QSCGAN: An Un-Supervised Quick Self-Attention Convolutional GAN for LRE Bearing Fault Diagnosis Under Limited Label-Lacked Data

For the fault diagnosis of rolling bearings in the liquid rocket engine (LRE), the fault data is scarce due to the high cost of doing experiments and lacks labels due to the unsure occurrence time of faults. Aiming at the above problem, in this article, an unsupervised fault diagnosis method based on quick self-attention convolutional generative adversarial network (QSCGAN) is proposed. QSCGAN consists of three convolutional sub-networks: a generator (G), a discriminator (D), and a classifier (C). G–D pair can map the noise distribution to the actual data distribution and then generate raw mechanical signals to enhance the training dataset of C. Finally, well-trained C finishes the task of fault diagnosis. By adding a self-attention (SA) layer to D and G, the network acquires a solid ability to mine features of the sample deeply. The spectral normalization (SN) to each layer parameter of G and D improves the stability and the convergence rate of the model. The experimental results on three cases of bearing fault diagnosis (CWRU, SQ, and the data of bearings in LREs) evaluate the effectiveness of the proposed method for fault diagnosis under small samples: get average accuracy of 99.73%, 98.74%, and 95.47%, respectively. The superiority of the proposed method is showed and discussed via comparing with related researches.