Multi-perception graph convolutional tree-embedded network for aero-engine bearing health monitoring with unbalanced data

In engineering, severely unbalanced data from aero-engine bearings leads data-driven methods to favor normal samples and disorganize decision boundaries, triggering poor performance. Although graph networks alleviate negative impact of unbalanced samples, they have limitations on single information transmission and graph adaptive updating. As such, a multi-perception graph convolutional tree-embedded network (MPGCTN) is developed. First, a dual-channel feature graph construction method is designed to convert high-dimensional mappings into feature distance and feature dynamic graphs, boosting diverse fault information. Then, multi-scale Chebyshev graph convolutional layers with multi-perception learning are constructed as the backbone network, capturing special and shared information through discrepancy and similarity constraints. Furthermore, a tree embedded decision layer is proposed as the rebuilt output layer to gradually recognize fault locations and sizes. Finally, a triple-loss training strategy is developed to update the parameters of the MPGCTN for deep feature extraction and hierarchical decision. Experimental results of two aero-engine bearing datasets demonstrate that the MPGCTN attains the classification accuracy of 97.54 % and 98.04 % with an unbalanced ratio of 20:1, outperforming state-of-the-art methods. From the above results, the MPGCTN exhibits excellent accuracy in gradually determining fault types and severities of aero-engine bearings with unbalanced data, consistent with the fundamental principles of maintenance.