Robust Vibration Signal Denoising and Diagnosis Using Encoder-Decoder Networks with Cross-layer Residual Connection

In the process of signal collection, the existence of noise is inevitable. However, noise brings great difficulties to subsequent signal analysis, fault diagnosis and life prediction. To solve the above problems, we propose an encoder-decoder network with cross-layer residual connection for signal denoising (CLR-CNN). In this network, the cross-layer residual connection can make CLRCNN predict noise and remove noise better. Non-local-block (NLB) is used to capture remote signal information. Leaky-ReLU can retain the negative information of the signal. CLR-CNN is trained by using the noisy bearing dataset, and effective signal is obtained by subtracting the noise from the original signal. Compared with traditional denoising methods and a deep learning method (NLFCNN), it has better denoising performance. In addition, CLRCNN can significantly improve the performance of fault diagnosis at the noise level.