MT-CDGAT: A multi-label diagnosis model for untrained planetary gearbox compound faults based on multi-task cross dynamic graph attention networks

Compound fault diagnosis of planetary gearboxes plays a vital role in ensuring the stable operation of mechanical equipment. However, in actual complex industrial environments, obtaining sufficient and labeled multiple components compound fault data is unpractical. Using single fault data from multiple components for compound fault diagnosis is currently a significant challenge. Moreover, compound faults are frequently regarded as a new type of fault in current fault diagnosis research, ignoring the connection between different fault labels. In order to improve the above problems, the multi-task cross dynamic graph attention network model (MT-CDGAT) is proposed in this paper, which can accurately diagnose unknown compound faults of gears and bearings based solely on the single fault data. First, the feature mode decomposition (FMD) method optimized by grey wolf optimization (GWO) is adopted to reduce noise interference and enhance fault features. Then, the multi-task cross-architecture combined with multi-head dynamic graph attention network is designed to update node information from various perspectives, extract, and separate unique node features of single faults. At the same time, the proposed method can also learn the common node features of two subtasks through a shared block. Finally, the output results of two subtasks are spliced to realize multi-label fault diagnosis. The test bench dataset under different operating conditions is utilized to demonstrate the effectiveness and superiority by comparing with multiple methods.