Abstract Aerospace assembly involves millions of drilling cycles per airframe, the majority still performed by hand, and results in many out-of-tolerance holes that need to be identified and corrected. This cost in time and quality means that achieving one-way assembly and using more automation are clear goals for the industry. A large obstacle to both of these is that the inherent variability in the quality of drilled holes requires frequent time consuming inspections, and necessitates breaking apart the structure to inspect and clean the inside of stacked components. This work proposes and implements a multi-step machine learning method to identify anomalies in aerospace drilling cycles, and classify them to allow for tailored mitigation strategies to be deployed. Time series data are collected for each cycle at 1 kHz and then segmented using an algorithm developed with domain-specific knowledge. These segments are used to train clustering models that detect anomalies for each segment. The results from all segments are then used in a decision tree or random forest model to classify the anomalies. The normal condition and the engineered anomalies are selected to represent likely industrial cases, with industry-relevant tolerances defining holes as anomalies to enable investigation of process variability even under ideal conditions. The results show performance equal to or better than existing methods, with an average F1 score of 0.91, and specific types of anomaly detected with perfect accuracy. Classification of anomaly type is achieved with an average F1 of 0.63, showing promise on particular anomalies but challenges with others.
Pamplin et al. (Mon,) studied this question.