反馈对比生成的零样本滚动轴承复合故障诊断

Feedback comparison for zero-shot rolling bearing composite fault diagnosis

关键词:  滚动轴承  复合故障诊断  零样本学习  特征生成  对比学习

Key words:  rolling bearing  composite fault diagnosis  zero-shot learning  feature generation  contrastive learning

摘要:

为解决复杂工业场景中滚动轴承复合故障数据难以采集所导致模型识别故障困难的问题,提出基于反馈对比生成的零样本滚动轴承复合故障诊断模型。首先,采用连续小波变换将振动信号转换为时频图,更好地保留故障的时频信息。其次,引入注意力引导ConvNeXt特征提取模块,利用通道和空间注意力机制强化故障特征的表征,消除无关信息干扰,增强故障特征的辨识性。然后,结合对抗训练和属性反馈对齐网络,确保生成的伪故障特征能够准确反映其对应的故障属性信息,实现高质量的故障特征生成。同时引入对比学习模块,生成接近正样本但远离其他样本的故障特征,进一步提高特征生成器的性能和特征的辨别力。最后,通过计算伪故障特征与未知复合故障特征的相似度,将相似度最高的类别标签作为未知复合故障标签,实现对未知复合故障的诊断。结果表明:加入注意力机制的特征提取网络相比其他网络,诊断精度提升8.42%；相比仅使用WGAN-GP生成模块,诊断精度提升14.67%；与其他模型相比,文中所提模型在故障诊断准确率上显著提高28.67%,从而验证了所提模型的有效性与优越性,为机械设备的智能维护提供了一种全新的解决方案。

Abstract:

To solve the problems of fault identification in models stemming from the difficulty of acquiring composite fault data in complex industrial environments, a novel zero-shot rolling-bearing composite fault diagnosis model based on contrastive feedback generation is proposed. Initially, continuous wavelet transform is employed to convert vibration signals into time-frequency images, thus preserving the temporal and spectral characteristics of faults more effectively. Subsequently, an attention-guided ConvNeXt feature extraction module is introduced, harnessing channel and spatial attention mechanisms to enhance fault feature representation, mitigate interference from extraneous information, and augment the distinctiveness of fault characteristics. Integrating adversarial training and attribute feedback alignment networks ensures that the generated pseudo-fault features accurately reflect their corresponding fault attribute information, achieving high-fidelity fault feature generation. A contrastive learning module is incorporated to produce fault features that are proximate to positive samples while maintaining distance from other samples, thereby further enhancing the performance of the feature generator and the discriminative power of the features. Finally, calculating the similarity between the pseudo-fault features and the unknown composite fault features, the category label with the highest similarity is assigned as the label for the unknown composite fault, thereby achieving its diagnosis. Experimental results demonstrate that the feature extraction network augmented with the attention mechanism improves diagnostic accuracy by 8.42% compared to other networks; enhances by 14.67% over using only the WGAN-GP generation module; and significantly elevates fault diagnosis accuracy by 28.67% compared to other models, thereby validating the effectiveness and superiority of the proposed model, offering an innovative solution for the intelligent maintenance of mechanical equipment.