Zero-shot fault diagnosis using soft semantic embedding of diffusion-encoded probability

Fault diagnosis is essential for ensuring the stable operation of industrial equipment. However, it faces significant challenges due to insufficient fault samples and the occurrence of unseen faults during operation. For this reason, a novel semantic embedding of diffusion-encoded probability (SEDEP) is proposed for zero-shot fault diagnosis (ZSFD) in this work. A diffusion-encoded convolutional autoencoder is first trained on abundant normal data to extract features for capturing essential fault patterns from raw data. A soft semantic learning network is then constructed using a Gaussian mixture model to generate probabilities as semantic representations. A variational autoencoder-based zero-shot learning framework incorporating cross-alignment loss and distribution-alignment loss is employed to accommodate diffusion-encoded features and soft semantic information. By training on normal data to increase inter-class distance among features and enhancing semantic representation through soft semantic learning, the proposed SEDEP has outstanding nature in ZSFD. Experiments on a benchmark bearing setup and a beam chopper bearing setup achieved accuracy values of 90.15% and 94.96%, respectively. This demonstrates its robustness across different fault scenarios. Compared to state-of-the-art methods, SEDEP provides a generalizable and effective approach for addressing ZSFD tasks.