Abstract—Few-shot samples and variable operating conditions have long been research challenges for intelligent health state recognition in industrial equipment. This paper proposes an intelligent health state recognition method based on a Denoising Diffusion Probabilistic Model (DDPM) embedded with an Inception-structured Self-taught Learning Network (ISLN). The proposed method first converts collected vibration signals into time-frequency representations using Synchrosqueezed Wavelet Transform (SWT). These time-frequency representations are then fed into the DDPM for sample augmentation, generating enhanced training data for the ISLN to perform health state recognition. The method is validated using a wind turbine drivetrain fault simulator. Experimental results demonstrate that the proposed method achieves the highest intelligent diagnosis accuracy in identifying various bearing and gear faults, outperforming other intelligent diagnosis models.