The gearbox plays a crucial role in the transmission system of mechanical equipment, yet its failure is frequent due to complex operational environments. Existing data-driven methods utilizing deep learning (DL) have yielded satisfactory results in intelligent fault diagnosis (IFD). However, most IFD methods rely on supervised learning, which requires a large number of labeled samples to train the DL model, posing a significant challenge in practical applications due to the difficulty of collecting sufficient annotated data. To address this issue, a pseudo-label assisted semi-supervised adversarial enhancement learning method is proposed for gearbox degradation IFD in this paper. Firstly, a semi-supervised adversarial learning framework with pseudo-label assistance is constructed, enabling full utilization of massive unlabeled data to capture representations of potential features. Subsequently, the generated labeled data, real unlabeled data as well as limited labeled data, are used simultaneously to train the adversarial model, thereby greatly reducing the requirement for data annotation. Additionally, an enhanced optimization strategy based on metric learning is integrated into the developed semi-supervised framework to align the distribution of generated and real data and then improve feature discriminability across categories. Finally, fault diagnosis experiments conducted on two self-built gearbox test rigs demonstrate that the proposed method achieves higher diagnosis accuracy and lower dependency on annotated samples than the compared ones. Comprehensive ablation studies further reveal the effectiveness and superiority of the presented approach. The related code was released at https://github.com/xinswjtu/Pseudo-label-SSAEL.

中文翻译：

---

伪标签辅助半监督对抗增强学习在有限数据下对变速箱退化进行故障诊断


齿轮箱在机械设备的传动系统中起着至关重要的作用，但由于运行环境复杂，其故障频发。利用深度学习 （DL） 的现有数据驱动方法在智能故障诊断 （IFD） 方面产生了令人满意的结果。然而，大多数 IFD 方法依赖于监督学习，这需要大量标记样本来训练 DL 模型，由于难以收集足够的注释数据，在实际应用中构成了重大挑战。针对这一问题，该文提出了一种针对变速箱退化IFD的伪标签辅助半监督对抗增强学习方法。首先，构建了一个具有伪标签辅助的半监督对抗学习框架，能够充分利用大量未标记的数据来捕获潜在特征的表示。随后，生成的标记数据、真实未标记数据和有限标记数据同时用于训练对抗模型，从而大大降低了对数据标注的要求。此外，基于度量学习的增强优化策略被集成到开发的半监督框架中，以协调生成数据和真实数据的分布，从而提高跨类别的特征可区分性。最后，在两个自建齿轮箱试验台上进行的故障诊断实验表明，与对比方法相比，所提方法实现了更高的诊断精度和更低的对注释样本的依赖性。全面的消融研究进一步揭示了所提出的方法的有效性和优越性。相关代码于 https://github.com/xinswjtu/Pseudo-label-SSAEL 发布。