control architectures for Cyber-physical Production Systems (CPPS).

U. Teicher studied manufacturing technology at the Technische Universität Dresden. He finished his Ph.D. about machinability of highly heterogeneous materials like cellular metals. He serves at the Fraunhofer Institute for Machine Tools and Forming Technology (IWU) as a group manager for smart manufacturing processes. His current research deals with the combination of complex machining processes and their digital representation.

A. Hellmich studied Mechatronics at the Technische Universität Chemnitz. He finished his Ph.D. on the subject of noninvasive identification of controlled system parameters for electromechanical axes. He holds the position of the Department Manager of Cyber-physical Production Systems (CPPS) at Fraunhofer Institute for Machine Tools and Forming Technology (IWU).

S. Ihlenfeldt is holder of the Chair of Machine Tools Development and Adaptive Controls at the Technische Universität Dresden. In addition, he is the Division Director of Cyber-physical Production Systems (CPPS) at Fraunhofer IWU, a member of the German Academic Association for Production Technology (WGP) and represents the Fraunhofer IWU in The International Academy for Production Engineering (CIRP).

W.-G. Drossel is holder of the Chair Adaptronics and Functional Lightweight Design in Production at Technische Universität Chemnitz. In addition, he is the Executive Director of the Fraunhofer Institute for Machine Tools and Forming Technology (IWU), the Chairman of the Fraunhofer Group for Production, and Associate Member of The International Academy for Production Engineering (CIRP).