The continuous aging of mechanical equipment, component wear, and technological upgrades have led to dynamic evolution in fault modes, posing significant challenges to traditional fault diagnosis models. There is an urgent need to develop fault diagnosis models with incremental learning capabilities. Although incremental learning offers a potential solution, its practical application faces three key challenges: few-shot learning, class imbalance, and corrupted historical data. To address these issues, this paper proposes an incremental fault diagnosis method based on diffusion model and wavelet transform convolution attention. we enhance the diffusion model generation process by leveraging both normal and fault samples, establishing transformation patterns from normal samples to various fault types. This improvement increases sample utilization efficiency and significantly reduces the generation difficulty for diverse fault categories. Furthermore, The proposed approach designs a wavelet transform convolution attention module with multi-scale analysis capabilities, enabling joint time-frequency domain feature extraction and multi-level feature fusion. This module effectively captures both local detail features and global trend characteristics of fault signals. Experimental results demonstrate that the proposed method achieves superior performance and low forgetting rates in every incremental learning stage, including scenarios with sufficient samples, few-shot learning, and class imbalance.