Anna-Maria Schmitt

The noise in sensor data has a substantial impact on the reliability and accuracy of (ML) algorithms. A comprehensive framework is proposed to analyze the effects of diverse noise inputs in sensor data on the accuracy of ML models. Through extensive experimentation and evaluation, this research examines the resilience of a LightGBM ML model to ten different noise models, namely, Flicker, Impulse, Gaussian, Brown, Periodic, and others. A thorough analytical approach with various statistical metrics in a Monte Carlo simulation setting was followed. It was found that the Gaussian and Colored noise were detrimental when compared to Flicker and Brown, which are identified as safe noise categories. It was interesting to find a safe threshold limit of noise intensity for the case of Gaussian noise, which was missing in other noise types. This research work employed the use case of changeover detection in (CNC) manufacturing machines and the corresponding data from the publicly funded research project (OBerA).

Monitoring the metal Additive Manufacturing (AM) process is an important task within the scope of quality assurance. This article presents a method to gain insights into process quality by comparing the actual and target layers. Images of the powder bed were captured and segmented using an Xception–style neural network to predict the powder and part areas. The segmentation result of every layer is compared to the reference layer regarding the area, centroids, and normalized area difference of each part. To evaluate the method, a print job with three parts was chosen where one of them broke off and another one had thermal deformations. The calculated metrics are useful for detecting if a part is damaged or for identifying thermal distortions. The method introduced by this work can be used to monitor the metal AM process for quality assurance. Due to the limited camera resolutions and inconsistent lighting conditions, the approach has some limitations, which are discussed at the end.