ABSTRACT

Widespread adoption of additive manufacturing (AM) is hindered by the low quality and reproducibility of printed parts. In-situ quality monitoring technologies provide early defect detection and correction capability for AM processes and are crucial for part quality assurance. Research on in-situ monitoring and control of AM is exceptionally challenging due to the numerous variations and complex reactions in AM. This study has developed a novel unsupervised recognition model for the quality of parts printed by directed energy deposition (DED). This model consists of a long short-term memory-based autoencoder (LSTM-Autoencoder) and a K-means clustering. The LSTM-Autoencoder automatically extracts features from spectra collected during the DED process, and the K-means clustering model is employed for the deposition quality classification. Experimental verifications are conducted on the quality recognition of Al7075 alloy depositions printed under different conditions by varying the laser power, printing speed, and powder delivery rate. The results show that the proposed method can discern between the unqualified depositions with rough surfaces and high porosity from the qualified depositions correctly. This proposed LSTM-Autoencoder is data-driven and thus may be applied to printing conditions not tested within this work.

摘要

打印部件的低质量和可重复性阻碍了增材制造 (AM) 的广泛采用。原位质量监控技术为增材制造工艺提供早期缺陷检测和纠正能力，对零件质量保证至关重要。由于 AM 的众多变化和复杂的反应，AM 的原位监测和控制研究极具挑战性。本研究开发了一种新的无监督识别模型，用于通过定向能量沉积 (DED) 打印的零件质量。该模型由一个基于长短期记忆的自动编码器（LSTM-Autoencoder）和一个 K-means 聚类组成。LSTM-Autoencoder 自动从 DED 过程中收集的光谱中提取特征，并采用 K-means 聚类模型进行沉积质量分类。通过改变激光功率、打印速度和粉末输送速率，对不同条件下打印的 Al7075 合金沉积物的质量识别进行了实验验证。结果表明,所提方法可以正确区分粗糙表面和高孔隙率的不合格沉积物和合格沉积物。这个提议的 LSTM-Autoencoder 是数据驱动的，因此可以应用于本工作中未测试的打印条件。