Machine learning for the intelligent analysis of 3D printing conditions using environmental sensor data to support quality assurance

Process and environmental parameters that influence manufacturing processes and results are of great importance in additive manufacturing processes such as Fused Deposition Modeling (FDM). The recording and analysis of these parameters is an important task of quality assurance (QA). For this purpose, sensors are increasingly used, which continuously record the environmental data during the printing process. Subsequently, algorithms for machine learning (ML) are suitable for the data analysis of data sequences as well as for the intelligent classification of the results in defined 3D printing condition classes. In this paper different state-of-the-art ML algorithms are presented, which enable a supervised learning classification approach of environmental sensor data (temperature, humidity, air pressure, gas particles) in the FDM process. For this purpose, a new data preparation method was developed which sequences different sensor time series data. FDM sensor parameters of various 3D printing conditions were recorded, preprocessed accordingly and saved in two differently sized datasets. Furthermore, a sensitivity analysis was carried out in order to examine the influence of the individual sensor parameters on the ML analyses. Interestingly, the air pressure values were characterized as being most relevant to the analyses. Better results were always achieved with the air pressure values than without. The air pressure values have a stabilizing effect on the analyses and reduce overfitting. In the further course of the investigations, tests were carried out on the two datasets of different sizes with all considered ML algorithms as well as tests with and without the air pressure values. There, the modern XceptionTime architecture has proven to be the most effective and robust against overfitting. XceptionTime can achieve excellent results with a minimum of 95% accuracy with both a small and a large database. The Macro F1-Scores are also always above 89% and indicate a good classification for all 3D printing conditions examined. The ML investigations were then compared in a proof of concept with 3D scan examinations established in quality assurance. The 3D scans of the printed FDM components could not provide any clear information about the different printing conditions and only the component surface could be analyzed. The ML analyses, especially with the XceptionTime architecture, enable an effective alternative to quickly and easily differentiate between different 3D printing conditions. The ML time series classification presented in this work is accordingly well suited for use in an industrial environment and, with special optimizations, can be effectively applied in practice to support quality assurance in additive manufacturing. This quality assurance approach is completely new and offers immense potential to increase trust in and acceptance of additive manufacturing processes.