Presentation: Test Oracle Generation for Audio Cues in Additive Manufacturing, 10th User Conference on Advanced Automated Testing

Additive manufacturing (AM), or 3D-printing, is one pillar of the 4th industrial revolution as it allows to rapidly develop prototypes that are highly customizable. One common AM-technique is Fused Deposition Modeling (FDM) in which molten filament is layered to additively create the desired object. Even though widely used, FDM is still error-prone, with each machine suffering from wear and tear. While computer vision is promising in testing a print’s quality, it can not detect issues of occluded printer parts. Therefore, automated tests are required that support the verification of a print's quality and the early detection of looming printer flaws without visual clarity. In our work, we fulfill this need for predictive maintenance of individual FDM machines by analyzing acoustic emission. With the early detection of printer flaws, the amount of failed prints and therefore wasted time and material can be reduced, increasing the suitability of FDM printing at industrial scale. Several approaches already utilize acoustic emission to detect small deviations in 3D-prints. Still, knowledge about the audio characterization of specific machine instructions for individual FDM printers is missing in the literature. By recording audio signals for a set of machine movements, we form an audio profile that characterizes a specific FDM machine and its environment. This profile represents a test oracle that can verify whether the motors of the printer sounds as expected for a certain movement. To generate the profile and perform the tests, we preprocess the audio data and apply Fourier transformations to determine the signal’s frequency spectrum. We identify audio features in the spectrum that best characterize specific machine movements. Based on the identified features, a mapping is created that matches feature values to specific print instructions and printers. In our work, we generate a test oracle for audio cues of specific machine instructions performed on individual FDM printers. With this oracle, print sounds can be dynamically tested and evaluated at runtime. Testing against a printer’s audio profiles may hint to its wear and tear and therefore support organizations with a predictive maintenance mechanism for FDM appliances. We plan to use the gained audio profile knowledge to automate maintenance notification, such as “apply lubricant”, and as a filter for defect audio cues such as scratching noises. In addition, the mismatching audio tests may be reused to record or filter visual information to train or enhance deep learning models. Overall, the identified audio features may help organizations by introducing predictive maintenance and improving the quality assurance in 3D print farms.