To address the issue of scarce fault samples in industrial fault diagnosis that limits the performance of deep learning models， as well as the training instability and mode collapse problems in the existing Generative Adversarial Network （GAN） -based data augmentation methods， a sample augmentation method for fault diagnosis with a high-frequency guided diffusion model was proposed. Firstly， the one-dimensional vibration signals were converted into two-dimensional gray-scale spectrograms using Short-Time Fourier Transform （STFT）， and a U-Net structure-based denoising diffusion probabilistic model was constructed. Then， a high-frequency guidance module was designed to enhance the extraction and fusion of high-frequency features via high-pass filtering and cross-attention mechanism， thereby preserving fine-grained texture information in fault signals. Meanwhile， an auxiliary classifier was introduced at each denoising step of the diffusion process to improve intra-class diversity and inter-class distinction of samples via class-conditional constraints. Finally， a multi-objective optimization strategy， combined with noise prediction and classifier guidance， was adopted to stabilize the generation process and enhance sample quality. Experimental results on two public industrial fault datasets showed that the classification model trained on samples generated by the proposed method achieved maximum diagnostic accuracies of 97.41% and 96.66%， respectively. Compared with the sub-optimal method on the two datasets， WGAN-GP （Wasserstein Generative Adversarial Network with Gradient Penalty） and Conditional Denoising Diffusion Probabilistic Model （CDDPM）， the improvements were 1.92 and 2.16 percentage points， respectively. The proposed method enhances fault diagnosis performance under few-shot conditions significantly and provides a stable and reliable solution to the class imbalance problem in industrial scenarios.