Compound Fault Transfer Diagnosis of Gearboxes Based on Improved Transformer Network Under Small Datasets and Variable Working Conditions

Due to the coupling and diversity of compound fault features, it is difficult to decoupling diagnose the compound fault under small datasets and variable working conditions. To enhance the compound fault diagnosis performance, this work proposed a compound fault transfer diagnosis method that uses the multilabel Transformer-based transfer learning network. Initially, the 1-D signal is converted into 2-D image using continuous wavelet transform (CWT) as the input of the constructed deep learning network. Subsequently, a convolutional neural network (CNN)-Transformer network structure is constructed, which simultaneously utilizes convolutional operation and self-attention mechanism to comprehensively capture the multiscale fault features. In addition, multilabel classification is adopted to identify specific single-fault types in the compound fault and output corresponding fault labels. Finally, a transfer learning strategy with a fine-tuning mechanism is creatively integrated to achieve the compound fault transfer diagnosis, which can learn fault knowledge from single-fault samples in the source domain for compound fault diagnosis in required target tasks. The experimental datasets of planetary gearbox and harmonic reducer were collected in our own laboratory to validate the diagnostic performance. The results show that our proposed method achieved 98% average decoupling identification accuracy in insufficient compound fault samples on both datasets, and the accuracy fluctuation was less than 1.5% in the six transfer tasks. It is verified that the proposed method achieved excellent decoupling identification accuracy and generalization ability in small datasets and variable working conditions, which has a significant advantage over the traditional local convolution-based multilabel classification methods and domain adaptation-based transfer learning methods.