Deep learning has made significant progress in fault diagnosis, but it is mostly an end-to-end intelligent diagnosis with limited application in signal denoising. This article proposes a denoising method based on fully convolutional network (FCN). Firstly, the overall model adopts the encoder decoder architecture, where the encoder part consists of three convolutional layers and the decoder part consists of four deconvolution layers. Secondly, residual connections were introduced to constrain the learning objectives of the model, allowing the model to focus more on noise information during propagation. And in order to enhance the feature extraction ability of the model, non-local blocks (NLB) are introduced in the encoder and decoder. Then, through simulation signal comparison experiments, select the hyperparameters of the network and compare them with current mainstream noise reduction methods to preliminarily verify the noise reduction effect of the model. Finally, the denoising effect of the proposed method was compared and verified through practical cases. The results showed that the method proposed in this paper achieved good application effects in both intuitive observation and denoising performance indicators, and can effectively improve the accuracy of fault diagnosis.