Fault diagnosis plays a critical role in maintaining the safety and reliability of mechanical equipment operations. However, vibration signals are often contaminated by irrelevant noise due to environmental factors and sensor interference, which adversely affects fault diagnosis. Convolutional Neural Networks (CNNs), leveraging their robust feature extraction capabilities and the effective noise reduction achieved by deconvolutional layers in signal reconstruction, demonstrate substantial potential in signal processing. Building upon this foundation, this paper introduce a fault diagnosis model based on convolution-deconvolution denoising and dual time-frequency path feature fusion. Designed with a modular architecture, the model comprises three components: a denoising module, a feature learning module, and a fault diagnosis module. Through convolutional and deconvolutional operations, it automatically learns temporal features of signals and reconstructs them for noise reduction. By combining characteristic fusion from both time and frequency domains, the model achieves effective fault diagnosis under strong impulsive noise interference. Experimental results on CWRU demonstrate that the proposed method not only achieves 100% classification accuracy on normal signals but also outperforms other state-of-the-art methods across various noisy environmental conditions.