Unsupervised Machine Learning for Blind Rivets Quality Inspection

Fastening plays a crucial role in aircraft manufacturing, and the demand for automated solutions has grown. Blind rivets are appealing for automation but require indirect assessment of the formed head for quality monitoring. Unsupervised machine learning holds potential for blind rivet inspection and extends to industrial data clustering/classification. In this context, labeling industrial data is challenging due to production focus and the need for NO OK labels. Unsupervised machine learning and advanced data analysis methods offer opportunities to optimize quality control processes without manual labeling or costly experiments. This paper proposes two approaches to address the issue by clustering time-dependent signals in the riveting process. After preprocessing the signals, different clustering techniques are applied to time-series and signal features to obtain OK and NO OK installation clusters. The first approach, using Euclidean distance and Dynamic Time Warping, yields poor clustering results. The second approach involves feature extraction using time domain and expert descriptors, along with dimensional reduction techniques (PCA, UMAP), followed by clustering techniques. UMAP combined with DBSCAN clustering achieves interesting results, with high precision and accuracy values (above 0.8) for both OK and NO OK clusters.