Zero-shot compound fault diagnosis method driven by semantic adaptive fusion

Abstract The scarcity of compound fault samples and the complexity of feature coupling in industrial scenarios severely limit the accuracy of traditional data-driven methods and zero-shot diagnosis models. Ensuring the distribution consistency and diversity of generated samples remains a significant challenge. To address this, this paper proposes a zero-shot compound fault diagnosis method driven by semantic adaptive fusion, aiming to achieve precise identification of unseen compound faults using only single-fault data. First, continuous wavelet transform is utilized to convert one-dimensional vibration measurement signals into two-dimensional time–frequency images containing rich transient information. A wavelet vision transformer network is then introduced to extract highly robust 128-dimensional high-order visual features from these images. Second, a quantile-based adaptive semantic discretization strategy is designed to construct physical semantic benchmarks from single-fault distributions. Furthermore, a semantic adaptive attention fusion module is developed to perform adaptive interpolation of single-fault semantics in the latent space via an attention mechanism, inferring compound fault semantic labels that align with distribution consistency. Third, a conditional diffusion model incorporating auxiliary classifier constraints is constructed. Conditioned on the inferred compound semantics, this model generates pseudo-compound fault samples with high fidelity and excellent inter-class separability, effectively overcoming the mode collapse problem of traditional generative models. Experimental results demonstrate that the proposed method achieves an 81.67% recognition rate for compound faults without using any real compound fault training samples, validating its effectiveness and strong generalization capability for intelligent monitoring applications.