Abstract
Fused deposition modelling, like other additive manufacturing methods, has largely remained an open loop process in the absence of rigorous process monitoring and diagnostic functionality. By creating a framework that integrates quantitative diagnostic tools whose measurements are coordinated with the printing process and the system which commands the printer hardware, this paper demonstrates the feasibility of closing the loop in additive manufacturing systems. Specifically, this paper introduces the use of ultrasonic excitation as a means of detecting filament bonding failures introduced by manipulating the print bed temperature during the fused deposition modelling build process. Furthermore, this work demonstrates the capability of correcting these filament bonding failures using a correction mechanism introduced through tunable control of another process parameter of the printer. By demonstrating the detection and correction of filament bonding failures in situ, this work has demonstrated the progress toward fully closed loop control for fused deposition modeling processes.
Keywords
14.1 Introduction
Recent advances in additive manufacturing (AM) capabilities have propelled its application to many exciting and complex manufacturing problems. Yet the AM process has largely remained an open loop in the absence of process feedback mechanisms. In traditional manufacturing, the absence of in-process part qualification was acceptable due to the wealth of information on the mechanical characteristics of the stock materials, which enabled part qualification through sampled destructive testing. Yet, one of the major advantages of AM is its ability to produce incredibly complex geometries with rapid turnaround for low production volumes. Because additive manufacturing processes produce structures with bulk properties which are often anisotropic and difficult to characterize, post-process qualification becomes considerably less feasible. Therefore, a framework to integrate data acquisition, signal processing, and control feedback into the print process has become necessary. Such a system would close the loop in AM systems, providing greater levels of confidence in AM products, by increasing part quality quantification and providing part defect correction. In this research, we demonstrate progress toward a novel fault detection data acquisition system (DAQ) and integrated control feedback for fault correction in fused deposition modeling.
Previous work in additive manufacturing research has explored a wide range of diagnostic sensors, including acoustic emissions sensors [1,2,3], accelerometers, temperature sensors, and cameras [4]. Post-process vibration-based inspection techniques have also been investigated [5]. Furthermore, in-process ultrasonic part characterization has been attempted via build plate piezoelectric transducers [6] and laser Doppler vibrometer measurements [7].
As AM techniques increase in popularity and application, control feedback is becoming more crucial to creating parts that meet mechanical properties specifications, as well as meet dimensional tolerances required to be effectively used in industry. Since material properties are determined from the build history in AM processes, and hence geometry dependent, qualification of AM parts will require control over process parameters most greatly affecting the material properties [8]. Several methods of detection and correction of faults in AM parts have been explored, such as correcting the part based off of discrepancies found by measuring height differences [9], and using scans of the geometry [10]. This work approaches the problem using ultrasonic sensors and a control feedback loop to adjust certain parameters in the print process in order to correct part failure. In order to accomplish this, a unified DAQ and control framework was developed to integrate real-time diagnostics with 3D printer commands. A repeatable fault mechanism was identified and a robust corrective mechanism was found to test the unified framework. Lastly, an experimental setup for full integration of the developed unified framework and data acquisition techniques for fault detection are discussed.
14.2 Unified DAQ and Control Framework
This work began from the perspective that both the issues of process monitoring and control should be simultaneously addressed in a unified system. This governed that the system responsible for controlling the printer must also communicate with the software running online diagnostics during the print process such that any detected failures of the print could be corrected or compensated for by sending additional commands to the printer or by modifying the commands in the queue. Because there are a multitude of AM platforms and a variety of diagnostic tools under development, the framework should be constructed in a manner that would allow quick adaptation to new printers and incorporation of multiple diagnostic tools.
The developed architecture is shown in Fig. 14.1, which demonstrates the flow of data and control in the system. At the beginning of a print, the G-code machine language instructions for the 3D printer are loaded into the print control software and the contents of each field are parsed and stored. The Print Control & Feedback software sends its commands to the printer over a USB (Serial) link—a common interface for open source, FDM 3D printers. The printer responds by executing the commands it was sent, while simultaneously a set of diagnostic sensors observe the print process resulting from each command. The Data Acquisition & Processing Software takes the data from those sensors and informs the Control Feedback Software about the status of the print via statistics from each process monitoring sensor. The Control Feedback Software would then use these statistics to make decisions about whether or not it needs to modify the instructions to be sent to the printer, and how to do so.
