A Novel Adversarial One-Shot Cross-Domain Network for Machinery Fault Diagnosis With Limited Source Data

In recent years, meta-learning-based methods have been widely used in cross-domain fault diagnosis and promising results can be obtained even with limited target training data. However, data scarcity problems can exist not only in the target domain but also in the source domain, which puts a damper on the meta-knowledge learning process since the source domain can not provide sufficient source tasks. In this study, a novel adversarial one-shot cross-domain network named AOCN for fault diagnosis is proposed, which requires only a few source samples and as low as one labeled target sample per class. The main idea of AOCN is to learn a domain invariant embedding and generate domain invariant prototypes without causing overfitting problems. AOCN consists of two modules: a feature generator and a sample-pair discriminator with four outputs. The optimization process is divided into three steps. The first step is the meta-learning of the feature generator. The second step is the pre-training of the sample-pair discriminator to distinguish four groups of sample pairs that are generated by the pairing strategy. The third step is the adversarial learning of the two modules to confuse the features between homogeneous pairs and the features between heterogeneous pairs, respectively. Experiment results on two datasets show that AOCN can achieve more satisfactory performance than the existing methods compared.