Missing data often occur in multi-channel vibration signals of rotating machinery due to sensor faults, communication interruptions, or environmental disturbances, which can degrade the performance of fault diagnosis models. To address this issue, this paper proposes an improved diffusion-model-based method for missing data imputation. A masked multi-scale conditional diffusion model is developed based on denoising diffusion probabilistic models, where observed data are incorporated as conditional information to guide the stepwise generation of missing values. The joint distribution of multi-channel vibration signals is modeled to effectively capture inter-channel correlations. Regarding network architecture, a U-Net backbone is employed, with multi-scale convolutional residual blocks and linear attention modules stacked in both the encoder and decoder. This design enhances the extraction of temporal dependencies and multi-scale features of vibration signals, improving the accuracy and stability of missing data imputation. Comparative experiments are conducted on bearing and gearbox multi-channel vibration datasets under random point missingness and random block missingness scenarios. The results demonstrate that, across different missing rates, the proposed method outperforms traditional imputation methods and existing deep learning models in terms of root mean square error (RMSE), mean absolute error (MAE), and symmetric mean absolute percentage error (SMAPE). Moreover, the imputed signals preserve temporal and frequency domain characteristics that are closer to the original signals. When applied to fault diagnosis tasks, the classification accuracies reach 95.85% and 94.85% on the bearing and gearbox datasets, respectively, validating the effectiveness of the proposed method in improving multi-channel vibration signal quality and ensuring reliable fault diagnosis performance.