Data Augmentation for Intelligent Mechanical Fault Diagnosis Based on Local Shared Multiple-Generator GAN

Deep learning based intelligent fault detection for mechanical equipment has become an important research fields. However, due to various equipment and working conditions, it is difficult to collect sufficient fault samples by monitoring sensors, which restricts the accuracy of existing intelligent fault diagnostic approaches. In this paper, we propose a novel generative adversarial networks (GAN) based on local weights-shared multi-generator for the generation of fault data and building a fault diagnosis model. Unlike existing GAN based data augmentation methods, the proposed model contains multiple generators which are divided into several groups with each groups’ prior layers of generators weights-shared. By adversarial training, the discriminator gradually improves the performance of judging authenticity and fault categories with each group of generators acquiring the ability to generate a type of fault data. In this way, the proposed local shared multiple-generator GAN method effectively reduces mode collapse and generates more diverse fault data. The performance of the proposed method was validated on the fault datasets that are not applied in training stage. Compared with traditional deep learning diagnosis methods, the accuracy of fault diagnosis could be improved up to 30%, which indicates that the data generation model can generate diverse data and expand the fault data space.