The system developed in this work was a code set written in MATLAB that interfaced with a Lulzbot Taz 6 3D printer and a National Instruments PXI-6250 M-Series Multifunction DAQ. Like any other print control software, it accepted G-code generated from a CAD model using toolpath-planning (“slicer”) software. In this work, Cura Lulzbot Edition configured for the Lulzbot Taz 6 3D printer was used to slice the test geometry. The G-code was parsed and stored by the MATLAB code, referred to as the Print Control & Feedback Software in Fig. 14.1. This software was responsible for interfacing with the printer via a serial port opened in MATLAB. The printer accepted ASCII G-code commands via the serial port, with a line number prepended and a checksum appended. The printer responded by completing the commands it had been sent and notifying the print control software when they were finished. This setup allowed the Print Control Software to send G-code line-by-line to the printer, allowing the software to synchronize data acquisition to the start and end of each command. Once the DAQ was integrated with the control of the printer and began taking data, that data was able to be mapped back to the position where it was taken. With this mapping ability, three-dimensional visualizations of the diagnostic data were able to be created in the form of MATLAB plots.
14.3 Part Failure and Correction Methods
To test this idea of a unified data acquisition and control feedback system, a test part geometry had to be selected. The approach taken was to induce a fault in the part by changing one of the printing parameters and then having the control feedback system change a different printing parameter to correct the induced fault. In addition to being able to induce a repeatable fault, it was also desired to restrict print times to less than 30 min, so that experimental design iterations could be done quickly. Ultimately, a single shell thick, hollow square box was chosen having dimension of 40 mm × 40 mm × 20 mm, Fig. 14.2.
Debonding is a common fault occurring in FDM parts, especially in larger parts in which thermal stresses and differential cooling rates can cause debonding great enough for complete print failure. Even small debonding events can significantly reduce the strength and integrity of the part when deployed in service. Therefore, a debonding type fault was targeted as the fault to be studied. In order to introduce a fault in the printing of the test geometry, several experiments varying extrusion speed, the print head temperature, and the print bed temperature outside the range of the manufacturer recommended settings for these parameters were conducted. It was discovered that increasing the print bed temperature from 25 °C to 60–80 °C and disabling the cooling fan on the print head resulted in a repeatable fault for this geometry. By keeping the bed temperature raised for the duration of the print, temperatures in regions of the part remained above the glass transition temperature for PLA (60–65 °C). Resultantly, the plastic did not solidify enough to support the layers being printed on top of it. This caused the part to slump away from the print head, resulting in a visually detectable fault in which sections of the walls were not bonded together, Fig. 14.2.
Since the fault is induced by temperature variation, a temperature control mechanism needed to be identified to correct the faults. As such, it was found this fault can be statically corrected by enabling the cooling fan for the duration of the print with the elevated print bed temperature. The fan actively cooled the extruded PLA layer to below its glass transition temperature and solidified the plastic layer to create a rigid foundation for the next layer to be printed on. The healthy part, the faulted part, and the statically corrected part are shown in Fig. 14.2.
14.4 Control Feedback Demonstration
The logic flow to be executed during the printing of the test geometry is shown in Fig. 14.3. A characteristic signal is collected during the print and compared to a known acceptance value from healthy prints. If the characteristic value is determined to be too far outside the acceptance range, the print will be aborted to save material and machine time being wasted on a failed print. Between the acceptance range and the abort value, the characteristic can be used to determine the control action - in this case, either tune the fan speed to control the level of cooling or simply turn the fan on/off using its max speed.
