A Multitask Crisscross Network for One-Shot Bearing Multiattribute Fault Diagnosis
针对轴承故障数据稀缺导致诊断性能差的问题，提出多任务交叉网络，利用故障语义属性作为迁移知识，通过多任务学习和属性分类器链预测多个属性，实验证明其高效稳健。
The scarcity of samples presents a significant challenge in data-driven fault diagnosis, particularly for bearing malfunctions where swift shutdowns are imperative to avert accidents. This operational constraint makes it difficult to obtain bearing fault data, resulting in catastrophic performance on the test set, particularly when only a single sample is available for each fault. This article proposes a multitask crisscross network (MTCCN) for one-shot fault diagnosis based on transfer learning theory. The semantic information of faults is treated as attributes, serving as knowledge to be transferred. A multitask learning approach is used to predict multiple attributes, which provides more valuable fault information. The horizontal structure of MTCCN is a task-sharing network, where global features are extracted for task-specific networks to predict attributes. The vertical structure consists of an attribute classifier chain (ACC), and the correlation between attributes is modeled by a directed acyclic graph (DAG) for the attribute prediction tasks. An information map integrates multiple heterogeneous sources of information to improve prediction accuracy. Finally, extensive experiments were conducted on the different datasets, and the transfer effect was quantified, which demonstrated the efficiency and robustness of MTCCN.