Hu Yunjin, Xie Qingsheng, Yang Xudong, Yang Hai, Zhang Yizong
Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education, Guizhou University, Guiyang 550028, China.
School of Mechanical Engineering, Guizhou University, Guiyang 550028, China.
Sensors (Basel). 2025 Jan 3;25(1):224. doi: 10.3390/s25010224.
Deep learning has performed well in feature extraction and pattern recognition and has been widely studied in the field of fault diagnosis. However, in practical engineering applications, the lack of sample size limits the potential of deep learning in fault diagnosis. Moreover, in engineering practice, it is usually necessary to obtain multidimensional fault information (such as fault localization and quantification), while current methods mostly only provide single-dimensional information. Aiming at the above problems, this paper proposes an Attention-based Multidimensional Fault Information Sharing (AMFIS) framework, which aims to overcome the difficulties of multidimensional bearing fault diagnosis in a small sample environment. Specifically, firstly, a shared network is designed to capture the common knowledge of the Fault Localization Task (FLT) and the Fault Quantification Task (FQT) and save it to the global feature pool. Secondly, two branching networks for performing FLT and FQT were constructed, and an attentional mechanism (AM) was used to filter out features from the shared network that were more relevant to the task to enhance the branching network's capability under small samples. Meanwhile, we propose an innovative Dynamic Adjustment Strategy (DAS) designed to adaptively regulate the training weights of FLT and FQT tasks to achieve optimal training results. Finally, extensive experiments are conducted in two cases to verify the effectiveness and superiority of AMFIS.
深度学习在特征提取和模式识别方面表现出色，并且在故障诊断领域得到了广泛研究。然而，在实际工程应用中，样本数量的不足限制了深度学习在故障诊断中的潜力。此外，在工程实践中，通常需要获取多维故障信息（如故障定位和量化），而目前的方法大多只提供单维信息。针对上述问题，本文提出了一种基于注意力的多维故障信息共享（AMFIS）框架，旨在克服小样本环境下多维轴承故障诊断的困难。具体而言，首先，设计了一个共享网络来捕捉故障定位任务（FLT）和故障量化任务（FQT）的共同知识，并将其保存到全局特征池中。其次，构建了两个用于执行FLT和FQT的分支网络，并使用注意力机制（AM）从共享网络中过滤出与任务更相关的特征，以增强分支网络在小样本下的能力。同时，我们提出了一种创新的动态调整策略（DAS），旨在自适应地调节FLT和FQT任务的训练权重，以实现最优训练结果。最后，在两种情况下进行了大量实验，以验证AMFIS的有效性和优越性。