To clearly demonstrate a simplified version of this process, a 40 mm × 40 mm × 20 mm single shell thick, hollow square box was built. A pause command was inserted into the G-code to temporarily stop the print once the print reached the 20 mm height and move the print bed forward. An excitation transducer (APC 90-4050) attached to the print bed was then driven at 50 kHz, and an acoustic emission sensor (APC PK15I) was placed on the print bed on the opposite side of the part from the excitation transducer. The sensor then collected fifty RMS values of the received signal based on 1.25 million samples each. This was done for twenty healthy prints with the bed temperature at 25 °C to determine the distribution of this characteristic signal. The acceptance range for the characteristic was set to be two standard deviations from the mean of the characteristic.
After this acceptance criteria was determined, the control logic could then be implemented. In this case, the same characteristic signal gathering procedure was used, then the average of the fifty RMS values was compared against the acceptance limit. Simple logic to turn the fan on to its maximum speed and continue building another 20 mm of box height was used for values which fell outside of the acceptance window, and the build proceeded with no fan cooling if it fell within. To exaggerate the level of fault previously discussed, the print bed temperature was increased to 80 °C for five prints with the control turned on. In Fig. 14.4, it is seen that the print was able to largely recover the desired geometrical shape of the part once the control is initiated. For reference the cubic, hollow structure was also printed with the elevated bed temperature and no control (Fig. 14.4, middle) and the structure continued to see deformation and unbonded layers. The diagnostic used to test the control logic was not meant to be robust, but rather to give a quick validation of the performance of fan control concept when implemented into a build. In doing so, this test also demonstrated the capability of the build to meet its prescribed shape at the top of the build when printing on a largely deformed base.
14.5 Process Monitoring System
Previous experimental investigations have used ultrasonic inspection techniques to monitor FDM builds [6, 7]. It was the intention of this work to build off the lessons learned from those studies to design a more robust system of measurement. The methodology explored in [6] involved ultrasonic transducers mounted to the build plate of the printer. One transducer pitched an excitation signal into the build plate that transmitted up into the part and the resulting vibrations were caught by three transducers on the far side of the part. Data analysis from that study indicated that either the sensors themselves or the adhesive used to bond them to the build plate degraded with repeated heat cycling, suggesting that a more robust set of sensors and sensor attachment methods would be appropriate.
A few items were identified to potentially improve on the setup used in [6] (1) improve the power of an ultrasonic excitation source, (2) use of the print head itself as a sensing location would allow automatic localization of the data and easily allow reconstruction 3D diagnostic information of the build process [11], (3) a simpler base measurement might lead to more reliable results and worthwhile statistics. To accommodate these proposed changes, the excitation source used in [7], namely the APC 90-4050, was acquired and attached to the build plate. The simpler base measurement chosen was the energy of a single frequency sinusoid excitation transmitted from the build plate up through the part into the print head. The choice of single frequency excitation tone was also informed by the success of LDV measurements of AM parts in [7], and because a single frequency would be an easy distinguishable signal in the frequency domain. The expectation was that gaps, voids, and debonding would attenuate the transmitted energy of the signal significantly, as shown in Fig. 14.5.
During each extruding command, the build plate was excited with a 40 kHz sinusoidal tone via a power ultrasonic transducer, shown in Fig. 14.6. This vibration propagated through the build plate and up into the part, which then excited the print head. An ultrasonic acoustic emissions sensor was mounted on a modified hot end mounting plate made of ABS plastic with the intention that it would convert the vibrations it experienced into a measurable voltage waveform. It was hypothesized that the strength of the measured 40 kHz tone would be strongly reduced in areas where the extruded filament did not properly bond with the layer beneath it, and that by adaptively estimating the distribution of the magnitude of the 40 kHz component of the measured signal, the integrated framework could be used to estimate the probability of the occurrence of a bond failure event and take action to turn on the cooling fan if necessary.
Figure 14.7 shows the 3D visualization for a healthy print with the setup described, including the printer head movement from its home location to start printing, and the skirt which was printed along with the test part geometry. It was found that this setup needs improvement to make this sensing technique robust. First, the ABS mounting plate for the sensor was susceptible to cracking and should be replaced by a machined aluminum plate instead. Also, the exciter/sensor pair used was not ideal. Excitation from the transducer through air was easily picked up by the sensor and washed out the expected differences in signal strength that were intended to be used to implement the feedback control continuously during the print. This aspect was actually taken advantage of in the previous section for the control logic demonstration, but is not desired in continuous inspection of the part. A more robust exciter/sensor pair for this experiment is currently being researched.
