Gradient flow-based meta generative adversarial network for data augmentation in fault diagnosis

To date, various meta-learning methods have been explored to face the data-scarcity problem in fault diagnosis. Almost without exception, these methods work on the premise that an auxiliary dataset close or related to the target data exists and can be employed as a background set to pre-train a deep network. However, in diversified industrial applications, such an appropriate dataset is not always available or requires heavy time and resource consumption for search. To address this problem, a Gradient Flow-based Meta Generative Adversarial Network (GFMGAN) is proposed for fault diagnosis of rotating machinery under the condition of insufficient training data. In the proposed method, a novel architecture taking an Unconditional Generative Model from A Single Natural Image (SinGAN) as the backbone is developed to capture the data distribution from several signal patches, permitting the training on few training data. Then, a gradient flow-based meta-learning technique is introduced into the GFMGAN to modify its learning property and further boost its generative ability. The welltrained GFMGAN can produce sufficient imitated samples. Thereafter, the quality of the generated data is evaluated by three methods, namely Fast Fourier transform (FFT)-based spectrum analysis, t-distributed stochastic neighbor embedding (t-SNE)-based feature visualization, and novel Gramian Angular Summation Fields (GASF)-based signal imaging. Ultimately, these newly generated vibration signals are employed as supplementary data to train the classification model and to finish diagnosis tasks downstream. Extensive experiments illustrate the effectiveness and superiority of the proposed framework over other state-of-the-art methods. (c) 2023 Elsevier B.V. All rights reserved.