An anti-noise one-dimension convolutional neural network learning model applying on bearing fault diagnosis. (December 2021)

• Record Type: Journal Article
• Authors: Zou, Fengqian; Zhang, Haifeng; Sang, Shengtian; Li, Xiaoming; He, Wanying; Liu, Xiaowei; Chen, Yufeng
• Is Part Of: Measurement. Volume 186(2021)
• Publication Date: 2021-12
• Subjects: Fault diagnosis -- 1DCNN -- Rotatory machinery -- Anti-noise; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110236
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Perform one-dimensional convolution to reduce the amount of calculation. Multiple different convolution kernels are used to convolve the same network layer. Using Adam optimization algorithm to continuously modify the descent direction. Performance comparisons and evaluation by feature visualization algorithm. Abstract: Since the continuous shrinking of the spatial distance between components from large-scale mechanized equipment which causes the physical characteristic signals transmitted by them to be buried in a large amount of interference noise. Aiming at the poor anti-noise performance and high computation complexity of conventional fault diagnosis methods toward rotatory machinery. This study proposes a one-dimension convolutional neural network (1DCNN) for fault diagnosis that directly worked on time-domain signals. A compact 1DCNN network structure is deeply optimized under the Pytorch software environment; features are automatically extracted from the background noise. Moreover, fully connected neural networks, softmax activation function, Adam optimization algorithm, and cross-entropy loss function are utilized for achieving the accuracy of over 99% to multi-classification tasks. Subsequently, when dealing with the noise-rich scenario, simulation results demonstrate that the accuracy rate can reach 98.31% when the signal-to-noise ratio is −8dB. What's more, the proposed model can still keep the valuable accuracy rate at 87.27% under −10 dB, which made a breakthrough contribution to the model's anti-noise performance by more than 7%. The result of this work verifies the effectiveness of the anti-noise robustness performance, thereby becoming a cornerstone inside the preventive fault diagnosis system for rotatory machinery.