Skip to content
Open access

CLDAWT-Net: Contrastive Learning-Based Deep Adaptive Weighted Transfer Network for Open-Set Cross-Domain Recognition

2026 · IEEE Transactions on Automation Science and Engineering · Vol 23, pp. 19101-19112 · 0 citations · 29 references

Abstract

Pattern recognition systems are essential for ensuring safety and efficiency in intelligent manufacturing, yet they rely heavily on labeled data. However, datasets collected across factories exhibit significantly different distributions due to variations in production environments and schedules. More critically, discrepancies in production schedules result in only partial class overlap between source and target factories, with the target domain containing classes entirely absent from the source. Conventional unsupervised domain adaptation (UDA) methods assume a shared label space between domains, and thus systematically misclassify unshared target-domain samples into known categories, introducing unacceptable safety risks. To address this challenge, this paper proposes a Contrastive Learning-based Deep Adaptive Weighted Transfer Network (CLDAWT-Net). Specifically, a Domain Adversarial Training (DAT) module initially aligns target samples with source shared classes or rejects them as unshared. A Reliable Category Alignment (RCA) module then applies cross-domain contrastive center loss to precisely align shared-class feature distributions across domains. Finally, an Adaptive Weighted Learning (AWL) module employs a two-head orthogonal classifier to generate a per-sample soft threshold, replacing the global threshold used in prior methods and effectively separating shared from unshared classes. The proposed method improves overall classification accuracy by at least 3.0% and achieves an unshared class detection accuracy exceeding 89% in challenging industrial wheel scenarios, outperforming existing state-of-the-art methods. Note to Practitioners—Motivated by the challenges of automated wheel recognition across different factories where production environments vary and unseen wheel types often appear, this paper proposes a contrastive learning-based deep adaptive weighted transfer network. The proposed method integrates a reliable category alignment module and an adaptive weighting mechanism to automatically distinguish between shared wheel types and unshared new ones without relying on fragile manual thresholds. Extensive experimental results on industrial wheel datasets demonstrate that the proposed approach achieves a high unshared class detection accuracy and outperforms state-of-the-art methods. Consequently, it provides a robust and efficient solution for cross-factory knowledge transfer, significantly reducing data annotation costs while ensuring operational safety in intelligent manufacturing.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.