An Intelligent Method for Assembly Action Recognition Under Data Scarcity

This paper presents an intelligent action recognition method to address data scarcity in industrial assembly tasks. The proposed approach enables effective assembly action recognition using only a small number of samples, eliminating the need for large training datasets to retrain or fine-tuning the model—a common challenge when traditional action recognition models struggle to generalize across assembly videos with high intra-class variability. Inspired by few-shot learning techniques, we create a dataset consisting of multiple few-sample assembly action classification tasks and train a metric model using an episode-based strategy. This model compares test samples against a limited set of reference samples and predicts labels by matching them to the most similar reference instance. Using the assembly of a small two-stroke engine as a case study, our experiments demonstrate that the proposed method achieves accurate assembly action classification even with very few training samples under data-scarce conditions.