A compound fault diagnosis model for gearboxes using correlation information between single faults

Abstract

Abstract Gearboxes are key components of rotating machinery. Performing intelligent fault diagnosis of gearboxes with condition-based monitoring information helps to make reliable decisions on equipment operation and maintenance. Besides single faults, compound faults also are common failure forms of gearboxes. Conventional intelligent diagnosis models (known as single-label models) generally treat a compound fault as a new fault type, ignoring the correlations between the compound fault and the corresponding single faults. To overcome this problem, multi-label learning has been introduced and developed into multi-label models. It is also possible that different single faults are not independent but correlated with each other. Existing multi-label models, however, usually ignore this aspect. Therefore, exploiting the correlation information between single faults can further improve multi-label models. To this end, every single fault is treated as a label node , resulting in a label graph. The feature vector of each label node is initialized by the word embedding of the corresponding single-fault label. All the word embeddings are mapped using graph convolutional networks (GCN) into the parameter vectors of a set of interdependent binary linear classifiers that can directly perform multi-label classification on health categories. Meanwhile, the adjacency matrix of the label graph is adaptively learned by self-attention (SA) from node feature vectors. In this way, a novel multi-label model based on SA and GCN (referred to as SA-GCN) is proposed for compound fault diagnosis of gearboxes. SA-GCN mainly consists of a ResNet-based fault feature learning module, an SA-based adjacency matrix learning module, and a GCN-based multi-label classifier learning module. The application results on two gearbox cases show that SA-GCN outperforms conventional single-label models as well as state-of-the-art multi-label models in terms of both the diagnostic accuracy of compound faults and the overall diagnostic accuracy. Moreover, the effects of internal modules and hyperparameters on SA-GCN are also investigated.