Knowledge Extraction in Additive Manufacturing: a Formal Concept Analysis Approach

In Additive Manufacturing (AM), it is still a major challenge to manage part quality, which is heavily influenced by feedstock materials, process settings, and in-process control. Deviations in these factors can lead to defects in the final product, particularly in Laser-Powder Bed Fusion (L-PBF), an AM process that builds parts with high precision but is too complex to control. On that account, in-situ monitoring becomes crucial for evaluating and ensuring process stability during L-PBF builds. By analyzing in-situ data, AM operators can detect process deviations and pause or stop a build as needed. Continuously monitoring in-situ data allows for immediate adjustments of process settings and avoids potential defects in the final part. However, the extensive amount of high dimensional data generated from advanced sensing technologies holds their benefit for anomaly detection and process control. The abundance and complexity of AM process monitoring data require a systematic approach to extract valuable insights and translate them into actionable knowledge. While machine learning-based techniques are nowadays essential for establishing AM process-structure-property relationships, they lack interpretability for quick actions from AM operators. In this context, we introduce a new methodology based on Formal Concept Analysis (FCA) to extract knowledge from AM multi-modal data as a set of association rules that could be used as a knowledge base to guide decisions and provide guidelines for AM operators. Specifically, for L-PBF, melt pool features are first extracted from in-situ melt pool images, and clustered into discrete categories through Machine Learning (ML) techniques. After that, FCA is allied and demonstrated to discover association rules revealing the intricate relationships between ranges of values of these features and process deviations. The final set of rules is analyzed and presented for users to adopt in monitoring their L-PBF AM processes.