2025 International Conference on Equipment Intelligent Operation and Maintenance (ICEIOM2025)

Fault diagnosis plays a crucial role in maintaining the reliable operation of industrial equipment. However, it remains challenging due to the scarcity of labeled fault data and the emergence of previously unseen fault types during operation. To address these issues, we propose a novel method termed Semantic Embedding of Diffusion-Encoded Probability for Zero-Shot Fault Diagnosis. In the proposed framework, a diffusion-encoded convolutional autoencoder is first trained on abundant normal data to extract meaningful features that capture intrinsic fault patterns. Next, a soft semantic learning module is introduced, employing a Gaussian Mixture Model to produce probabilistic semantic representations. These are integrated into a variational autoencoder-based ZSFD framework, enhanced by a cross-alignment loss and a distribution-alignment loss, which enables the effective fusion of diffusion-encoded features with soft semantics. By training solely on normal data, the method increases inter-class separability in the feature space and enriches semantic discrimination. Experimental results on two benchmark datasets demonstrate the robustness and effectiveness of the proposal, achieving accuracies of 90.15% and 94.96%, respectively. Compared with state-of-the-art approaches, our proposal offers a generalizable and efficient solution for zero-shot fault diagnosis tasks.