14.6 Fault Detection Discussion
Although the diagnostics for the experimental setup were not robust enough to demonstrate the full capability of the integrated DAQ and control framework, ultrasonic techniques show great promise at detecting and localizing faults in the printed parts. To further demonstrate this point, a new analysis of the data collected for [6] is presented here. The previous work mainly focused on the FRF behavior. However, if excitation energy is transferred into the printed part, then a unique signature of the pulse reflecting off the top surface of the part should be present at the receiver transducer locations. In order to identify this signature, cross-correlations of the transmitted signal with the one of the time series data samples from each measurement were computed. To highlight time dependent trends, the temporal average of the computed correlations is removed.
To give a sense of scale for the correlation plots, the pulse sent during the experiment was 0.1 s long and consisted of a sine chirp from 50–100 kHz, but the scale of the temporal separation which is plotted is 0–0.0005 s, and therefore largely describes only the front edge of the pulse.
Figure 14.8 is the plot of the correlation described above for an “empty” print, i.e. the G-code commands for the desired geometry (a square column) were executed, but no filament was extruded. The plot demonstrates relatively constant behavior across the samples. The data in this study was not integrated with the 3D printer, so sample size index was used instead of build height. A linear relationship existed between build height and sample index, as the area of each layer in the column was constant, but it has not been computed directly. The vertical lines seen in the cross-correlation plot correspond to where the transmitted signal correlates well with the measured signal. This happened repeatedly, because the pulse is long and its frequency did not change rapidly, meaning the first few sinusoidal cycles align well. These lines suggest the pulse traveled straight from the transmitter to the sensor. Theoretically, any stationary behavior like that seen above should have been removed by the demeaning process, but the data was not exactly stationary during the build possibly caused by the samples being taken while the print bed was in different locations.
When filament was included and a healthy part was built, Fig. 14.9, in addition to the artifacts of the pulse passing straight through the build plate, a diagonal trend was also observed, indicating a pulse whose arrival was becoming more and more delayed as the part grew. This trend appears to be nearly linear as well, suggesting this second pulse return corresponds to the reflection off of the top of the part as it was being built up and its height grew linearly with time.
The cross-correlation for an intentionally faulted part is shown in Fig. 14.10. The fault was induced by increasing the z-step height at layer 300 for only that layer, introducing a debonding event at that layer. In the plot above, this event corresponds roughly to sample index 280. The near linearly delayed return is seen in the cross-correlation up until the point in which this debonding event was introduced. The pulse continued to be reflected off of the fault in the part, and did not appear to propagate strongly further into the top section of the part.
These results suggest that the pulse transmitted into the plate travelled up into the part, was reflected off the top surface or the defect layer, and traveled back into the plate, where it was picked up at a later time than the signal that took the direct route to the sensors through the build plate. This phenomenon could also help explain the strange “start-up” behaviors with the general magnitudes of the FRF data presented in [6]. The constructive and destructive interference of shifted pulses could have caused strange behavior in the magnitude of the FRFs.
This data validates the idea that an ultrasonic pulse can travel through a part being built on a 3D printer and be measured by sensors on the build plate. It also validates that a fault can stop the excitation signal from propagating into layers further above the fault, which is fundamental idea proposed in the previous section for fault identification. Future work could improve the clarity of these plots by using a much shorter ultrasonic pulse with characteristics that would create a much sharper cross-correlation profiles, e.g. a random burst. This could allow an algorithm to track the delay of the peak corresponding to the reflected pulse, and compare it with the expected time of arrival for the current height of the part/print head. A stronger excitation source, such as the one used in the experiment presented in this paper and in [7], and a more sensitive transducer than the ones used in [6] may also improve the results.
