Shevchik S, Wrobel R, Quang T Le, Pandiyan V, Hoffmann P, Leinenbach C, Wasmer K
Empa, Swiss Federal Laboratories for Material Science and Technology, Thun, Switzerland.
ETH Zürich, Department of Materials, Laboratory for Nanometallurgy, Zürich, Switzerland.
Heliyon. 2024 Jun 6;10(12):e32656. doi: 10.1016/j.heliyon.2024.e32656. eCollection 2024 Jun 30.
Metal additive manufacturing is a recent breakthrough technology that promises automated production of complex geometric shapes at low operating costs. However, its potential is not yet fully exploited due to the low reproducibility of quality in mass production. The monitoring of parts quality directly during manufacturing promises to solve this problem, while machine learning showed efficient performance correlating versatile manufacturing measurements with different quality grades. Today, most monitoring algorithms are based on semi- or supervised learning, thus, requiring a collection and ground-truth validation of training sets. This is costly and time consuming in real-life conditions. Our work is a feasibility study of the application of unsupervised machine learning to monitor different manufacturing regimes and quality in metal additive manufacturing. The algorithm combines the kernel Bayes rule for inference and Bayesian adaptive resonance for structuring the incoming data. Airborne acoustic emission from laser powder bed fusion is used as an algorithm input. The recognition of the main manufacturing regimes (conduction mode, stable, and unstable keyholes) are shown on real-life data, while the self-learning accuracy of developed algorithm exceeds 88 %. Our approach promises future development of plug-and-play quality monitoring systems for laser processing technology, requiring minimum modifications of the existing machines, reducing time/cost for algorithm preparation and providing continuous data driven adaptation of the algorithm to changes in manufacturing conditions.
金属增材制造是一项最新的突破性技术，有望以较低的运营成本自动生产复杂的几何形状。然而，由于大规模生产中质量的可重复性较低，其潜力尚未得到充分发挥。在制造过程中直接监测零件质量有望解决这一问题，而机器学习在将多种制造测量与不同质量等级相关联方面表现出了高效性能。如今，大多数监测算法基于半监督或监督学习，因此需要收集训练集并进行真值验证。在实际情况下，这既昂贵又耗时。我们的工作是一项关于应用无监督机器学习来监测金属增材制造中不同制造方式和质量的可行性研究。该算法将用于推理的核贝叶斯规则与用于构建输入数据的贝叶斯自适应共振相结合。来自激光粉末床熔融的空气声发射用作算法输入。在实际数据上展示了对主要制造方式（传导模式、稳定和不稳定小孔）的识别，而所开发算法的自学习准确率超过88%。我们的方法有望推动激光加工技术即插即用质量监测系统的未来发展，只需对现有机器进行最少的修改，减少算法准备的时间/成本，并使算法能够根据制造条件的变化进行持续的数据驱动自适应调整。
