DenseNet Extension under the Generalized Zero-shot Learning Framework: A Novel strategy for Bogie Fault Diagnosis of High-speed Train

In the field of deep learning, the quality and quantity of data directly affect the effectiveness of models. However, the lack of historical samples on target fault modes is a common phenomenon in hi-tech industries. A challenging topic is discussed in this paper: Building an efficient method of fault diagnosis without involving target fault samples in the training process. Firstly, a set of attribute descriptions is defined for bogie fault modes manually. Then, the independent attribute classifier for each attribute is designed. The principle of these classifiers is to integrate the Squeeze module as an attention mechanism into DenseNet, playing a role in enhancing useful features and suppressing useless ones. Subsequently, the results of all attribute classifiers are mapped to the fault modes to achieve generalized zero-shot learning (GZSL). Finally, the effectiveness of the proposed GZSL-DenseNet-Squeeze model is verified by simulation experiments. The results show that the accuracy rate of the method proposed for known single faults is 94.77%, and for unknown composite faults is 96.04%.