14.7 Conclusions
This work demonstrates the progress toward the development of a closed loop feedback control framework in an FDM application. A repeatable fault and correction mechanism was found to introduce slump and debonding in the parts and fix them. The control logic for this test case was demonstrated in a pseudo-static application of the correction control and can easily be altered in the MATLAB code once the excitation and sensing diagnostics are improved. Although the in-process measurement attempted did not successfully detect part debonding or slumping, it was shown that the unified framework is capable of visualizing detailed 3D data collected during the build process. As this tool matures, it will enable the identification of in process diagnostics for material characterization applications and part testing. Furthermore, new analysis of data from ultrasonic inspection of FDM builds showed that the hypothesis behind the proposed setup has a large chance of success, once a better exciter/sensor combination is acquired, because excitation energy is reflected off of the debonded layer back down into the build plate and should not be received at the printer head. Eventually, the developed unified DAQ and control framework could be applied to closed loop control across a variety of AM processes, not just FDM. However, meaningful diagnostics need to drive the control and may need to be developed for the particular AM technique of interest. Further development may include incorporation of process simulation databases to use for control when diagnostic information for a healthy dataset of built parts is not available.
References
Wu, H., et al.: In situ monitoring of FDM machine condition via acoustic emission. Int. J Adv. Manuf. Technol. (84), 1483–1495 (2016)
Strantza, M., et al.: Evaluation of SHM system produced by additive manufacturing via acoustic emission and other NDT methods. Sensors (15), 26709–26725 (2015)
Clavette, et al.: Chapter 4: Real time NDE of cold spray processing using acoustic emission. In: Structural Health Monitoring and Damage Detection, Volume 7, Proceedings of the 33rd IMAC, A Conference and Exposition on Structural Dynamics, 2015
Rao, P.K., et al.: Online real-time quality monitoring in additive manufacturing processes using heterogeneous sensors. J. Manuf. Sci. Eng. 137 (2015)
Sbriglia, L.R., et al.: Embedding sensors in FDM plastic parts during additive manufacturing. In: International Modal Analysis Conference XXXIV, Orlando, FL, 2016, pp. 205–214
Cummings, I.T., et al.: In-process ultrasonic inspection of additive manufactured parts. In: International Modal Analysis Conference XXXIV, Orlando, FL, 2016, pp. 235–247
Koskelo, E.C., Flynn, E.B.: Scanning laser ultrasound and wavenumber spectroscopy for in-process inspection of additively manufactured parts. In: SPIE Smart Structures/NDE 2016, Las Vegas, NV, 2016
Gao, W., et al.: The status, challenges, and future of additive manufacturing in engineering. Comput. Aided Des. 69, 65–89 (2015). doi:10.1016/j.cad.2015.04.001
Politics & Government Week. Patents; Researchers Submit Patent Application, “Additive Manufacturing with Virtual Planarization Control”, for Approval (USPTO 20150266242) [Online]. Available: http://search.proquest.com/docview/1719986055?accountid=27865
Cohen, D.L.: Additive manufacturing of functional constructs under process uncertainty. Ph.D. dissertation, Cornell Univ., Ithaca, NY (2010)
Flynn, E.B.: Personal correspondence (2016, May)
Acknowledgements
The authors would like to extend their gratitude to Eric B. Flynn, Eliseanne C. Koskelo, and Niall O’Dowd for their thoughtful discussions on this work. This work was performed at Los Alamos National Laboratory (LANL). LANL is operated by the Los Alamos National Security, LLC for the U.S. Department of Energy NNSA under Contract No. DE-AC52-06NA25396.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2017 The Society for Experimental Mechanics, Inc.
About this paper
Cite this paper
Cummings, I.T., Bax, M.E., Fuller, I.J., Wachtor, A.J., Bernardin, J.D. (2017). A Framework for Additive Manufacturing Process Monitoring & Control. In: Mains, M., Blough, J. (eds) Topics in Modal Analysis & Testing, Volume 10. Conference Proceedings of the Society for Experimental Mechanics Series. Springer, Cham. https://doi.org/10.1007/978-3-319-54810-4_14
Download citation
DOI: https://doi.org/10.1007/978-3-319-54810-4_14
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-54809-8
Online ISBN: 978-3-319-54810-4
eBook Packages: EngineeringEngineering (R0)