A Transfer-Learning-Assisted Role-Differentiated Approach for Industrial Outlier Detection under Label Noise

Abstract

In this paper, a novel role-differentiated learning with noisy label (RD-LNL) approach is proposed for industrial outlier detection. A leader-follower-inspired sample selection (LFSS) strategy is introduced to choose relatively "clean samples" for establishing a robust outlier detector against label noise. Specifically, a pre-trained deep learning model is employed as the leader network to guide the training of two follower deep learning models via a joint training manner, where a selection metric is designed to facilitate sample selection by leveraging both the training dynamics and the prediction discrepancy among the models. To further enhance the possibility of selecting potential clean samples, an adaptive selection scheme is put forward to adaptively adjust the clean sample selection ratio throughout the model training process by making full use of the loss characteristics of the samples. The proposed outlier detection approach is exploited in a real-world industrial outlier detection task with application to wire arc additive manufacturing (WAAM). Experimental results demonstrate the effectiveness of the developed RD-LNL approach for WAAM outlier detection in terms of detection accuracy.