Zero-shot intelligent fault diagnosis via semantic fusion embedding

Most fault diagnosis studies rely on the man-made data collected in laboratory where the operation conditions are under control and stable. However, they can hardly adapt to the practical conditions since the man-made data can hardly model the fault patterns across domains. Aiming to solve this problem, this paper proposes a novel deep fault semantic fusion embedding model (DFSFEM) to realize zero-shot intelligent fault diagnosis. The novelties of DFSFEM lie in two aspects. On the one hand, a novel semantic fusion embedding module is proposed to enhance the representability and adaptability of the feature learning across domains. On the other hand, a neural network-based metric module is designed to replace traditional distance measurements, enhancing the transferring capability between domains. These novelties jointly help DFSFEM provide prominent faithful diagnosis on unseen fault types. Experiments on bearing datasets are conducted to evaluate the zero-shot intelligent fault diagnosis performance. Extensive experimental results and comprehensive analysis demonstrate the superiority of the proposed DFSFEM in terms of diagnosis correctness and adaptability.