Elevator vibration signal denoising by deep residual U-Net

Vibration signals from elevators contain critical information pertinent to condition monitoring and fault diagnosis. However, the presence of noise in real-world data acquisition environments invariably contaminates these signals, thus compromising the effectiveness of condition monitoring and fault diagnosis. This study proposes a novel denoising method based on deep residual U-Net to mitigate noise in the vertical vibration signals of elevator cabins. The proposed network is a convolutional neural network with skip connection and multi-scale convolution structure, which can automatically learn the potential mapping between noisy and clean signals. The robustness and effectiveness are verified through experiments using real-world vibration signals compared with three conventional denoising methods in both linear and non-linear systematic indicators. Moreover, the proposed method exhibits higher accuracy and promising prospects in practical applications when applied to elevator travel distance monitoring.